Navigating AI conformity: A design framework to assess fairness, explainability, and performance

on Zahn, Mo i z; Zacha ias, Jan; Lowin, Maximilian; Chen, Johannes; Hinz, Oli e
A icle — Published Ve sion
Na iga ing AI con o mi y: A design amewo k o assess
ai ness, explainabili y, and pe o mance
Elec onic Ma ke s
P o ided in Coope a ion wi h:
Sp inge Na u e
Sugges ed Ci a ion: on Zahn, Mo i z; Zacha ias, Jan; Lowin, Maximilian; Chen, Johannes; Hinz,
Oli e (2025) : Na iga ing AI con o mi y: A design amewo k o assess ai ness, explainabili y, and
pe o mance, Elec onic Ma ke s, ISSN 1422-8890, Sp inge , Be lin, Heidelbe g, Vol. 35, Iss. 1,
h ps://doi.o g/10.1007/s12525-025-00770-2
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/323620
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h p://c ea i ecommons.o g/licenses/by/4.0/
Vol.:(0123456789)
Elec onic Ma ke s (2025) 35:24
h ps://doi.o g/10.1007/s12525-025-00770-2
RESEARCH PAPER
Na iga ing AI con o mi y: Adesign amewo k oassess ai ness,
explainabili y, andpe o mance
Mo i z onZahn1· JanZacha ias1· MaximilianLowin1· JohannesChen1· Oli e Hinz1
Recei ed: 25 May 2024 / Accep ed: 12 Feb ua y 2025
© The Au ho (s) 2025
Abs ac
A i icialin elligence (AI) sys ems c ea e alue bu can pose subs an ial isks, pa icula ly due o hei black-box na u e and
po en ial bias owa ds ce ain indi iduals. In esponse, ecen legal ini ia i es equi e o ganiza ions o ensu e hei AI sys ems
con o m o o e a ching p inciples such as explainabili y and ai ness. Howe e , conduc ing such con o mi y assessmen s
poses signi ican challenges o o ganiza ions, including a lack o skilled expe s and ambiguous guidelines. In his pape , he
au ho s help o ganiza ions by p o iding a design amewo k o assessing he con o mi y o AI sys ems. Speci ically, building
upon design science esea ch, he au ho s conduc expe in e iews, de i e design equi emen s and p inciples, ins an ia e
he amewo k in an illus a i e so wa e a i ac , and e alua e i in i e ocus g oup sessions. The a i ac is designed o
bo h enable a as , semi-au oma ed assessmen o p inciples such as ai ness and explainabili y and acili a e communica ion
be ween AI owne s and hi d-pa y s akeholde s (e.g., egula o s). The au ho s p o ide esea che s and p ac i ione s wi h
insigh s om in e iews along wi h design knowledge o AI con o mi y assessmen s, which may p o e pa icula ly aluable
in ligh o upcoming egula ions such as he Eu opean Union AI Ac .
Keywo ds Machine lea ning· Algo i hmic ai ness· Explainable AI· Ce i ica ion· AI audi ing· Impac assessmen
JEL Classi ica ion M15 · L86 · O30
In oduc ion
Nowadays, o ganiza ions adop a i icial in elligence (AI)
sys ems in a ious applica ion domains, including hi ing
(Van den B oek e al., 2021), heal hca e (Topuz e al.,
2018), and c edi isk assessmen (Moula e al., 2017).
Con empo a y AI sys ems can each p edic ion pe o mance
ha exceeds human capabili ies by a . As a consequence,
AI sys ems bene i socie y in a ious ways, e.g., by
ad ancing en i onmen al ini ia i es ( on Zahn e al.,
2024) o by guiding he de elopmen o new medica ions
(Fleming, 2018).
Despi e he bene i s o AI, o ganiza ions mus be awa e
o he po en ial ha ms associa ed wi h i s adop ion. These
include secu i y b eaches and da a leaks (Michael e al.,
2023), p i acy-in asi e p ac ices (Mökande & Flo idi,
2022), declining sys em pe o mance leading o poo e deci-
sion-making (dos Reis e al., 2016), biased ou pu s ha dis-
ad an age ce ain demog aphic g oups (Ba ocas & Selbs ,
2016), and a lack o accoun abili y o decisions made by
he AI (Raji e al., 2020).
In ecen yea s, esea che s and jou nalis s ha e
pa icula ly highligh ed wo majo conce ns wi h AI ha can
cause ha m: bias and opaci y (Angwin e al., 2016; Baue
e al., 2023; B ingas Colmena ejo e al., 2022; B own e al.,
2021; Jobin e al., 2019). Bias in AI sys ems can mani es
as algo i hmic disc imina ion, as esea che s and jou nalis s
ha e e ealed in nume ous cases whe e AI sys ems
ha e yielded dispa a e ou comes based on indi iduals’
sociodemog aphic cha ac e is ics (see, e.g., Angwin e al.,
2016; Cho, 2021; Fu e al., 2021; Lamb ech & Tucke ,
2019). Women, o example, a e o en sys ema ically
pu a a disad an age agains men when applying o a
Responsible Edi o : Ioanna Cons an iou.
* Mo i z on Zahn
[email p o ec ed] .de
1 In o ma ion Sys ems andIn o ma ion Managemen , Goe he
Uni e si y, F ank u , Theodo -W.-Ado no Pla z 4, F ank u
Am Main, 60323Hesse, Ge many
Elec onic Ma ke s (2025) 35:24 24 Page 2 o 24
bank loan (Fu e al., 2021) o ecei ing heal hca e (Cho,
2021). Biased AI sys ems may hus ha m subg oups o he
popula ion, ein o ce exis ing inequali ies, and impose legal
and epu a ional isks on o ganiza ions. The opaci y o AI
sys ems e e s o he black-box cha ac e o mos s a e-o -
he-a AI sys ems, such as deep neu al ne wo ks (K aus
e al., 2020). He e, o ganiza ions ypically ace a ade-
o , as s a e-o - he-a AI sys ems o e high p edic ion
pe o mance bu a e incomp ehensible o humans. The
opaci y o AI sys ems implies an inabili y o o ganiza ions
o explain c i ical decisions o ex e nal s akeholde s, such as
egula o s. Fo example, a bank may ejec a loan applica ion
as a esul o a c edi de aul p edic ion made by a deep
neu al ne wo k. Humans (e.g., audi o s o he banking
supe ision) canno unde s and why his p edic ion has been
made, which may cause he bank o ace se ious legal and
epu a ional isks (Langenbuche , 2020). In summa y, while
his pape is b oadly conce ned wi h he po en ial ha ms o
AI, we con ine ou ocus o wo speci ic issues: bias and
opaci y. These challenges a e among he mos discussed and
complex in he ield, making hem pa icula ly illus a i e o
he b oade isks posed by AI sys ems.
In o de o coun e ac he po en ial ha ms o AI sys ems,
especially wi h ega d o bias and opaci y, o ganiza ions
will inc easingly be equi ed o pe o m con o mi y assess-
men s (Mökande & Flo idi, 2022; Thelisson & Ve ma,
2024). Con o mi y assessmen s de e mine whe he he AI
sys em con o ms o pa icula p inciples, such as cybe -
secu i y (Junklewi z e al., 2023) and— he ocus o his
pape — ai ness (as opposed o bias) and explainabili y (as
opposed o opaci y). In ecen yea s, a ious egula o s and
p ac i ione s ha e called o con o mi y assessmen s ha
encompass he p inciples o ai ness and explainabili y, as
is he case o he Algo i hmic Accoun abili y Ac in he US
(117 h Cong ess, 2022), he Na ional New Gene a ion A i-
icial In elligence Go e nance Expe Commi ee in China
(Robe s e al., 2021), o he AI Ac in he Eu opean Union
(EU) (Eu opean Union, 2023). Fo example, he EU’s AI
Ac demands ce ain AI sys ems o unde go “con o mi y
assessmen p ocedu es be o e hose sys ems can be placed
on he Union ma ke ” highligh ing o e a ching e hical p in-
ciples such as “non-disc imina ion and ai ness” (Eu opean
Union, 2023). In he US, he Algo i hmic Accoun abili y Ac
demands o ganiza ions o assess whe he and how hey can
imp o e hei AI sys ems wi h ega d o “ ai ness, includ-
ing bias and nondisc imina ion” as well as “explainabili y”
and o he c i e ia (117 h Cong ess, 2022). Simila ly, he
US Ins i u e o Elec ical and Elec onics Enginee s (IEEE)
ad oca es assessing he “compliance in he de elopmen o
use o a i icial in elligence wi hin o ganiza ions” (IEEE
S anda ds Associa ion, 2022). These examples illus a e
he main eason o o ganiza ions o conduc AI con o m-
i y assessmen s, namely, o comply wi h egula ion and
indus y-wide s anda ds. Ano he eason o o ganiza ions
o conduc con o mi y assessmen s is o gain a compe i i e
ad an age. In his case, o ganiza ions may sel -commi o
con o mi y assessmen s and communica e he esul s o sup-
plie s, cus ome s, and o he s akeholde s, hus signaling he
use o con o ming AI sys ems (Cihon e al., 2021; Roski
e al., 2021).
While he necessi y seems wi hou ques ion, o ganiza-
ions s uggle wi h he implemen a ion o AI con o mi y
assessmen s. A majo eason ela es o he di e se ange o
de ini ions p oposed by esea che s ega ding AI con o m-
i y. While his holds ue o many aspec s o con o mi y
(Mökande & Flo idi, 2022), i especially applies o ai ness
and explainabili y. Fo example, de ini ions o AI ai ness
a e numbe ing in he double digi s (c. . Ba ocas e al., 2019),
and some de ini ions a e ma hema ically impossible o ul ill
a he same ime (Kleinbe g e al., 2017). Needless o say, he
di e se ange o de ini ions poses a pi o al ye challenging
ask o o ganiza ions in selec ing an app op ia e de ini ion o
ai ness ha sui s he speci ic con ex (Dola a e al., 2022).
This challenge becomes e en mo e p onounced as he cho-
sen de ini ion subs an ially in luences bo h he impac o AI
ai ness on p edic ion pe o mance (Co be -Da ies e al.,
2017) and he inancial cos s incu ed by he o ganiza ion
( on Zahn e al., 2022). Simila ly, he concep o explain-
abili y in AI sys ems is a om s aigh o wa d, as i encom-
passes a di e se ange o ideas and app oaches (Dwi edi
e al., 2023; Meske e al., 2022) and obus measu es o
objec i ely assess anspa ency s ill need o be de eloped
(F esz e al., 2024). The unce ain y su ounding de ini-
ions o ai ness and explainabili y is u he compounded
by b oad and o en gene ic egula ions. I is mos ly unclea
how ai ness and explainabili y as o e a ching p inciples
ansla e o speci ic echnical equi emen s wi hin AI sys-
ems (Mökande & Flo idi, 2022; Veale & Zuide een
Bo gesius, 2021). Adding o he di icul y, AI egula ions
a e ypically no s a ic and may be upda ed pe iodically. Fo
ins ance, as s a ed in A icle 42, Sec s.5 and 6 o he AI Ac ,
he Eu opean Commission can adap he equi emen s o
con o mi y assessmen s h ough delega ed ac s (Eu opean
Union, 2023). Pu di e en ly, AI sys ems shall con o m o
o e a ching p inciples, bu how exac ly and o wha ex en
needs s ill o be de e mined. Consequen ly, o ganiza ions
may be awa e o he need o AI con o mi y assessmen s o
comply wi h egula ions bu s ill lack guidance on he neces-
sa y s eps o mo e o wa d.
Ano he eason why o ganiza ions s uggle wi h con-
duc ing AI con o mi y assessmen s is he lack o skilled
expe s (A in e al., 2021; Benbya e al., 2020). Manually
assessing he con o mi y o AI sys ems equi es expe ise in
he domain in which he AI sys em is deployed, as well as
knowledge o law, algo i hms o a i icial in elligence, and
a guably e en e hics. Conside ing he exis ing challenges
Elec onic Ma ke s (2025) 35:24 Page 3 o 24 24
o ganiza ions ace in inding expe s o hei AI sys ems
(Chui e al., 2022), he dis inc combina ion o skills and
knowledge necessa y o conduc ing AI con o mi y assess-
men s is expec ed o exace ba e he di icul y in secu ing
quali ied p o essionals.
In his pape , we aim o suppo o ganiza ions in o e -
coming he a o emen ioned challenges by p o iding a design
amewo k and illus a i e p o o ype o conduc ing AI con-
o mi y assessmen s in a (semi-)au oma ed manne . To he
bes o ou knowledge, we a e he i s o de elop a ame-
wo k ha includes me a- equi emen s and design p inciples
o sys ems ha de elop, sha e, and assess c i e ia o AI
con o mi y, and subsequen ly implemen his amewo k in
an illus a i e a i ac ha we es and e ine h ough i e a-
i e e alua ion. The amewo k hus lays he ounda ion o
de eloping, sha ing, and assessing indus y-wide s anda ds
o AI con o mi y assessmen s. Fo his, we ollow design
science esea ch: we e iew he li e a u e and legal docu-
men s, conduc se en expe in e iews ex ac ing concep s
and hemes, de elop he design amewo k, ins an ia e ou
amewo k in an illus a i e unning so wa e a i ac , and
e alua e he a i ac in i e ocus g oup sessions. Gi en he
high ele ance o ai ness and explainabili y in egula o y
amewo ks and b oade discou se, as well as he pa icu-
la challenges o ganiza ions ace in con o ming o hese
p inciples, ou ocus is cen e ed on hese wo aspec s o AI
con o mi y.
Ou wo k makes impo an con ibu ions o bo h heo y
and p ac ice. Fi s , we p opose a design amewo k ha con-
ibu es no el design heo y on building sys ems o de el-
oping, sha ing, and assessing AI con o mi y. Resea che s
and p ac i ione s can le e age ou amewo k as a ounda-
ion o implemen ing con o mi y assessmen s, including
hose o sel -audi ing pu poses. Second, ou unning so -
wa e a i ac p o ides o ganiza ions and egula o s wi h an
illus a i e p ac ical implemen a ion o ou amewo k. In
ac , o ganiza ions can di ec ly apply ou so wa e a i ac
o hei own da ase s and models wi hin he scope o i s
cu en implemen a ion, p o iding an ini ial assessmen ha
can se e as a ounda ion o u he analysis and e inemen .
Thi d, we con ibu e o he imely opic o AI con o mi y by
p o iding quali a i e e idence on he necessi y, challenges,
and p omises o AI con o mi y assessmen s o p ac i ione s
building and le e aging AI sys ems. This quali a i e e i-
dence can suppo p ac i ione s in pe suading managemen
o he impo ance o add essing AI con o mi y, while also
p o iding a s uc u ed o e iew o an icipa ed challenges
and enabling p oac i e measu es o mi iga e hem e ec-
i ely. Las bu no leas , ou design amewo k ep esen s an
ini ial s ep owa d he collabo a i e de elopmen o indus y-
wide s anda ds o sys ema ic AI sys em e alua ions, con-
ibu ing o he li e a u e on (collabo a i e) go e nance o AI
sys ems (Bi ks ed e al., 2023) and on he s anda diza ion
o in o ma ion and communica ion echnology (Hanse h &
Bygs ad, 2015).
The emainde o his wo k is s uc u ed as ollows. Sec-
ion2 p o ides backg ound in o ma ion, and Sec .3 p esen s
ou design science me hodology. Following ha , Sec .4
desc ibes he design amewo k and ou a i ac in de ail,
including he de elopmen and esul s o hee alua ion.
Sec ion5 p oceeds by discussing ou indings and Sec .6
concludes.
Backg ound
AI con o mi y assessmen s
The concep o AI con o mi y assessmen s1 o ms a key pa
o many egula o y amewo ks. Fo example, he Eu opean
Commission’s AI Ac p oposal de ines con o mi y assess-
men s as “ he p ocess o e i ying whe he he equi e-
men s […] o his Regula ion ela ing o an AI sys em ha e
been ul illed” (Eu opean Union, 2023). Pu di e en ly, a
con o mi y assessmen p esen s he p ocess o echnically
de e mining o which ex en an AI sys em achie es di e en
c i e ia. These assessmen c i e ia can be o b oad a ie y,
including cybe secu i y, echnical obus ness, en i onmen al
sus ainabili y, human o e sigh and agency, explainabili y,
and ai ness (c. . Thelisson & Ve ma, 2024).
AI con o mi y assessmen s a e mul i ace ed p ocesses
ha a e non- i ial o o ganiza ions o implemen . As ou -
lined by p io esea ch (B own e al., 2021; Mökande &
Flo idi, 2022; Thelisson & Ve ma, 2024), o ganiza ions
aiming o conduc AI con o mi y assessmen s mus i s
cla i y bo h he pu pose and he con ex o i . The pu pose
may a y, including egula o y au ho i ies ensu ing legal
compliance, endo s o AI sys ems de ec ing mal unc ions
and epu a ional isks, and s akeholde s assessing he AI
con o mi y o o ganiza ions be o e engaging wi h hem.
The con ex is c ucial, as o ganiza ions mus g asp he
socio echnical sys em in which he AI sys em ope a es,
including in ended use s and he o ganiza ional se ing.
The AI Ac ecognizes he impo ance o con ex , whe e
con o mi y assessmen s’ necessi y and igo a e de e -
mined by he isk he AI sys em poses o human use s
1 The e m “con o mi y assessmen ” is closely ela ed o he con-
cep o algo i hmic audi ing (B own e al., 2021). Fo example, he
IEEE de ines algo i hmic audi s as “an independen e alua ion o
con o mance o so wa e p oduc s and p ocesses o applicable egu-
la ions, s anda ds, guidelines, plans, speci ica ions, and p ocedu es”
(IEEE, 2008), a de ini ion synonymous wi h he concep o con-
o mi y assessmen s. In his pape , we adop he e m “con o mi y
assessmen ” as a b oade concep ha includes, among o he e ms,
“algo i hmic audi ing” (B own e al., 2021) and “impac assessmen s”
(117h Cong ess, 2022).
Elec onic Ma ke s (2025) 35:24 24 Page 4 o 24
(Eu opean Union, 2023). Mo eo e , al hough o ganiza-
ions may p ima ily conduc con o mi y assessmen s o
ex e nal s akeholde s, he e is g owing suppo among
p ac i ione s and esea che s o he implemen a ion o
in e nal AI assessmen amewo ks du ing he de elop-
men s age (Raji e al., 2020). This would ensu e ha
model de elope s p io i ize no only p edic i e pe o -
mance bu also a b oade ange o c i e ia and e hical
p inciples. Howe e , his app oach equi es cla i ying
con ex - ela ed dependencies ea lie in he p ocess, mak-
ing implemen a ion e en mo e challenging.
Beyond he challenges o con ex and pu pose, he ope -
a ionaliza ion o assessmen c i e ia p esen s ano he di -
icul y. P io esea ch highligh s ha he p ac ical imple-
men a ion o assessmen c i e ia emains la gely unclea
(Mökande & Flo idi, 2022; Thelisson & Ve ma, 2024).
Fo ins ance, while he e is consensus on he impo ance
o ai ness as a key equi emen o AI sys ems, he e is li -
le ag eemen on how o ope a ionalize ai ness in p ac ice
(Feue iegel e al., 2020). These disag eemen s, along wi h
a lack o coo dina ion, ha e hinde ed he de elopmen o
s anda dized me hods o con o mi y assessmen s and le
o ganiza ions wi hou he means o assess he con o mi y o
hei AI sys ems.
To he bes o ou knowledge, we a e among he i s o
p o ide quali a i e e idence on he need o and po en ial
o s anda dized, con ex -dependen con o mi y assessmen s
om a p ac i ione ’s iewpoin and o de elop a design
amewo k o such assessmen s. We hus lay he ounda ion
o mo e accessible and e ec i e AI con o mi y assessmen s
and he eby help o ganiza ions communica e he compliance
o hei AI sys ems o s akeholde s. Gi en he wide ange o
con ex s and c i e ia o AI sys ems, ou me hodology neces-
si a es a ocus on speci ic cases ha e lec cu en p ac ical
challenges. In his pape , we ocus on he inancial con ex
o aud de ec ion, a domain whe e AI has di ec implica-
ions o human li es, such as he isk o unjus ly lagging
indi iduals as audulen , and may hus quali y as a “high-
isk” applica ion in egula ions such as he EU’s AI Ac . We
u he ocus on wo cen al and highly ele an assessmen
c i e ia beyond p edic i e pe o mance: ai ness and explain-
abili y (Baue e al., 2023; B ingas Colmena ejo e al., 2022;
B own e al., 2021; Jobin e al., 2019). We ocus on hese
wo c i e ia because hey exempli y he b oade challenges
in de ining and ope a ionalizing assessmen c i e ia o AI
con o mi y. Bo h ai ness and explainabili y a e b oad con-
cep s wi h compe ing—and some imes con lic ing—de ini-
ions and me hods o ope a ionaliza ion (Baue e al., 2023;
Feue iegel e al., 2020). Despi e hese complexi ies, hey
a e conside ed essen ial by policymake s (see, e.g., Thelis-
son & Ve ma, 2024) and ha e spa ked ex ensi e academic
discussions, leading o he es ablishmen o dedica ed con-
e ences such as he Con e ence on Fai ness, Accoun abili y,
and T anspa ency (FAccT). The ollowing wo subsec ions
will de ail he c i e ia in ocus.
Fai ness inAI sys ems
The i s assessmen c i e ion we ocus on is ai ness, as
esea che s ha e inc easingly emphasized he need o i
in ligh o g owing e idence o AI sys ems exhibi ing bias
agains ce ain indi iduals. Fo example, p io esea ch has
shown ha AI sys ems in inance may pu women a a dis-
ad an age by g an ing hem disp opo iona ely less c edi
(Fu e al., 2021) and showing hem ewe ad e isemen s o
high-paying jobs (Lamb ech & Tucke , 2019). Simila ly,
jou nalis s and esea che s ha e demons a ed how AI in he
c iminal jus ice sys em has alsely classi ied black de end-
an s as “a isk” mo e equen ly han non-black de endan s
(Angwin e al., 2016). Such bias can pe pe ua e exis ing
inequali ies, hinde social p og ess, and expose o ganiza-
ions o legal isks.
As a emedy o bias, esea che s ha e p oposed me hods
o measu e and p omo e ai ness in AI sys ems (see Ba ocas
e al., 2019, o an o e iew). O e ecen yea s, a ious
de ini ions o ai ness in AI sys ems eme ged and a e ypi-
cally desc ibed ei he a he le el o indi iduals o he le el
o g oups (Dola a e al., 2022; Feue iegel e al., 2020). The
o me , indi idual ai ness, elies on a concep o simila i y:
indi iduals wi h simila p ope ies should ecei e simila
ou comes (Dwo k e al., 2012). Howe e , in p ac ice, de in-
ing a sui able measu e o simila i y can be challenging. The
la e , g oup-le el ai ness, s ipula es ha ou comes associ-
a ed wi h he AI sys em should be equally dis ibu ed inside
and ou side o a g oup (Dwo k e al., 2012; Ha d e al.,
2016). G oups a e ypically iden i ied by an a ibu e deemed
sensi i e, such as ace, age, o gende (Ba ocas & Selbs ,
2016; Ba ocas e al., 2019).
G oup-le el ai ness is o pa icula ele ance bo h in aca-
demia (see, e.g., Feue iegel e al., 2020) and in egula ion
(see, e.g., Ba ocas & Selbs , 2016). Howe e , e en wi hin
g oup-le el ai ness exis s a b oad a ie y o de ini ions o
choose om (c. . Ba ocas e al., 2019). The mos p ominen
examples a e s a is ical pa i y (Dwo k e al., 2012), which
ep esen s independence be ween sensi i e a ibu es and
he dis ibu ion o p edic ions, and equalized odds (Ha d
e al., 2016), which ep esen s independence be ween sensi-
i e a ibu es and he dis ibu ion o p edic ion e o s. C u-
cially, some de ini ions o g oup-le el ai ness a e compe -
ing and e en ma hema ically impossible o ul ill a he same
ime (Kleinbe g e al., 2017). Mo eo e , a a ie y o op ions
a ises when conside ing he sensi i e a ibu e ha de ines
he g oups in luencing g oup-le el ai ness. He e oo exis s
a la ge a ie y o di e en sensi i e a ibu es ha depend
on he ( egula o y) con ex . Fo example, unde he Equal
C edi Oppo uni y Ac in he US, nine di e en a ibu es

Elec onic Ma ke s (2025) 35:24 Page 5 o 24 24
a e deemed sensi i e (Smi h, 1977). O e all, g oup-le el
ai ness p esen s a mul i ace ed concep wi h high egula-
o y ele ance ha equi es ca e ul conside a ion.
Explainabili y inAI sys ems
When adop ing AI sys ems, o ganiza ions may nowadays
encoun e a ade-o be ween high p edic ion pe o mance
and model in e p e abili y (Meske e al., 2022). S a e-o -
he-a models, such as a i icial neu al ne wo ks, o en
exhibi he highes accu acy in complex p edic ion asks
which makes hem pa icula ly a ac i e o o ganiza ions
(K aus e al., 2020). Howe e , hese models a e ypically
opaque, ha is, hey a e o “black-box” cha ac e impeding
he abili y o human use s o unde s and hei ou comes
(Meske e al., 2022). The opaci y o AI sys ems can ha e
conside able downsides, such as impai ed use us and
es ic ed con es abili y (c. . Rosen eld & Richa dson, 2019).
Resea che s ha e de eloped me hods o explainabili y
as a emedy o opaci y in AI sys ems (Meske e al., 2022).
Typically, esea che s conside ea u e-based explana ions as
s a e-o - he-a (Baue e al., 2023; Hsieh e al., 2020), such
as SHAP alues (Lundbe g & Lee, 2017). SHAP is a model-
agnos ic me hod ha uses addi i e ea u e a ibu ions o p o-
ide in e p e able explana ions o black-box model p edic-
ions, o e ing con as i e explana ions on an indi idual le el.
As a consequence, human use s a e able o be e in e p e he
ou come o AI p edic ions (Baue e al., 2023).
The explainabili y o AI ou comes p omises a ious socie al
and business- ela ed bene i s (Coussemen & Benoi , 2021),
such as inc eased use us owa ds AI sys ems (Rosen eld &
Richa dson, 2019) and he abili y o communica e he a ionale
behind AI ou comes o s akeholde s (Wang e al., 2022). As
a consequence, ecen legal ini ia i es demand o ganiza ions
o make hei da a-d i en decisions explainable (see, e.g., he
p oposed Algo i hmic Accoun abili y Ac in he US, 117 h
Cong ess, 2022).
Rela ed ools and amewo ks
To he bes o ou knowledge, he e a e no well-es ablished
AI con o mi y assessmen ools add essing p ac i ione s’
needs and p omo ing indus y-wide s anda diza ion.
Howe e , he e a e ea ly wo ks on go e nance amewo ks
and open-sou ce oolki s aimed a a gene al assessmen o AI
sys ems using ques ionnai es o common me ics o assess
p edic ion pe o mance, ai ness, and o he aspec s. In he
ollowing, we p o ide an o e iew o hese ea ly wo ks.
CapAI, in oduced by Flo idi e al. (2022), is a p ocedu e
de eloped o align AI sys ems wi h he con o mi y c i e ia o
he EU’s AI Ac . I p o ides o ganiza ions wi h a s uc u ed
app oach o e alua ing he e hical, legal, and echnical
obus ness o AI sys ems in he o m o a s ep-by-s ep
assessmen guide. Use s can wo k h ough he p o ided
checklis , answe key ques ions, pe o m he necessa y
analyses manually, and submi hei esul s o egula o y
au ho i ies using in o ma ion empla es. Howe e , o he bes
o ou knowledge, he app oach is de eloped solely om a
egula o y pe spec i e, lacks a use -cen ic design, and is no
implemen ed as a so wa e.
Ano he ela ed ool is Fai X (Sikde e al., 2024). Fai X
is an open-sou ce benchma king ool designed o e alua e AI
models in e ms o ai ness, pe o mance, and explainabili y.
The use begins by loading ei he abula , image, o cus om
da ase s in o he ool. The ool au oma ically p ep ocesses
he da a, ains a model based on he uploaded da ase , and
compu es di e en e alua ion me ics. No ably, he ool
also in ol es bias mi iga ion echniques, such as ad e sa ial
de-biasing, o coun e ac ai ness p oblems. Fu he mo e,
he e a e se e al o he assessmen ools ha a e simila o
Fai X (see, e.g., AI Fai ness 360 by Bellamy e al. (2019),
o Fai lea n by Bi d e al. (2020)). The ocus o hese ools,
howe e , lies p ima ily in implemen ing exis ing me ics
and me hods om he li e a u e, wi hou he aim o guiding
o ganiza ions in deeply explo ing speci ic aspec s, c ea ing and
sha ing use cases, o documen ing he e olu ion o AI sys ems
o e ime.
Ou amewo k and so wa e a i ac o con o mi y assess-
men s di e s om exis ing app oaches in se e al ways. Ou
design amewo k is based on he academic li e a u e, legal
documen s, and expe in e iews, hus add essing AI con-
o mi y assessmen s mo e comp ehensi ely. The amewo k
places a s ong emphasis on ( he de elopmen o ) indus y-
wide s anda ds wi h he ul ima e goal o no jus implemen ing
me ics and me hods bu guiding o ganiza ions in applying
hem. Fo example, we add unc ionali ies ha p omo e he
es ablishmen and u iliza ion o s anda ds and bes p ac ices o
he echnical implemen a ion o con o mi y assessmen s in di -
e en use cases. Mo eo e , as opposed o ela ed app oaches,
ou amewo k builds upon dis inc use oles h oughou he
assessmen p ocess, allowing di e en s akeholde s, such as
AI owne s and audi o s, o pe o m speci ic asks sui ed o
hei expe ise. Addi ionally, he so wa e ins an ia ion o ou
amewo k includes a g aphical use in e ace, making he
assessmen accessible o a b oade ange o use s. The in e -
ace p esen s he esul s in a clea and in e p e able manne
so ha no only echnical expe s bu also managemen and
egula o s can assess he le el o AI con o mi y.
Resea ch me hodology
O e all esea ch design
We de elop he amewo k o AI con o mi y assessmen s
ollowing design science esea ch, which “c ea es and
Elec onic Ma ke s (2025) 35:24 24 Page 6 o 24
e alua es a i ac s in ended o sol e iden i ied o ganiza ional
p oblems” (He ne e al., 2004). In ou case, we c ea e a
design amewo k o o ganiza ions o conduc AI con o m-
i y assessmen s and, he eby, signal AI con o mi y o he
ma ke and adhe e o upcoming egula ions. We u he
ins an ia e his amewo k in o a unning so wa e a i ac
and e alua e he a i ac in ocus g oup sessions.
We adop he design esea ch cycle p oposed by Kuechle
and Vaishna i (2008) as ou me hodological ounda ion, hus
ollowing an i e a i e p ocess o con inuous e alua ion and
adap a ion o he a i ac . The design cycle consis s o i e
phases: p oblem awa eness, sugges ion, de elopmen , e alu-
a ion, and conclusion. In he p oblem awa eness phase, he
esea che s iden i y and de ine he p oblem by e iewing
li e a u e and in e iewing expe s. In he sugges ion phase,
hey de i e me a- equi emen s and o mula e design p in-
ciples g ounded in scien i ic heo ies and expe ise. The
de elopmen phase in ol es ins an ia ing he design ame-
wo k, in he o m o a so wa e a i ac , me hods, models,
o cons uc s (Ma ch & Smi h, 1995). The a i ac is hen
e alua ed using ocus g oup sessions wi h expe s (Me h
e al., 2015), labo a o y expe imen s (Gnewuch e al., 2017),
o o he es ablished e alua ion me hods. Finally, he p ojec
concludes, and e alua ion esul s in o m subsequen i e a-
ions i needed.
Design science esea ch is pa icula ly well-sui ed o
add ess he challenge o how a so wa e ool can suppo
AI con o mi y assessmen s, as his challenge is inhe en ly
socio- echnical in na u e. I s unique abili y o b idge echno-
logical and o ganiza ional pe spec i es enables i o ackle
such challenges, wi h leading schola s e en desc ibing
design science esea ch as “essen ial” o add essing socio-
echnical ques ions (Abbasi e al., 2024). Fu he mo e, by
in eg a ing heo y wi h p ac ical conside a ions wi hin i s
design cycle, design science esea ch enables he de elop-
men o a i ac s ha no only ad ance academic knowl-
edge bu also p o ide ac ionable solu ions o p ac i ione s.
This dual ocus is pa icula ly aluable in he con ex o
AI con o mi y assessmen s, whe e egula o y, e hical, and
ope a ional conside a ions mus be balanced wi hin o gani-
za ional en i onmen s.
Design cycle implemen a ion
In he ollowing, we de ail he applica ion o he gene al
design cycle o ou speci ic use case. In Fig.1, we show he
i e phases o he gene al design cycle (le ) along wi h ou
speci ic implemen a ion ( igh ).
P oblem awa eness
In he p oblem awa eness s age, we conduc semi-s uc u ed
in e iews wi h se en expe s in he ield o AI con o mi y
(see Table1). We ca e ully selec expe s who employ an
o ganiza ional, p ac ice-o ien ed pe spec i e on AI sys ems.
Acco dingly, we choose se en expe s based in Ge many
and Swi ze land who ei he consul companies on aspec s o
AI (Expe s 1, 3, 4, 5, 6) o e iew companies’ AI sys ems
om unding (Expe 2) o egula o y pe spec i es (Expe
7). Impo an ly, all expe s, wi h he excep ion o Expe 7,
wo k in e na ionally, engaging wi h AI sys ems a leas a
he Eu opean le el. We conduc all in e iews online and
ollow a semi-s uc u ed p o ocol. We ini ially explo e he
ele ance o AI con o mi y h ough open-ended ques ions,
a oiding any p iming. Subsequen ly, we del e in o hypo-
he ical scena ios o gauge expe insigh s on AI assessmen
c i e ia. Finally, we in oduce and sea ch o eedback on he
concep o an ea ly- e sion AI con o mi y assessmen ool,
ensu ing i does no in luence ea lie esponses. We p o-
ide he de ailed p o ocol in Appendix A. Fo he quali a i e
analysis o he in e iew esponses, we c ea e eco dings and
ansc ip s. We subsequen ly in es iga e codes and gene al
hemes o he in e iew ia con en ional con en analysis.
No ably, con en ional con en analysis is pa icula ly sui ed
o s udy he meaning o ex da a when exis ing heo y and
esea ch a e limi ed (Hsieh & Shannon, 2005). Speci ically,
Fig. 1 Ou esea ch me hod-
ology (based on Kuechle &
Vaishna i, 2008)
Elec onic Ma ke s (2025) 35:24 Page 7 o 24 24
we ollow Gioia e al. (2013) o cap u e i s -o de concep s
and cons uc second-o de as well as agg ega e hemes by
clus e ing and in e p e a ion. To do so, h ee esea che s
engage in coding ansc ip s and subsequen ly in e p e ing
codes o highe -o de concep s. They ollow an i e a i e con-
sensual p ocess, i.e., wo esea che s conduc he coding and
in e p e a ion independen ly, a e which hey discuss hei
esul s unde he guidance o a hi d, independen esea che
o each a consensus.
Ou coding p ocedu e can be illus a ed by wo ela ed
ye dispa a e quo a ions. The i s quo a ion says ha “mos
companies, especially small and medium-sized, simply do
no deal wi h [de eloping con o ming AI] a all because hey
don’ ha e he capaci y o i .” The second quo a ion e e s
o AI audi ing, speci ically, a “bo leneck, o example, in
he medical de ices sec o , whe e inspec o s a e s uggling
o ge he job done a all, especially in he con ex o he
sho age o skilled wo ke s.” One code assigned bo h quo-
a ions o one holis ic i s -o de concep e e ing o he
gene al need o highly-skilled pe sonnel in he con ex o
AI con o mi y. By con as , ano he code assigned hem o
wo dispa a e ye ela ed i s -o de concep s: “Complexi y
and mul i ude o con o mi y aspec s ha equi e specialized
expe s ( om, e.g., compu e science and law),” which e e s
o he need o a a ie y o expe s o success ully de elop
con o ming AI, and “Highly specialized (in e nal and ex e -
nal) AI audi o s a e cu en ly lacking”, which e e s o he
sca ci y o expe s o conduc ing con o mi y assessmen s.
A e discussing he con lic wi hin he esea ch eam, we
ag eed on hese wo mo e speci ic i s -o de concep s.
The esea che s epea he p ocess un il he e is a sa is-
ac o y con e gence in he in e p e a ions o highe -le el
codes, as de e mined by he independen esea che . This
app oach allows us o sys ema ically de i e insigh s om he
in e iews o la e o m he basis o ou design amewo k.
Sugges ion
In he sugges ion s age, we de i e he me a- equi emen s and
design p inciples o ou amewo k based on ou indings
o he p oblem awa eness. Design science esea ch li e a-
u e p oposes expe in e iews combined wi h insigh s
om exis ing li e a u e as an impo an sou ce o design
knowledge (Miah & Genemo, 2016) and, mo e speci ically,
he o mula ion o me a- equi emen s (Heinz e al., 2024).
Thus, we ansla e each p e iously iden i ied p oblem in o
one me a- equi emen acco ding o ou unde s anding which
high-le el needs a so wa e a i ac should ul ill o add ess
hese p oblems. Nex , we de i e design p inciples o each
me a- equi emen , based on ou pe spec i e o wha speci ic
design componen s a e needed o ul ill all me a- equi e-
men s. Fo he o mula ion o he design p inciples, we d aw
on he amewo k o ac ion and ma e iali y-o ien ed design
p inciples acco ding o Chand a e al. (2015).
De elopmen
In he de elopmen s age, we ins an ia e he p e iously
de i ed me a- equi emen s and design p inciples in o a
unning so wa e a i ac , ac ing as a p o o ype o conduc -
ing con o mi y assessmen s. We implemen he a i ac as a
web-based applica ion based on Py hon o compu ing he
me ics, Flask 2.3—a ligh weigh Py hon amewo k— o
he web se e , MySQL as a da abase managemen sys em,
and de aul Boo s ap hemes o he on end design. This
con igu a ion allowed us o add and al e unc ionali ies eas-
ily. Designing he a i ac as a web applica ion allows easy
so wa e sha ing wi h ex e nal expe s, o ins ance, du ing
la e ocus g oup sessions. The a i ac is publicly a ailable
and can be accessed a h ps:// gi hub. com/ mlowin/ con o
mi y_ asses smen . O cou se, i is also possible o un he
web se e in a p i a e ne wo k wi hou access o he In e -
ne . This is especially aluable o companies and use cases
in ol ing con iden ial da a.
E alua ion
In he e alua ion s age, we aim o assess ou design ame-
wo k and so wa e a i ac on AI con o mi y assessmen s
using ocus g oups in line wi h p io wo k in design science
Table 1 O e iew o
in e iewed expe s Expe Ti le O ganiza ion Yea s o
ele an expe-
ience
1 Managing Di ec o AI es ing cen e 6
2 Di ision Manage Technology esea ch und 5
3 Head o AI Technical inspec ion associa ion 2
4 AI solu ion a chi ec Technical inspec ion associa ion 7
5 Manage Technical consul ancy 9
6 Manage Technical consul ancy 12
7 Senio Risk Manage Na ional banking supe ision 17
Elec onic Ma ke s (2025) 35:24 24 Page 8 o 24
esea ch (e.g., Hi e al., 2019; Me h e al., 2015; Zacha ias
e al., 2022). We ollow Venable e al. (2012) o conduc
a o ma i e e alua ion, ha is, we aim o de i e po en ial
imp o emen s o ou a i ac and, he eby, o ou design
amewo k (He ne e al., 2004). Wi h ega d o he pa a-
digm o ou e alua ion, we build upon an a i icial se ing,
ha is, we conduc bo h emo e and on-si e ocus g oup
sessions in which we e alua e ou a i ac in a hypo he ical
scena io. The main eason is ha he a i icial se ing allows
us o limi he in e e ence o con ounding a iables and
esou ce cons ain s. We conduc i e sepa a e ocus g oup
sessions, each co esponding o a di e en indus y pa ne ,
yielding a o al o 21 expe ienced p ac i ione s as pa ici-
pan s e alua ing ou a i ac . Table2 p o ides an o e iew
and key desc ip i es o ou conduc ed ocus g oup sessions.
In he ocus g oup sessions, pa icipan s e alua e he
p oposed so wa e a i ac by assessing he con o mi y o
an AI sys em, ha is, a machine lea ning classi ie ha we
implemen ed using eal-wo ld da a om aud de ec ion.
The classi ie employs g adien boos ing (Chen & Gues in,
2016) and ollows s anda d p ac ices o ain- es spli ing
and hype pa ame e uning (Has ie e al., 2017). The ses-
sions las be ween 45 and 90min and consis o h ee pa s:
Pa 1. We in oduce he hypo he ical scena io in
which pa icipan s need o assess he con o mi y o
an AI sys em o aud de ec ion. Fo his, we b ie ly
ecap he p oblem o bias and opaci y in AI as well as
upcoming AI egula ions o make pa icipan s cognizan
o he p oblem. We hen p esen he AI sys em o be
assessed, including he de elopmen p ocess, de ails
on he unde lying machine lea ning model and da a,
and he con ex o he applica ion in aud de ec ion.
Subsequen ly, we ask he pa icipan s o assess he
con o mi y o he AI sys em.
Pa 2. To in oduce ou so wa e a i ac , we demons a e
all unc ionali ies o he so wa e o he pa icipan s on
ou own sc een. Following ha , we sha e he so wa e
URL and le each pa icipan conduc he con o mi y
assessmen sepa a ely on he own lap op. The pa icipan
uploads he p edic ion model and da a, answe s he se
o ques ions, and ul ima ely explo es he ou come o he
assessmen on he in e ac i e dashboa d. Du ing he con-
o mi y assessmen , he pa icipan sha es he hough s
wi h us ollowing a hink-aloud p o ocol (Van Some en
e al., 1994).
Pa 3. We commence he g oup discussion once all
pa icipan s a e inished wi h explo ing ou so wa e.
Following he ecommenda ion o He ne e al. (2010)
and Abdel-Ka im e  al. (2023) o use open-ended
ques ions in ocus g oup sessions, we open he discussion
by asking pa icipan s open-ended ques ions on how
hey expe ienced he use o ou a i ac . We join ly
discuss he upsides and downsides o ou a i ac as
well as u u e oppo uni ies and isks. We hen conclude
he ocus g oup session. In sessions D and E, we send
a pos -e en su ey o ga he s uc u ed eedback on
ou me a- equi emen s, design p inciples, and hei
implemen a ion wi hin he so wa e.
Conclusion
Finally, in he conclusion s age, we consolida e he ind-
ings om he e alua ion and e ine ou a i ac based on
he newly de i ed design p inciples, hus s a ing he sec-
ond design cycle. We discuss he s eng hs o ou app oach
and limi a ions iden i ied du ing he ocus g oup sessions.
The eby, we posi ion ou wo k as a ounda ion o u u e
esea ch and p ac ical applica ion in he de elopmen o sys-
ema ic AI con o mi y assessmen amewo ks.
Designing AI con o mi y assessmen s
Awa eness o  hep oblem
In his phase, we build upon expe in e iews
complemen ed by insigh s om legal documen s and he
scien i ic li e a u e o iden i y bo h ac o s ha mo i a e
he in oduc ion o con o mi y assessmen s in p ac ice
and p oblems ha complica e he implemen a ion o
such assessmen s. Figu e2 p esen s he esul s o he
con en ional con en analysis. We u he p o ide ou
Table 2 O e iew o ocus
g oup sessions. No e ha all
pa icipan s held posi ions
wi hin he espec i e
o ganiza ions ha we e ele an
o he de elopmen o u iliza ion
o AI sys ems
Focus g oup Nmb . pa ici-
pan s
Indus y O ganiza ion size Loca ion o session
A 3 FinTech 10–49 employees On-si e
B 2 Financial consul ancy > 250 employees Remo e
C 4 Bank > 250 employees Remo e
D 5 Banking IT-Se ice > 250 employees Remo e
E 7 P o ide
Manu ac u e
> 250 employees Remo e
Elec onic Ma ke s (2025) 35:24 Page 15 o 24 24
Fo he subsequen e alua ion o ou a i ac , we ocus
on he use case o aud de ec ion in inance. AI o aud
de ec ion is well-sui ed o explo e aspec s o AI con o mi y
o se e al easons. Fi s , AI o aud de ec ion is well
es ablished (Abdallah e al., 2016) and many o ganiza ions
ely on i (among o he s, he ones we collabo a e wi h in
ou e alua ion). Second, he domain o aud de ec ion
is legally pe inen , as he ou comes o AI-d i en aud
de ec ion—whe e ansac ions may be au onomously
lagged and blocked—ha e a di ec impac on indi iduals’
abili y o execu e inancial ope a ions. As a consequence,
such scena ios a e likely o a ac egula o y sc u iny. Fo
ins ance, unde ce ain condi ions, he EU AI Ac could
designa e hese AI applica ions as “high- isk,” subjec ing
hem o mo e igo ous con o mi y s anda ds (Eu opean
Union, 2023). Finally, aud de ec ion algo i hms equen ly
inco po a e sensi i e a ibu es such as gende (Deepak
& Ab aham,2021) and ace he challenge o high-class
imbalance, wi h legi ima e ansac ions as ly ou numbe ing
audulen ones (Abdallah e al., 2016). This complexi y
heigh ens he po en ial o ai ness issues (Ba ocas &
Selbs , 2016), necessi a ing a ho ough e alua ion o ensu e
con o mi y.
Fo ou use case, we conside a hypo he ical company
ha uses AI o bina y classi ica ion, ha is, classi ies
ansac ions as ei he audulen o legi ima e using a
machine lea ning model. The model is ained on mo e
han 150,000 his o ical ansac ions spanning 12mon hs4
and elies on a g adien -boos ed o es (Chen & Gues in,
2016) o p edic aud. Simila o cases om he li e a u e
(e.g., Deepak & Ab aham, 2021; on Zahn e al., 2022), he
company suspec s sys ema ically dispa a e p edic ions o
ansac ions o men s. women, po en ially esul ing in legal
and epu a ional isks as ou lined p e iously. Mo eo e , he
i m le e ages ea u e-based explana ions o indi idual
p edic ions (an inc easingly adop ed app oach in he ield,
Bha e al., 2020) o be e unde s and and po en ially
mi iga e biases. This use case p esen s us wi h a angible
con ex o e alua e ou a i ac ’s e ec i eness.
While we choose he con ex o aud de ec ion o dem-
ons a ion pu poses, we no e ha ou a i ac does imple-
men he unc ionali y o add u he use cases and me ics
ia a sepa a e in e ace o audi o s (as p oposed by Fig.4).
In ac , we conside ed se e al o he use cases o he e alu-
a ion, such as hi ing and c edi sco ing. In hi ing, o exam-
ple, p ac i ione s o en le e age AI (Van den B oek e al.,
2021) wi h se e al documen ed iola ions o AI con o mi y
(see, e.g., Pa asu ama & Sedoc, 2022). In acco dance wi h
employmen and an i-disc imina ion laws (e.g., he Gene al
Equal T ea men Ac in Ge many), audi o s may de ine use
cases o hi ing. Simila ly, in c edi sco ing, AI equen ly
c ea es alue (Khandani e al., 2010) bu o en con lic s wi h
legal amewo ks (e.g., wi h he Equal C edi Oppo uni y
in he US, Smi h, 1977). The audi o s may de ine addi ional
use cases o c edi sco ing o add ess he unique equi e-
men s and laws applicable in his con ex .
Con o mi y me ics
Fo he exempla y use case o aud de ec ion, we
de e mine he se o sui able me ics o measu e ai ness,
explainabili y, and pe o mance ha is p esen ed in Table3.
Ou a i ac au oma ically e alua es hese me ics using he
p o ided da ase , model, and ques ionnai e answe s. The
audi o es ima es he ele ance o each me ic o a speci ic
use case sepa a ely. We b ie ly desc ibe hese me ics in
he ollowing.
Fai ness: s a is ical pa i y di e ence We include he me ic
co esponding o s a is ical pa i y (Dwo k e al., 2012) due
o i s high ele ance wi hin legal amewo ks (Ba ocas
& Selbs , 2016). We measu e he de ia ion o s a is ical
Table 3 O e iew o me ics
implemen ed in a i ac Fai ness S a is ical pa i y di e ence
Equalized odds di e ence
Dwo k e al., 2012
Ha d e al., 2016
Explainabili y Ques ionnai e on explainabili y
S abili y o global explana ions
Inspi ed by Mi chell e al.,
2019 and Geb u e al.,
2021
Inspi ed by Hsieh e al.,
2020
Pe o mance Accu acy
Balanced accu acy
P ecision
Recall
F1-sco e
Da a d i
Sokolo a e al., 2006
Sokolo a e al., 2006
Sokolo a e al., 2006
Sokolo a e al., 2006
Sokolo a e al., 2006
Inspi ed by Reis e al., 2016
4 The da a is a ailable a h ps:// www. kaggle. com/ da as e s/ de mi s i /
aud- ans ac io ns- da as e .

Elec onic Ma ke s (2025) 35:24 24 Page 16 o 24
pa i y ( ha is, bias) as he di e ence in he sha es o he
a o able p edic ion wi hin a g oup s. ou side his g oup.
Impo an ly, a g oup is ypically de ined based on an
a ibu e deemed sensi i e (c . Ba ocas e al., 2019). We
compu e he “s a is ical pa i y di e ence” in ou a i ac
ia
In ou exempla y use case, we conside gende as a sen-
si i e a ibu e and measu e he s a is ical pa i y di e ence
be ween women and he es o he cus ome s. No ably,
in e p e ing he alue esul ing om Eq.1 and de e mining
whe he o no i is accep able depends on he con ex . I
aspi ing o pe ec ai ness as de ined by s a is ical pa i y,
Eq.1 mus equal 0. I ollowing he so-called “80% ule,” o
example, which is common in many legal amewo ks (Ba o-
cas & Selbs , 2016; Feldman e al., 2015), ai ness would
only imply ha Eq.1 yields a alue g ea e han − 0.2. In
ou a i ac , audi ing expe s ha de ine use cases can o
cou se de ine he ange o alues conside ed accep able o
he s a is ical pa i y di e ence and, he eby, ailo i o he
gi en con ex .
Fai ness: equalized odds di e ence We include he me ic
co esponding o equalized odds (Ha d e al., 2016) which
shi s he emphasis om he dis ibu ion o he p edic ions
hemsel es o he p edic ion e o s. I is one o he mos
widely used me ics o algo i hmic ai ness and, in mos
cases, does no imply ai ness when s a is ical pa i y di e -
ence does (Ba ocas e al., 2019; Ga g e al., 2020), making i
a complemen a y and na u al choice o ou implemen a ion.
Wi h equalized odds, we measu e bias as he a e age
absolu e di e ence be ween e o a es om wi hin s. ou -
side a g oup. Simila o s a is ical pa i y, g oups a e ypically
de ined based on sensi i e a ibu es (Ha d e al., 2016). We
compu e he “equalized odds di e ence” in ou a i ac ia
Analogously, we conside gende as a sensi i e a ibu e
in ou exempla y use case. He e,
FPRing oup
e e s o he alse
posi i e a e o women,
FPRou g oup
e e s o he alse posi-
i e a e o he es o he cus ome s, and
FNRing oup
and
FNRou g oup
a e he coun e pa s wi h ega d o he alse neg-
a i e a e. Simila o s a is ical pa i y di e ence, he audi o
may manually se he ange o alues deemed accep able
o he equalized odds di e ence and ailo i o he con ex
a hand.
Impo an ly, in his s udy, we ocus on s a is ical pa i y
di e ence and equalized odds di e ence due o hei legal
ele ance, complemen a i y, and widesp ead ecogni ion.
Fu u e i e a ions o he a i ac will na u ally inco po a e
(1)
# a o able p edic ions in g oup
# all p edic ions in g oup
−
# a o able p edic ions ou side g oup
# all p edic ions ou side g oup .
(2)
1
2
[∣ FPRing oup −FPRou g oup ∣+∣FNRou g oup −FNRing oup ∣]
.
a b oad ange o ai ness me ics o enhance lexibili y and
add ess a a ie y o con ex s and egula o y needs.
Explainabili y: ques ionnai e Explainabili y ep esen s
a special c i e ion as i is ha dly easible o measu e he
explainabili y o AI sys ems solely based on quan i a i e
me ics. The e o e, inspi ed by exis ing wo ks aiming
o measu e he explainabili y o AI sys ems (see, e.g.,
Geb u e al., 2021; Mi chell e al., 2019), we de elop
a ques ionnai e ha he AI owne ills ou as pa o he
con o mi y assessmen . This ques ionnai e con ains
ques ions pe aining o he p edic ion model unde lying
he AI sys em (e.g., Do you communica e he echnical
limi a ions and po en ial isks o he AI sys em o use s,
such as i s le el o accu acy and/o e o a es?) and he
used explana ions o p edic ion model ou pu s (e.g., Does
you model p o ide any explana ions o p edic ions? I
yes, wha kind o explana ions?). The audi o de e mines
o each ques ion whe he i is ele an o a use case and
which answe o he ques ion ep esen s which colo in
he a ic ligh (i.e., g een, yellow, and ed). Ou ini ial
implemen a ion displays he mos c i ical colo -coded
esponse. Fo example, i he answe s o a ca ego y span
g een, yellow, and ed, he ca ego y would ul ima ely be
assigned a ed a ing o e lec he highes le el o conce n.
Explainabili y: s abili y o ea u e‑based explana ions He e,
s abili y e e s o he sensi i i y o ea u e-based explana-
ions owa ds small pe u ba ions in he aining da a. A low
s abili y sco e indica es ha small pe u ba ions in he ain-
ing da a se e ely impac he se o ea u es ha a e epo ed
o be highly impo an o he AI sys em, making he global
explana ions less eliable. Ou implemen a ion o his me ic
is inspi ed by Hsieh e al. (2020).
Pe o mance We measu e he p edic ion pe o mance based
on he mos common me ics o bina y classi ica ion. These
include accu acy, balanced accu acy, p ecision, ecall, and he
F1-sco e. Fo a de ailed explana ion o each lis ed pe o mance
me ic, we e e o he wo k o Sokolo a e al. (2006). Mo eo e ,
o accoun o ou in e iewed expe s’ conce n abou de e io a -
ing pe o mance o e ime, we implemen a me ic ha meas-
u es da a d i based on he Kolmogo o –Smi no es (Da ling,
1957; dos Reis e al., 2016). In sho , he Kolmogo o –Smi no
es e alua es whe he wo da a se ies a e d awn om he same
Elec onic Ma ke s (2025) 35:24 Page 17 o 24 24
p obabili y dis ibu ion; unde da a d i , he s a is ical p ope ies
o da a change o e ime, leading o a change in he unde lying
p obabili y dis ibu ion. Ou me ic builds upon his e alua ion
o show he p opo ion o a iables being a ec ed by da a d i .
E alua ion
We e alua e ou design amewo k and so wa e a i ac
ollowing he F amewo k o E alua ion in Design Science
Resea ch (Venable e al., 2012) and build upon ocus g oup
sessions o bo h alida e, e ine, and add o ou design
amewo k. No ably, he e alua ion is di ided in o wo dis inc
i e a ions. In he i s i e a ion, ocus g oups A, B, and C
assess he ini ial p o o ype om he i s design cycle, which
inco po a es DP 1 o DP 10. We use he eedback o hese
h ee sessions o o mula e wo addi ional design p inciples,
DP 11 and DP 12. In he second i e a ion, ocus g oups D and
E assess hese wo newly in eg a ed DP in he upda ed so wa e
a i ac . We de ail he esul s o bo h i e a ions in he ollowing.
E alua ion o  i s i e a ion
In he ini ial h ee ocus g oup sessions, ou a i ac ecei es
o e all posi i e eedback. Pa icipan s exp ess ha ou a i ac
p o ides a aluable app oach o conduc ing con o mi y
assessmen s, highligh ing i s e ec i eness in assessing ai ness
and explainabili y in AI sys ems. They app ecia e he use -
iendly and in e ac i e in e ace as well as he comp ehensi e
insigh s and suppo ing in o ma ion p o ided by he a i ac , as
exp essed by he ollowing commen : “[The ool is] easy, clean,
and nicely p esen ed. The e is no oo much in o ma ion o
cogni i e o e load” (Manage , ocus g oup B). The hink-aloud
p o ocols, cap u ing pa icipan s’ usage o he so wa e du ing he
sessions, and he subsequen g oup discussions u he con i m
he alidi y o he me a- equi emen s and design p inciples ha
we ha e in e ed based on ou awa eness o he p oblem. Fo
example, many pa icipan s pa icula ly app ecia e he in ui i e
na u e o he quali y labels (see DP 7) displayed h ough he
a ic ligh sys em, as i enables hem o quickly g asp he le el
o con o mi y o he AI sys em: “The s uc u e is good, clea a
he beginning, and he a ic ligh sys em immedia ely shows he
pe o mance in each a ea.” (Depa men head, ocus g oup C).
O e all, he posi i e ecep ion o he a i ac and he alignmen
be ween i s unc ionali y and he in ended objec i es p o ides
s ong e idence o he ca e ul conside a ion and success ul
implemen a ion o ou design choices.
Pa icipan s u he make aluable sugges ions ha
subs an ially con ibu e o he e inemen and enhancemen
o ou design amewo k. Many pa icipan s highly alue
he abundance o in o ma ion a ailable on he dashboa d bu
exp ess hei desi e o he inclusion o wha -i scena ios. They
emphasize ha ai ness and pe o mance a e in luenced by
pa ame e s chosen by he AI owne , and ha ing he abili y o
explo e he impac o hese choices on speci ic ai ness and
pe o mance me ics wi hin he dashboa d is highly bene icial.
“I would like o see how ai ness and pe o mance mo e when
I a y [AI sys em] pa ame e s like he classi ica ion h eshold.”
(Da a Scien is , ocus g oup A). Fo example, conside he
scena io whe e he AI owne iewing he dashboa d no ices
ha he AI sys em disp opo iona ely labels ansac ions made
by women as audulen , pu ing hem a a disad an age.
To add ess his, he AI owne may wish o in es iga e he
impac o adjus ing he classi ica ion h eshold speci ically
o women. By inco po a ing, e.g., a slide ha allows o
pa ame e adjus men s and obse ing he esul ing changes in
ai ness and pe o mance me ics, he a i ac could acili a e
he explo a ion o such hypo he ical scena ios. The e o e, we
p opose an addi ional design p inciple: The a i ac should
enable he explo a ion o wha -i scena ios (DP 11).
Pa icipan s in he ocus g oup sessions also sha e hei
hough s on he p esen a ion o esul s on he dashboa d. Ou
cu en app oach in ol es a a ic ligh sys em o communica e
esul s. Acco ding o eedback om pa icipan s, his sys em
is highly e ec i e o con eying in o ma ion o non- echnical
s akeholde s, including uppe managemen , egula o s, and
cus ome s. Howe e , de elope s u ilizing ou so wa e o
p oblem iden i ica ion and op imiza ion exp ess he need o
mo e de ailed quali y labels ha p o ide deepe insigh s. “The
a ic ligh s a e sui able o uppe managemen , bu hey a e
no op imal o de elope s.” (Senio Da a Scien is , ocus g oup
C). The pa icipan u he a gues ha o in e nal pu poses, he
a i ac should communica e he con o mi y esul s in a mo e
de ailed and ac ionable way, unde lining he impo ance o a
use-case-speci ic de ini ion o addi ional use oles (see DP 2).
Finally, pa icipan s in he ocus g oup sessions show
en husiasm owa d assessing he esul s on an in e ac i e
dashboa d and seem o app ecia e he abili y o explo e
di e en le els o g anula i y. Howe e , hey also oice hei
conce n ega ding he need o pe sis en esul s and p ope
documen a ion. In p ac ice, he so wa e would need o
gene a e a documen a ion epo summa izing all esul s in
a comp ehensi e way. “Fo alida ions o in e nal e isions,
[ he so wa e] should gene a e expo documen s ha a e
ampe -p oo […]” (Risk Manage , ocus g oup C). P ope
documen a ion o esul s ensu es he pe sis ence o con o mi y
esul s and, hus, enables o ganiza ions o ack hei p og ess
o e ime e ec i ely. This leads o ou las design p inciple:
The a i ac should ensu e pe sis en accessibili y o esul s
and p o ec agains manipula ion (DP 12).
E alua ion o second i e a ion
We in eg a e he new design p inciples DP 11 (A i ac should
enable in e ac i e explo a ion (wha -i scena ios)) and DP 12
(A i ac should ensu e pe sis en accessibili y o esul s (e.g.,
Elec onic Ma ke s (2025) 35:24 24 Page 18 o 24
ia a epo ) and p o ec agains manipula ion) in o ou ool.
To add ess DP 11, we in oduce a slide ha allows pa ici-
pan s o iew how pe o mance and ai ness me ics change
as hey adjus he classi ica ion h eshold, ha is, speci ying
he p obabili y a which an obse a ion is p edic ed o be o
he posi i e class. To add ess DP 12, we include a bu on
ha , when clicked, gene a es a PDF epo wi h a s uc u ed
o e iew o he esul s o he con o mi y assessmen , hus
ensu ing easily accessible and pe sis en documen a ion.
In he inal wo ocus g oup sessions, ou a i ac once
again ecei ed o e all posi i e eedback. In pa icula , he
newly in eg a ed slide (implemen ing DP 11) was ound
o be use ul o da a scien is s in e es ed in explo ing
di e en classi ica ion h esholds (Requi emen s Enginee ,
ocus g oup E). Ano he ea u e ha ecei ed pa icula ly
posi i e eedback is ou sepa a ion o he asks o he AI
owne and he audi o . A manage om ocus g oup E,
o example, s a ed: “[The sepa a ion o he wo use
oles] makes o al sense… Fo me, he i s impulse he e
is immedia ely he ou -eyes p inciple.” In iguingly, we
ecei e ambi alen eedback ega ding ou cen al use case
eposi o y, as illus a ed by he ollowing quo e: “E e y hing
ha is wi hin a company is ce ainly possible, e e y hing
ha is o e a ching is mo e di icul ” (Manage , ocus g oup
E). On he one hand, p ac i ione s seem eluc an o sha e
de ails on con o mi y assessmen s wi h compe i o s, e en
i hey do no necessa ily include c i ical in o ma ion o
model speci ics. On he o he hand, some expe s ag ee ha
collabo a ing wi h o he o ganiza ions wi hin an indus y
may be necessa y in some con ex s: “I depends on he use
case. When i comes o egula ions ha e e yone has o
comply wi h [ he use case eposi o y] makes sense o be
able o compa e hem” (Da a Scien is , ocus g oup E). S ill,
he eluc ance o sha e ele an in o ma ion wi h compe i o s
may become a c ucial ba ie o de eloping bes p ac ices
and, ul ima ely, es ablishing assessmen s anda ds.
In ou pos -e en su ey on he design amewo k and i s
implemen a ion in he so wa e a i ac , pa icipan s s ongly
ag eed on he necessi y o he me a- equi emen s and design
p inciples and ag eed, hough o a sligh ly lesse ex en , ha
he p o o ype e ec i ely implemen s hese p inciples. We
p esen an o e iew o he alida ed design p inciples and
su ey esul s in Appendix C.
Discussion
In e p e a ion o key indings
Ou analysis o li e a u e, legal documen s, and expe
in e iews sheds ligh on he p essing need o o ganiza ions
o conduc AI con o mi y assessmen s, highligh ing bo h
hei necessi y and inhe en complexi y. In e es ingly,
ou in e iewed expe s emphasize ha he mo i a ion
o o ganiza ions o pu sue con o mi y assessmen s may
go beyond egula o y compliance and also s em om he
ecogni ion ha sel -commi men o AI con o mi y can o e a
compe i i e ad an age. This sel -imposed commi men can be
iewed as one mani es a ion o co po a e digi al esponsibili y
(CDR), whe e companies emb ace esponsible p ac ices
ela ed o digi al p oduc s, se ices, and echnologies (Mihale-
Wilson e al., 2021). Resea ch on CDR also suppo s he
insigh s p o ided by he in e iewed expe s, highligh ing he
signi icance o CDR ac i i ies in shaping consume pe cep ion.
This in luence can ha e a di ec impac on consume s’
opinions, consump ion decisions, and choices o adop ion
(Ca l e al., 2024; Sch eck & Rai hel, 2018), and ul ima ely
p omo e a compe i i e edge in he ma ke . To explo e whe he
conduc ing AI con o mi y assessmen s indeed imp o es
companies’ compe i i e ad an age, u he esea ch is needed
es ing ou design amewo k in an o ganiza ional se ing.
Ano he ema k om he expe in e iews ha is wo h
discussing conce ns democ a iza ion. Speci ically, so wa e
wi h a g aphical use in e ace may imp o e he accessibili y
o AI con o mi y assessmen s o non- echnical s akehold-
e s, po en ially shi ing he con ol and owne ship o he
assessmen s om a limi ed numbe o expe s o a la ge
g oup o people. In line wi h p e ious esea ch on democ-
a iza ion in he ield o in o ma ion sys ems (see, e.g.,
Awas hi & Geo ge, 2020; Zacha ias e al., 2022), his shi
in con ol and owne ship can subs an ially bene i o ganiza-
ions by pu ing mo e employees in he posi ion o ac i e
con ibu o s o da a-d i en solu ions. Mo eo e , democ a i-
za ion holds he po en ial o os e an o ganiza ional cul u e
ha p omo es in o ma ion sha ing and emb aces di e se
pe spec i es as well as o ganiza ional agili y (Hyun e al.,
2020). No ably, d awing he igh conclusions p io esea ch
has highligh ed ha a lack o echnical expe ise may gene -
ally limi b oade pa icipa ion in AI (Bi hane e al., 2022),
sugges ing he need o cau iously assess whe he g aphi-
cal use in e aces can genuinely enable all s akeholde s o
e ec i ely conduc con o mi y assessmen s.
The esul s o ou ocus g oup sessions bo h con i m he
app op ia eness o ou a i ac and design amewo k o
conduc ing con o mi y assessmen s and e eal he po en ial
o u he imp o emen . Pa icipan s emphasize ha he
a i ac holds he capabili y o p omo e con o mi y and, mo e
b oadly, e hics in AI sys ems. Howe e , hey emphasize he
need o e sa ili y in i s p ocedu es and p esen a ions. On
he one hand, (uppe ) managemen seeks in ui i e quali y
labels o acili a e decision-making ega ding he deploy-
men o AI sys ems in o p oduc ion, which is in line wi h
esea ch on manage ial decision-making (c. . Cla k J e al.,
2007). On he o he hand, de elope s need o con inuously
assess con o mi y and e hical impac h oughou he de el-
opmen p ocess o adhe e o “e hics by design” s anda ds
Elec onic Ma ke s (2025) 35:24 Page 19 o 24 24
(Ipho en & K i ikos, 2021; Kieslich e al., 2022). In his
ega d, a so wa e solu ion eme ges as highly sui able, o e -
ing comp ehensi e in o ma ion and adap able suppo ai-
lo ed o accommoda e hese di e se needs.
In e es ingly, pa icipan s exp essed cu iosi y abou he
a ailabili y o in o ma ion explo ed wi hin he in e ac i e
dashboa d in a du able, accessible, and non-manipula able
documen . Such a documen would hold alue in jus i ying he
use o AI sys ems and e ec i ely communica ing con o mi y o
ex e nal s akeholde s. This obse a ion aligns wi h he g owing
demand o AI ce i ica ion om esea che s (Cihon e al.,
2021; Ma us & Veale, 2022) and policymake s (e.g., Eu opean
Union, 2023, p. 41), who ad oca e o i s de elopmen . In
his ega d, ou indings can se e as a sp ingboa d o he
ad ancemen o AI ce i ica ion e o s, p o iding aluable
insigh s o guide i s de elopmen and implemen a ion.
Con ibu ion op ac ice
Ou wo k makes se e al con ibu ions o p ac ice. Fi s ,
we p opose a design amewo k and a so wa e a i ac
o AI con o mi y assessmen s, which we p o ide online.5
O ganiza ions aiming o p epa e o upcoming egula ions,
such as he AI Ac , may adop and build upon ou a i ac o
conduc con o mi y assessmen s and demons a e compliance
wi h hese egula ions. Second, ou a i ac o e s a s uc u ed
app oach o assess AI con o mi y, enabling o ganiza ions o
s eamline he es ablishmen o an AI con o mi y pipeline
wi h p ede ined, scien i ically de i ed e alua ion me ics. This
educes he e o equi ed du ing pipeline de elopmen and
allows o ganiza ions o gain an o e iew o ele an me ics
and pa ame e s. Thi d, by au oma ing much o he echnical
analysis, ou ool educes eliance on skilled expe s and
imp o es accessibili y o AI con o mi y assessmen s, which in
u n may educe cos s. Fou h, he ool acili a es collabo a ion
among a ious s akeholde s wi hin o ganiza ions, such as AI
de elope s, p oduc owne s, and senio managemen , and may
hus os e a sha ed unde s anding o AI con o mi y challenges
and oppo uni ies. Fi h, he quali a i e e idence ga he ed om
expe in e iews p o ides o ganiza ions wi h a s uc u ed
o e iew o bo h he necessi y and complexi y o AI con o mi y
assessmen s. This o e iew may help pe suade managemen
o he impo ance o add essing AI con o mi y and ake
p oac i e measu es o mi iga e an icipa ed implemen a ion
challenges (e.g., p epa ing a da a pipeline o con inuous
moni o ing o con o mi y me ics). Las ly, he a i ac may
e ol e in o a pla o m o sha ing well-de ined use cases ac oss
o ganiza ions in he same indus y, po en ially inco po a ing a
ecommenda ion sys em o guide o ganiza ions and p omo e
indus y-wide s anda ds e ec i ely.
Con ibu ion o heo y
Ou s udy makes impo an con ibu ions o heo y. Fi s
and o emos , ou s udy con ibu es p esc ip i e design
knowledge h ough he p oposed design amewo k (Fig.3)
con aining me a- equi emen s and design p inciples. This
design knowledge se es as a aluable basis o esea che s
o ad ance he unde s anding o so wa e o con o mi y
assessmen s and AI audi s in a b oade sense. Acco ding
o Baske ille e al. (2018), such no el design knowledge
cons i u es he main heo e ical con ibu ion o design
science esea ch.
Ou wo k u he con ibu es o he li e a u e on AI
go e nance. E ec i e go e nance o AI sys ems demands
obus ools ha ensu e compliance, p omo e e hical
s anda ds, and enhance accoun abili y ac oss applica ions
(Ab aham e al., 2019; Bi ks ed e al., 2023; Schneide e al.,
2023). Resea che s ha e equen ly called o a “collabo a i e
go e nance” o AI sys ems (see, e.g., Bi ks ed e al., 2023)
which ex ends he ocus beyond indi idual o ganiza ions
owa ds collabo a i e ne wo ks wi h in e nal and ex e nal
s akeholde s, each wi h dis inc oles and esponsibili ies.
Ou design amewo k can be iewed as a s ep owa ds
collabo a i e go e nance: i assesses AI sys ems agains
di e en socie al and business- ele an c i e ia while
conside ing dis inc oles, such as AI owne s and audi o s,
and acili a ing collabo a ion among mul iple o ganiza ions
h ough sha ing use cases and benchma king AI sys ems
agains eme ging s anda ds.
Ou wo k also makes a con ibu ion o he li e a u e
on s anda diza ion o in o ma ion and communica ion
echnology (see, e.g., Cos abile e al., 2022; Hanse h &
Bygs ad, 2015; Lyy inen & King, 2006). A cen al ene
o his li e a u e is ha s anda diza ion plays a i al ole
in managing echnologies, including AI sys ems, by
ensu ing compa abili y and in e ope abili y. This aligns
wi h insigh s om ou expe in e iews, which unde line
he necessi y o s anda dizing AI con o mi y assessmen s
o acili a e consis en and eliable e alua ions ac oss
o ganiza ions wi hin an indus y. By con as , indings
om ou ocus g oup sessions e eal a p ac ical p oblem:
while p ac i ione s acknowledge he impo ance o
s anda diza ion, hey a e eluc an o sha e audi - ela ed
in o ma ion wi h compe i o s, e en when i excludes
sensi i e cus ome da a o speci ics abou hei AI sys ems.
Ye , sha ing insigh s and expe iences wi h pas con o mi y
assessmen con igu a ions wi h o he o ganiza ions is
essen ial o he eme gence o common bes p ac ices and,
ul ima ely, indus y s anda ds. Ou wo k highligh s he
impo ance o add essing his ension and demons a es
ha inno a i e app oaches a e needed o p omo e
s anda diza ion while espec ing he conce ns o indi idual
o ganiza ions.
5 The so wa e a i ac and an exempla y use case a e a ailable a
h ps:// gi hub. com/ mlowin/ con o mi y_ asses smen .
Elec onic Ma ke s (2025) 35:24 24 Page 20 o 24
T ans e abili y and u u e esea ch oppo uni ies
As ou lined p e iously, ocusing on AI o bina y classi ica-
ion as well as ce ain aspec s o AI con o mi y was essen ial
o de eloping a angible, unc ional a i ac . Ou choice o
ai ness and explainabili y s ems om bo h aspec s being
no only highly ele an in cu en egula o y discussions bu
also ep esen ing he co e challenges ha we iden i ied om
he expe in e iews in he p oblem awa eness phase (e.g.,
con ex -dependen equi emen s and a lack o s anda diza-
ion). Howe e , a key ques ion emains: Can he indings
based on hese wo aspec s o AI con o mi y and he use
case o bina y classi ica ion be ans e ed o AI con o mi y
in di e en con ex s?
We a gue ha mos indings om ou s udy a e indeed
ans e able o o he con ex s. Fo ins ance, he eedback
om p ac i ione s on he u ili y o a so wa e-suppo ed
assessmen , i s usabili y, and ea u es, such as he “d ill-
down” app oach on speci ic me ics and he a ic ligh sys-
em o simpli ying communica ion, a e no ied exclusi ely
o ai ness and explainabili y no o bina y classi ica ion.
These design ea u es acili a e he sha ing o expe s and-
a ds and in e p e a ions, enabling hem o di use quickly
ac oss indus ies. This sugges s ha so wa e ools such as
ou a i ac could suppo con o mi y assessmen s in o he
domains, independen o he speci ic aspec s unde e alu-
a ion. Ano he no able poin is ha ai ness and explain-
abili y sha e simila cha ac e is ics and challenges wi h
o he aspec s o AI con o mi y, such as p i acy, AI sa e y,
and cybe secu i y. These aspec s, like ai ness and explain-
abili y, a e ma ked by mul iple, o en compe ing de ini ions
and me ics. Fo ins ance, p i acy can ha e widely di e -
en in e p e a ions depending on he con ex (Chua e al.,
2021), and bes p ac ices o AI sa e y and cybe secu i y
a e con inually e ol ing (Laza & Nelson, 2023). The e-
o e, ou insigh s on so wa e-based assessmen s—designed
o acili a e he sha ing o s anda ds and enable semi-au o-
ma ed e alua ions—may also p o e aluable in hese a eas.
Wha e e addi ional c i e ia a e inco po a ed in he u u e
will likely ei he be based on quan i iable indica o s, such
as hose e lec ing ai ness, o su ey-based measu es, as
demons a ed wi h explainabili y, bo h o which we e suc-
cess ully implemen ed and es ed du ing ou ocus g oup
sessions.
Wi h ega d o ex ending ou indings beyond bina y clas-
si ica ion o asks such as eg ession o con en gene a ion
(e.g., h ough gene a i e AI sys ems, see Feue iegel e al.,
2024), we adop a mo e nuanced pe spec i e. Ou indings
can a guably be ans e ed o eg ession asks, as eg ession
is no undamen ally di e en om classi ica ion. This is
e iden in he applicabili y o simila g oup-le el ai ness
measu es (c . Ba ocas e al., 2019) and compa able me hods
o explainabili y (e.g. Lundbe g & Lee, 2017). Howe e ,
applying ou indings o gene a i e AI poses mo e complex
challenges. While he undamen al issues emain simila ,
such as e idence o bias in la ge language modelsand in
image gene a ion (Ananya, 2024) and ongoing us issues
ela ed o opaci y (Wang e al., 2023) he na u e o gen-
e a i e sys ems in oduces addi ional laye s o complexi y.
While he gene al app oach o AI con o mi y p oposed in
his pape could se e as a aluable s a ing poin , ou so -
wa e a i ac would equi e signi ican ex ension o accom-
moda e he unique cha ac e is ics o gene a i e AI. Fu he
esea ch is pa icula ly c ucial gi en he widesp ead adop-
ion o sys ems based on gene a i e AI, such asCha GPT,
DeepSeek, and MidJou ney, which ampli ies he socie al
impac and calls o ailo ed solu ions o add ess he dis inc
challenges in his domain.
O he con ex ual ac o s could also in luence he ans-
e abili y o ou indings. Fo ins ance, ou expe sample,
while ope a ing in e na ionally, is en i ely based in Ge many
and Swi ze land. This geog aphic concen a ion migh unde
ce ain ci cums ances a ec he applicabili y o ou esul s
o o he egions wi h di e en egula o y en i onmen s o
indus y p ac ices. While we belie e ou indings emain
alid and make a s ong case o “po able p inciples”
(Magnani & Gioia, 2023) ha can be ans e ed o di e -
en con ex s, u u e esea ch should include b oade es ing
ac oss di e se geog aphic and egula o y con ex s o u he
subs an ia e his claim. Simila ly, ou a i ac was e alua ed
h ough ocus g oup sessions using a clickable p o o ype in a
con olled, a i icial se ing. In eal-wo ld applica ions, in e-
g a ing he a i ac in o an o ganiza ion’s IT in as uc u e
could signi ican ly in luence use in e ac ion and adop ion.
To alida e i s p ac ical applicabili y, u u e esea ch should
implemen ou a i ac (o a compa able sys em) wi hin
o ganiza ional se ings. This would allow o an in es iga-
ion o i s long- e m impac on he socio echnical en i on-
men , including i s e ec i eness, usabili y, and in luence on
o ganiza ional p ocesses and decision-making.
Conclusion
In his pape , we add ess he p essing need o o ganiza ions
o assess he con o mi y o hei AI sys ems. By le e aging
design science esea ch, we ha e de eloped a design
amewo k and so wa e a i ac ha se es as a ool o
semi-au oma ed AI con o mi y assessmen s, enabling
e ec i e communica ion be ween AI owne s and
s akeholde s such as egula o s. As we look o he u u e, we
en ision u he ad ancemen s in ou amewo k, ex ending
i s scope om assessmen s o ce i ica ion. These e o s will
con ibu e o adhe ing o egula ions on AI sys ems, such as
he EU AI Ac , and empowe o ganiza ions o na iga e he
e ol ing landscape o AI con o mi y wi h con idence.

Elec onic Ma ke s (2025) 35:24 Page 21 o 24 24
Supplemen a y In o ma ion The online e sion con ains supplemen a y
ma e ial a ailable a h ps:// doi. o g/ 10. 1007/ s12525- 025- 00770-2.
Funding Open Access unding enabled and o ganized by P ojek
DEAL. S a e Hesse,Ge many: LOEWE 3, p ojec "T ans AI ".
Open Access This a icle is licensed unde a C ea i e Commons A i-
bu ion 4.0 In e na ional License, which pe mi s use, sha ing, adap a-
ion, dis ibu ion and ep oduc ion in any medium o o ma , as long
as you gi e app op ia e c edi o he o iginal au ho (s) and he sou ce,
p o ide a link o he C ea i e Commons licence, and indica e i changes
we e made. The images o o he hi d pa y ma e ial in his a icle a e
included in he a icle’s C ea i e Commons licence, unless indica ed
o he wise in a c edi line o he ma e ial. I ma e ial is no included in
he a icle’s C ea i e Commons licence and you in ended use is no
pe mi ed by s a u o y egula ion o exceeds he pe mi ed use, you will
need o ob ain pe mission di ec ly om he copy igh holde . To iew a
copy o his licence, isi h p://c ea i ecommons.o g/licenses/by/4.0/.
Re e ences
117 h Cong ess. (2022). Algo i hmic accoun abili y ac . h ps:// www.
cong ess. go / bill/ 117 h- cong ess/ house- bill/ 6580/ e x . Accessed
02 Jan 2024.
Abbasi, A., Pa sons, J., Pan , G., Sheng, O. R. L., & Sa ke , S. (2024).
Pa hways o design esea ch on a i icial in elligence. In o ma-
ion Sys ems Resea ch, 35(2), 441–459.h ps:// doi. o g/ 10. 1287/
is e. 2024. edi o ial. 35. n2
Abdallah, A., Maa o , M. A., & Zainal, A. (2016). F aud de ec ion sys-
em: A su ey. Jou nal o Ne wo k and Compu e Applica ions,
68, 90–113.h ps:// doi. o g/ 10. 1016/j. jnca. 2016. 04. 007
Abdel-Ka im, B. M., P eu e , N., Ca l, K. V., & Hinz, O. (2023).
How AI-based sys ems can induce e lec ions: The case o
AI-augmen ed diagnos ic wo k. MIS Qua e ly, 47(4), 1395–
1424.h ps:// doi. o g/ 10. 25300/ MISQ/ 2022/ 16773
Ab aham, R., Schneide , J., & Vom B ocke, J. (2019). Da a go e n-
ance: A concep ual amewo k, s uc u ed e iew, and esea ch
agenda. In e na ional Jou nal o In o ma ion Managemen , 49,
424–438.h ps:// doi. o g/ 10. 1016/j. ijin omg . 2019. 07. 008
Ananya,. (2024). AI image gene a o s o en gi e acis and sexis
esul s: Can hey be ixed? Na u e News Fea u e, 627, 722–
725.h ps:// doi. o g/ 10. 1038/ d41586- 024- 00674-9
Angwin, J., La son, J., Ma u, S., & Ki chne , L. (2016). Machine bias. In
E hics o da a and analy ics(pp. 254–264). Aue bach Publica ions.
A in, S., Bel ield, H., B undage, M., K uege , G., Wang, J., Welle ,
A., Ande ljung, M., K awczuk, I., K uege , D., Lebensold,
J., Maha aj, T., & Zilbe man, N. (2021). Filling gaps in
us wo hy de elopmen o AI. Science, 374(6573), 1327–
1329.h ps:// doi. o g/ 10. 1126/ scien ce. abi71 76
Awas hi, P., & Geo ge, J. J. (2020). A case o da a democ a iza ion.
In AMCIS. h ps:// aisel. aisne . o g/ amcis 2020/ da a_ scien ce_
analy ics_ o _ decis ion_ suppo / da a_ scien ce_ analy ics_ o _
decis ion_ suppo / 23
Ba ocas, S., Ha d , M., & Na ayanan, A. (2019). Fai ness and machine
lea ning: Limi a ions and oppo uni ies. ai mlbook.o g. h p://
www. ai m lbook. o g
Ba ocas, S., & Selbs , A. D. (2016). Big da a’s dispa a e impac . Cali o -
nia Law Re iew671–732.h ps:// doi. o g/ 10. 2139/ ss n. 24778 99
Baske ille, R., Baiye e, A., G ego , S., He ne , A., & Rossi, M.
(2018). Design science esea ch con ibu ions: Finding a balance
be ween a i ac and heo y. Jou nal o he Associa ion o In o -
ma ion Sys ems, 19(5), 3.h ps:// aisel. aisne . o g/ jais/ ol19/ iss5/3
Baue , K., on Zahn, M., & Hinz, O. (2023). Expl(AI)ned: The
impac o explainable a i icial in elligence on use s’ in o ma-
ion p ocessing. In o ma ion Sys ems Resea ch, 34(4), 1582–
1602.h ps:// doi. o g/ 10. 1287/ is e. 2023. 1199
Bellamy, R. K., Dey, K., Hind, M., Ho man, S. C., Houde, S., Kan-
nan, K., Lohia, P., Ma ino, J., Meh a, S., Mojsilo ic, A., Naga ,
S., Na esan Ramamu hy, K., Richa ds, J., Saha, D., Sa ige i,
P., Singh, M., Va shney K. R., & Zhang, Y. (2019). Ai ai ness
360: An ex ensible oolki o de ec ing and mi iga ing algo i h-
mic bias. IBM Jou nal o Resea ch and De elopmen , 63(4/5),
4–1.h ps:// doi. o g/ 10. 1147/ JRD. 2019. 29422 87
Benbya, H., Da enpo , T. H., & Pachidi, S. (2020). A i icial in elligence
in o ganiza ions: Cu en s a e and u u e oppo uni ies. MIS
Qua e ly Execu i e, 19(4).h ps:// doi. o g/ 10. 2139/ ss n. 37419 83
Bha , U., Xiang, A., Sha ma, S., Welle , A., Taly, A., Jia, Y., Ghosh,
J., Pu i, R., Mou a, J. M., & Ecke sley, P. (2020). Explainable
machine lea ning in deploymen . In Con e ence on Fai ness,
Accoun abili y, and T anspa ency (FAT)(pp. 648–657).h ps://
doi. o g/ 10. 1145/ 33510 95. 33756 24
Bi d, S., Dudík, M., Edga , R., Ho n, B., Lu z, R., Milan, V., Sameki,
M., Wallach, H., & Walke , K. (2020). Fai lea n: A oolki o
assessing and imp o ing ai ness in AI. Mic oso , Tech. Rep.
MSR-TR-2020–32. h ps:// www. mic o so . com/ en- us/ esea
ch/ publi ca ion/ ai l ea n-a- oolk i - o - asses sing- and- imp o
ing- ai n ess- in- ai/
Bi hane, A., Isaac, W., P abhaka an, V., Diaz, M., Elish, M. C., Gab iel,
I., & Mohamed, S. (2022). Powe o he people? Oppo uni ies
and challenges o pa icipa o y AI. EAAMO ’22. Associa ion o
Compu ing Machine y.h ps:// doi. o g/ 10. 1145/ 35516 24. 35552 90
Bi ks ed , T., Minkkinen, M., Tandon, A., & Män ymäki, M. (2023).
AI go e nance: Themes, knowledge gaps and u u e agendas.
In e ne Resea ch, 33(7), 133–167.h ps:// doi. o g/ 10. 1108/
INTR- 01- 2022- 0042
B ingas Colmena ejo, A., Nannini, L., Riege , A., Sco , K. M.,
Zhao, X., Pa o, G. K., Kasneci, G., & Kinde -Ku landa, K.
(2022). Fai ness in ag eemen wi h Eu opean alues: An in e -
disciplina y pe spec i e on AI egula ion. In AAAI/ACM Con-
e ence on AI, E hics, and Socie y (AIES).h ps:// doi. o g/ 10.
1145/ 35140 94. 35341 58
B own, S., Da ido ic, J., & Hasan, A. (2021). The algo i hm audi :
Sco ing he algo i hms ha sco e us. Big Da a & Socie y,
8(1).h ps:// doi. o g/ 10. 1177/ 20539 51720 983865
Bu ns, J., & Foga y, J. (2010). App oaches o audi ing s anda ds
and hei possible impac on audi o beha io . In e na ional
Jou nal o Disclosu e and Go e nance, 7(4), 310–319.h ps://
doi. o g/ 10. 1057/ jdg. 2010. 21
Ca l, K. V., Mihale-Wilson, C., Zibuschka, J., & Hinz, O. (2024).
A consume pe spec i e on co po a e digi al esponsibili y:
An empi ical e alua ion o consume p e e ences. Jou nal o
Business Economics94, 979–1024. h ps:// doi. o g/ 10. 1007/
s11573- 023- 01142-y
Cho, M. K. (2021). Rising o he challenge o bias in heal h ca e AI.
Na u e Medicine, 27(12), 2079–2081.h ps:// doi. o g/ 10. 1038/
s41591- 021- 01577-2
Chand a, L., Seidel, S., & G ego , S. (2015). P esc ip i e knowledge
in IS esea ch: Concep ualizing design p inciples in e ms o
ma e iali y, ac ion, and bounda y condi ions. In Hawaii in e -
na ional con e ence on sys em sciences (pp. 4039–4048).
ieee.h ps:// doi. o g/ 10. 1109/ HICSS. 2015. 485
Chen, T., & Gues in, C. (2016). Xgboos : A scalable ee boos ing sys em.
In Acm sigkdd in e na ional con e ence on knowledge disco e y and
da a mining.h ps:// doi. o g/ 10. 1145/ 29396 72. 29397 85
Chua, H. N., Ooi, J. S., & He bland, A. (2021). The e ec s o di e -
en pe sonal da a ca ego ies on in o ma ion p i acy conce n
and disclosu e. Compu e s & Secu i y, 110, 102453.h ps://
doi. o g/ 10. 1016/j. cose. 2021. 102453
Elec onic Ma ke s (2025) 35:24 24 Page 22 o 24
Chui, M., Hall, B., Mayhew, H., Singla, A., & Sukha e sky,
A. (2022). The s a e o AI in 2022—and a hal dec-
ade in e iew. McKinsey & Company, 6. h ps:// www.
mckin sey. com/ capab ili i es/ quan umbla ck/ ou - insig h s/
he- s a e- o - ai- in- 2022- and-a- hal - decade- in- e iew
Cihon, P., Kleinal enkamp, M. J., Schue , J., & Baum, S. D. (2021).
Ai ce i ica ion: Ad ancing e hical p ac ice by educing in o -
ma ion asymme ies. IEEE T ansac ions on Technology and
Socie y, 2(4), 200–209.h ps:// doi. o g/ 10. 1109/ TTS. 2021.
30775 95
Cla k J , T. D., Jones, M. C., & A ms ong, C. P. (2007). The dynamic
s uc u e o managemen suppo sys ems: Theo y de elopmen ,
esea ch ocus, and di ec ion. MIS Qua e ly, 31(3), 579–
615.h ps:// doi. o g/ 10. 2307/ 25148 808
Co be -Da ies, S., Pie son, E., Felle , A., Goel, S., & Huq, A. (2017).
Algo i hmic decision making and he cos o ai ness. In Acm
sigkdd in e na ional con e ence on knowledge disco e y and da a
mining.h ps:// doi. o g/ 10. 1145/ 30979 83. 30980 95
Cos abile, C., Iden, J., & Bygs ad, B. (2022). Building digi al pla -
o m ecosys ems h ough s anda diza ion: An ins i u ional wo k
app oach. Elec onic Ma ke s, 32(4), 1877–1889.h ps:// doi. o g/
10. 1007/ s12525- 022- 00552-0
Coussemen , K., & Benoi , D. F. (2021). In e p e able da a science o
decision making. Decision Suppo Sys ems, 150, 113664.h ps://
doi. o g/ 10. 1016/j. dss. 2021. 113664
Da ling, D. A. (1957). The Kolmogo o -Smi no , C ame - on Mises
es s. The Annals o Ma hema ical S a is ics, 28(4), 823–
838.h ps:// www. js o . o g/ s able/ 22370 48
Deepak, P., & Ab aham, S. S. (2021). Fai lo : Fai ness in ou lie de ec-
ion. Da a Science and Enginee ing, 6(4), 485–499.h ps:// doi.
o g/ 10. 1007/ s41019- 021- 00169-x
Dola a, M., Feue iegel, S., & Schwabe, G. (2022). A socio echnical
iew o algo i hmic ai ness. In o ma ion Sys ems Jou nal, 32(4),
754–818.h ps:// doi. o g/ 10. 1111/ isj. 12370
dos Reis, D. M., Flach, P., Ma win, S., & Ba is a, G. (2016). Fas unsu-
pe ised online d i de ec ion using inc emen al Kolmogo o -
Dmi no es . In Acm sigkdd in e na ional con e ence on knowl-
edge disco e y and da a mining.h ps:// doi. o g/ 10. 1145/ 29396
72. 29398 36
Dwi edi, R., Da e, D., Naik, H., Singhal, S., Ome , R., Pa el, P., Qian,
B., Wen, Z., Shah, T., Mo gan, G., & Ranjan, R. (2023). Explain-
able AI (XAI): Co e ideas, echniques, and solu ions. ACM Com-
pu ing Su eys, 55(9), 1–33. h ps:// doi. o g/ 10. 1145/ 35610 48
Dwo k, C., Ha d , M., Pi assi, T., Reingold, O., & Zemel, R. (2012).
Fai ness h ough awa eness. In Inno a ions in heo e ical compu e
science con e ence.h ps:// doi. o g/ 10. 1145/ 20902 36. 20902 55
Eu opean Union. (2023). AI ac . h ps:// a i i cial in el ligen ceac . eu/ he-
ac /. Accessed 08 Oc 2024.
Feldman, M., F iedle , S. A., Moelle , J., Scheidegge , C., & Ven-
ka asub amanian, S. (2015). Ce i ying and emo ing dispa a e
impac . In Acm sigkdd in e na ional con e ence on knowledge
disco e y and da a mining.h ps:// doi. o g/ 10. 1145/ 27832 58.
27833 11
Feue iegel, S., Dola a, M., & Schwabe, G. (2020). Fai AI: Challenges
and oppo uni ies. Business & In o ma ion Sys ems Enginee ing,
62, 379–384.h ps:// doi. o g/ 10. 1007/ s12599- 020- 00650-3
Feue iegel, S., Ha mann, J., Janiesch, C., & Zschech, P. (2024). Gen-
e a i e AI. Business & In o ma ion Sys ems Enginee ing, 66(1),
111–126.h ps:// doi. o g/ 10. 1007/ s12599- 023- 00834-7
Fleming, N. (2018). How a i icial in elligence is changing d ug
disco e y. Na u e, 557, S55–S57. h ps:// doi. o g/ 10. 1038/
d41586- 018- 05267-x
Flo idi, L., Holweg, M., Taddeo, M., Amaya, J., Mökande , J., & Wen,
Y. (2022). CapAI-A p ocedu e o conduc ing con o mi y assess-
men o AI sys ems in line wi h he EU A i icial In elligence
Ac . SSRN 4064091.h ps:// doi. o g/ 10. 2139/ ss n. 40640 91
F esz, B., Göbels, V. P., Om i, S., B ajo ic, D., Aichele, A., Ku z, J.,
Neuhü le , J., & Hube , M. F. (2024). The con ibu ion o xai
o he sa e de elopmen and ce i ica ion o AI: An expe -based
analysis.h ps:// doi. o g/ 10. 48550/ a Xi . 2408. 02379
Fu, R., Huang, Y., & Singh, P. V. (2021). C owds, lending, machine,
and bias. In o ma ion Sys ems Resea ch, 32(1), 72–92.h ps://
doi. o g/ 10. 1287/ is e. 2020. 0990
Ga g, P., Villaseno , J., & Foggo, V. (2020). Fai ness me ics: A com-
pa a i e analysis. InIEEE in e na ional con e ence on big da a
(Big Da a) (pp. 3662–3666).h ps:// doi. o g/ 10. 1109/ BigDa a500
22. 2020. 93780 25
Geb u, T., Mo gens e n, J., Vecchione, B., Vaughan, J. W., Wallach, H., Iii,
H. D., & C aw o d, K. (2021). Da ashee s o da ase s. Communica-
ions o he ACM, 64(12), 86–92.h ps:// doi. o g/ 10. 1145/ 34587 23
Gioia, D. A., Co ley, K. G., & Hamil on, A. L. (2013). Seeking quali a-
i e igo in induc i e esea ch: No es on he Gioia me hodology.
O ganiza ional Resea ch Me hods, 16(1), 15–31.h ps:// doi. o g/
10. 1177/ 10944 28112 452151
Gnewuch, U., Mo ana, S., & Maedche, A. (2017). Towa ds designing
coope a i e and social con e sa ional agen s o cus ome se ice.
In ICIS.h ps:// aisel. aisne . o g/ icis2 017/ HCI/ P ese n a i ons/1
Hanse h, O., & Bygs ad, B. (2015). Flexible gene i ica ion: ICT s and-
a diza ion s a egies and se ice inno a ion in heal h ca e. Eu o-
pean Jou nal o In o ma ion Sys ems, 24(6), 645–663.h ps://
doi. o g/ 10. 1057/ ejis. 2015.1
Ha d , M., P ice, E., & S eb o, N. (2016). Equali y o oppo uni y
in supe ised lea ning. In: Ad ances in neu al in o ma ion p o-
cessing sys ems (NIPS)(Vol. 29).h ps:// doi. o g/ 10. 48550/ a Xi .
1610. 02413
Has ie, T., Tibshi ani, R., F iedman, J. H., & F iedman, J. H. (2017).
The elemen s o s a is ical lea ning: Da a mining, in e ence, and
p edic ion. Sp inge .
Heinz, D., Fassnach , M., Röh lee , J. H., Sagnie Ecke , L., & Sa zge ,
G. (2024). Designing digi al indus ial pla o ms o he ci cula
economy: A equi emen s ca alog. In ICIS.h ps:// aisel. aisne .
o g/ icis2 024/ go _ s a egy/ go _ s a egy/5
He ne , A., Cha e jee, S., T emblay, M. C., & Be nd , D. J. (2010).
The use o ocus g oups in design science esea ch. In Design
esea ch in in o ma ion sys ems: Theo y and p ac ice(pp.
121–143).Sp inge .
He ne , A. R., Ma ch, S. T., Pa k, J., & Ram, S. (2004). Design sci-
ence in in o ma ion sys ems esea ch. MIS Qua e ly, 28(1),
75–105.h ps:// doi. o g/ 10. 2307/ 25148 625
Hi , R., Kühl, N., & Sa zge , G. (2019). Cogni i e compu ing o
cus ome p o iling: Me a classi ica ion o gende p edic ion.
Elec onic Ma ke s, 29(1), 93–106.h ps:// doi. o g/ 10. 1007/
s12525- 019- 00336-z
Hsieh, C.-Y., Yeh, C.-K., Liu, X., Ra ikuma , P., Kim, S., Kuma , S., and
Hsieh, C.-J. (2020). E alua ions and me hods o explana ion h ough
obus ness analysis.h ps:// doi. o g/ 10. 48550/ a Xi . 2006. 00442
Hsieh, H.-F., & Shannon, S. E. (2005). Th ee app oaches o quali a-
i e con en analysis. Quali a i e Heal h Resea ch, 15(9), 1277–
1288.h ps:// doi. o g/ 10. 1177/ 10497 32305 276687
Hyun, Y., Kamioka, T., & Hosoya, R. (2020). Imp o ing agili y
using big da a analy ics: The ole o democ a iza ion cul u e. In
PACIS.h ps:// doi. o g/ 10. 17705/ 1pais. 12202
IEEE. (2008). IEEE s anda d o so wa e e iews and audi s. IEEE S d
1028–2008.h ps:// s and a ds. ieee. o g/ ieee/ 1028/ 4402/
IEEE S anda ds Associa ion. (2022). P2863 o ganiza ional go e n-
ance o a i icial in elligence wo king g oup. h ps:// sag o ups.
ieee. o g/ 2863
Ipho en, R., & K i ikos, M. (2021). Regula ing a i icial in elligence
and obo ics: E hics by design in a digi al socie y. Con empo a y
Social Science, 16(2), 170–184.h ps:// doi. o g/ 10. 1080/ 21582
041. 2018. 15638 03
Elec onic Ma ke s (2025) 35:24 Page 23 o 24 24
Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape o
AI e hics guidelines. Na u e Machine In elligence, 1(9), 389–
399.h ps:// doi. o g/ 10. 1038/ s42256- 019- 0088-2
Junklewi z, H., Hamon, R., And é, A., E as, T., Sole Ga ido, J., &
Sanchez Ma in, J. (2023). Cybe secu i y o a i icial in elligence
in he AI ac . Publica ions O ice Eu . Union, Luxembou g, UK,
Tech. Rep. JRC134461.h ps:// doi. o g/ 10. 2760/ 271009
Khandani, A. E., Kim, A. J., & Lo, A. W. (2010). Consume c edi - isk
models ia machine-lea ning algo i hms. Jou nal o Banking &
Finance, 34(11), 2767–2787.h ps:// doi. o g/ 10. 1016/j. jbank in.
2010. 06. 001
Kieslich, K., Kelle , B., & S a ke, C. (2022). A i icial in elligence
e hics by design. E alua ing public pe cep ion on he impo ance
o e hical design p inciples o a i icial in elligence. Big Da a &
Socie y, 9(1).h ps:// doi. o g/ 10. 1177/ 20539 51722 10929 56
Kleinbe g, J., Mullaina han, S., & Ragha an, M. (2017). Inhe en ade-
o s in he ai de e mina ion o isk sco es. In Con e ence on
Inno a ions in Theo e ical Compu e Science (ITCS).h ps:// doi.
o g/ 10. 48550/ a Xi . 1609. 05807
K aus, M., Feue iegel, S., & Oz ekin, A. (2020). Deep lea ning in
business analy ics and ope a ions esea ch: Models, applica ions
and manage ial implica ions. Eu opean Jou nal o Ope a ional
Resea ch, 281(3), 628–641.h ps:// doi. o g/ 10. 1016/j. ejo . 2019.
09. 018
Kuechle , B., & Vaishna i, V. (2008). On heo y de elopmen in design
science esea ch: Ana omy o a esea ch p ojec . Eu opean Jou -
nal o In o ma ion Sys ems, 17(5), 489–504.h ps:// doi. o g/ 10.
1057/ ejis. 2008. 40
Lamb ech , A., & Tucke , C. (2019). Algo i hmic bias? An empi ical
s udy o appa en gende -based disc imina ion in he display o
s em ca ee ads. Managemen Science, 65(7), 2966–2981.h ps://
doi. o g/ 10. 1287/ mnsc. 2018. 3093
Langenbuche , K. (2020). Responsible AI-based c edi sco ing–A legal
amewo k. Eu opean Business Law Re iew, 31(4).h ps:// doi.
o g/ 10. 54648/ eul 2 020022
Laza , S., & Nelson, A. (2023). AI sa e y on whose e ms?Science,
381(6654), 38.h ps:// doi. o g/ 10. 1126/ scien ce. adi89 82
Lundbe g, S. M., & Lee, S.-I. (2017). A uni ied app oach o in e -
p e ing model p edic ions. In Neu al In o ma ion P ocessing
Sys ems (NIPS).h ps:// doi. o g/ 10. 48550/ a Xi . 1705. 07874
Lyy inen, K., & King, J. L. (2006). S anda d making: A c i ical
esea ch on ie o in o ma ion sys ems esea ch. MIS Qua -
e ly, 30, 405–411.h ps:// doi. o g/ 10. 2307/ 25148 766
Magnani, G., & Gioia, D. (2023). Using he Gioia Me hodology in
in e na ional business and en ep eneu ship esea ch. In e -
na ional Business Re iew, 32(2), 102097.h ps:// doi. o g/ 10.
1016/j. ibus e . 2022. 102097
Ma ch, S. T., & Smi h, G. F. (1995). Design and na u al science
esea ch on in o ma ion echnology. Decision Suppo Sys-
ems, 15(4), 251–266.h ps:// doi. o g/ 10. 1016/ 0167- 9236(94)
00041-2
Ma us, K. J., & Veale, M. (2022). Ce i ica ion sys ems o machine
lea ning: Lessons om sus ainabili y. Regula ion & Go e n-
ance, 16(1), 177–196.h ps:// doi. o g/ 10. 1111/ ego. 12417
Meske, C., Abedin, B., Klie , M., & Rabhi, F. (2022). Explainable
and esponsible a i icial in elligence. Elec onic Ma ke s, 32,
2103–2106.h ps:// doi. o g/ 10. 1007/ s12525- 022- 00607-2
Me h, H., Muelle , B., & Maedche, A. (2015). Designing a equi e-
men mining sys em. Jou nal o he Associa ion o In o ma-
ion Sys ems, 16(9), 2.h ps:// doi. o g/ 10. 17705/ 1jais. 00408
Miah, S. J., & Genemo, H. (2016). A design science esea ch me h-
odology o expe sys ems de elopmen . Aus alasian Jou nal
o In o ma ion Sys ems, 20.h ps:// doi. o g/ 10. 3127/ ajis. 20i0.
1329
Michael, K., Abbas, R., & Roussos, G. (2023). Ai in cybe secu i y:
The pa adox. IEEE T ansac ions on Technology and Socie y,
4(2), 104–109.h ps:// doi. o g/ 10. 1109/ TTS. 2023. 32801 09
Mihale-Wilson, C. A., Zibuschka, J., Ca l, K. V., & Hinz, O. (2021).
Co po a e digi al esponsibili y-ex ended concep ualiza ion and
empi ical assessmen . In ECIS.h ps:// aisel. aisne . o g/ ecis2 021_
p/ 80
Mi chell, M., Wu, S., Zaldi a , A., Ba nes, P., Vasse man, L., Hu ch-
inson, B., Spi ze , E., Raji, I. D., & Geb u, T. (2019). Model
ca ds o model epo ing. In Con e ence on Fai ness, Accoun -
abili y, and T anspa ency (FAT).h ps:// doi. o g/ 10. 1145/ 32875
60. 32875 96
Mökande , J., & Flo idi, L. (2022). F om algo i hmic accoun abili y
o digi al go e nance. Na u e Machine In elligence, 4(6), 508–
509.h ps:// doi. o g/ 10. 1038/ s42256- 022- 00504-5
Moula, F. E., Guo ai, C., & Abedin, M. Z. (2017). C edi de aul p edic ion
modeling: An applica ion o suppo ec o machine. Risk Manage-
men , 19, 158–187.h ps:// doi. o g/ 10. 1057/ s41283- 017- 0016-x
Pa asu ama, P., & Sedoc, J. (2022). Gende ed in o ma ion in esumes
and i s ole in algo i hmic and human hi ing bias. In Academy
o managemen p oceedings (Vol. 2022, p. 1713). Academy o
Managemen B ia cli Mano .h ps:// doi. o g/ 10. 5465/ AMBPP.
2022. 285
Raji, I. D., Sma , A., Whi e, R. N., Mi chell, M., Geb u, T., Hu chin-
son, B., Smi h-Loud, J., The on, D., & Ba nes, P. (2020). Closing
he AI accoun abili y gap: De ining an end- o-end amewo k
o in e nal algo i hmic audi ing. In Con e ence on Fai ness,
Accoun abili y, and T anspa ency (FAT).h ps:// doi. o g/ 10. 1145/
33510 95. 33728 73
Robe s, H., Cowls, J., Mo ley, J., Taddeo, M., Wang, V., & Flo idi,
L. (2021). The Chinese app oach o a i icial in elligence: An
analysis o policy, e hics, and egula ion. In E hics, go e nance,
and policies in a i icial in elligence(pp. 47–79).h ps:// doi. o g/
10. 1007/ 978-3- 030- 81907-1_5
Rosen eld, A., & Richa dson, A. (2019). Explainabili y in human–
agen sys ems. Au onomous Agen s and Mul i-Agen Sys ems,
33(6), 673–705.h ps:// doi. o g/ 10. 1007/ s10458- 019- 09408-y
Roski, J., Maie , E. J., Vigilan e, K., Kane, E. A., & Ma heny, M. E.
(2021). Enhancing us in AI h ough indus y sel -go e nance.
Jou nal o he Ame ican Medical In o ma ics Associa ion, 28(7),
1582–1590.h ps:// doi. o g/ 10. 1093/ jamia/ ocab0 65
Scan ambu lo, T., Co és, A., & Schach , M. (2020). P og essing
owa ds esponsible AI. h ps:// doi. o g/ 10. 48550/ a Xi . 2008.
07326
Schneide , J., Ab aham, R., Meske, C., & Vom B ocke, J. (2023).
A i icial in elligence go e nance o businesses. In o ma ion
Sys ems Managemen , 40(3), 229–249.h ps:// doi. o g/ 10. 1080/
10580 530. 2022. 20858 25
Sch eck, P., & Rai hel, S. (2018). Co po a e social pe o mance, i m
size, and o ganiza ional isibili y: Dis inc and join e ec s on
olun a y sus ainabili y epo ing. Business & Socie y, 57(4),
742–778.h ps:// doi. o g/ 10. 1177/ 00076 50315 613
Sikde , M. F., Ramachand anpillai, R., de Leng, D., & Hein z, F.
(2024). Fai x: A comp ehensi e benchma king ool o model
analysis using ai ness, u ili y, and explainabili y.h ps:// doi. o g/
10. 48550/ a Xi . 2406. 14281
Smi h, J. F. (1977). The equal c edi oppo uni y ac o 1974: A cos /
bene i analysis. The Jou nal o Finance, 32(2), 609–622.h ps://
doi. o g/ 10. 2307/ 23267 94
Sokolo a, M., Japkowicz, N., & Szpakowicz, S. (2006). Beyond
accu acy, -sco e and oc: A amily o disc iminan measu es
o pe o mance e alua ion. In AI 2006: Ad ances in a i icial
in elligence (pp. 1015–1021). Sp inge .h ps:// doi. o g/ 10. 1007/
11941 439_ 114
Elec onic Ma ke s (2025) 35:24 24 Page 24 o 24
Thelisson, E., & Ve ma, H. (2024). Con o mi y assessmen unde
he EU AI ac gene al app oach. AI and E hics, 4(1), 113–
121.h ps:// doi. o g/ 10. 1007/ s43681- 023- 00402-5
Topuz, K., Zengul, F. D., Dag, A., Almehmi, A., & Yildi im, M.
B. (2018). P edic ing g a su i al among kidney ansplan
ecipien s: A Bayesian decision suppo model. Decision Sup-
po Sys ems, 106, 97–109.h ps:// doi. o g/ 10. 1016/j. dss. 2017.
12. 004
Van den B oek, E., Se gee a, A., & Huysman, M. (2021). When he
machine mee s he expe : An e hnog aphy o de eloping AI o
hi ing. MIS Qua e ly, 45(3).h ps:// doi. o g/ 10. 25300/ MISQ/
2021/ 16559
Van Some en, M., Ba na d, Y. F., & Sandbe g, J. (1994). The hink
aloud me hod: A p ac ical app oach o modelling cogni i e(Vol.
11). Ci esee .
Veale, M., & Zuide een Bo gesius, F. (2021). Demys i ying he d a
EU A i icial In elligence Ac —Analysing he good, he bad, and
he unclea elemen s o he p oposed app oach. Compu e Law
Re iew In e na ional, 22(4), 97–112.h ps:// doi. o g/ 10. 9785/
c i- 2021- 220402
Venable, J., P ies-Heje, J., & Baske ille, R. (2012). A comp ehensi e
amewo k o e alua ion in design science esea ch. In Design
science esea ch in in o ma ion sys ems. Ad ances in heo y and
p ac ice(pp. 423–438). Sp inge .h ps:// doi. o g/ 10. 1007/ 978-3-
642- 29863-9_ 31
on Zahn, M., Baue , K., Mihale-Wilson, C., Jagow, J., Speiche , M.,
& Hinz, O. (2024). Sma g een nudging: Reducing p oduc
e u ns h ough digi al oo p in s and causal machine lea ning.
Ma ke ing Science, o hcoming.h ps:// doi. o g/ 10. 1287/ mksc.
2022. 0393
on Zahn, M., Feue iegel, S., & Kuehl, N. (2022). The cos o ai -
ness in AI: E idence om e-comme ce. Business & In o ma-
ion Sys ems Enginee ing, 64(3), 335–348.h ps:// doi. o g/ 10.
1007/ s12599- 021- 00716-w
Wang, B., Chen, W., Pei, H., Xie, C., Kang, M., Zhang, C., Xu, C.,
Xiong, Z., Du a, R., Schae e , R., T uong, S. T., A o a, S.,
Mazeika, M., Hend ycks, D., Lin, Z., Cheng, Y., Koyejo, S.,
Song, D., & Li, B. (2023). Decoding us : A comp ehensi e
assessmen o us wo hiness in GPT models. Ad ances in neu-
al in o ma ion p ocessing sys ems (NIPS).h ps:// doi. o g/ 10.
48550/ a Xi . 2306. 11698
Wang, L., Gopal, R., Shanka , R., & Panc as, J. (2022). Fo ecas ing
enue popula i y on loca ion-based se ices using in e p e able
machine lea ning. P oduc ion and Ope a ions Managemen ,
31(7), 2773–2788.h ps:// doi. o g/ 10. 1111/ poms. 13727
Wilson, C., Ghosh, A., Jiang, S., Mislo e, A., Bake , L., Sza y, J., T in-
del, K., & Polli, F. (2021). Building and audi ing ai algo i hms:
A case s udy in candida e sc eening. In Con e ence on Fai ness,
Accoun abili y, and T anspa ency (FAccT).h ps:// doi. o g/ 10.
1145/ 34421 88. 34459 28
Zacha ias, J., on Zahn, M., Chen, J., & Hinz, O. (2022). Designing
a ea u e selec ion me hod based on explainable a i icial in el-
ligence. Elec onic Ma ke s, 32(4), 2159–2184.h ps:// doi. o g/
10. 1007/ s12525- 022- 00608-1
Publishe 's No e Sp inge Na u e emains neu al wi h ega d o
ju isdic ional claims in published maps and ins i u ional a ilia ions.