Am i , Chin an; Na ayanappa, Ashwini Kola
A icle
An analysis o he challenges in he adop ion o MLOps
Jou nal o Inno a ion & Knowledge (JIK)
P o ided in Coope a ion wi h:
Else ie
Sugges ed Ci a ion: Am i , Chin an; Na ayanappa, Ashwini Kola (2025) : An analysis o he challenges
in he adop ion o MLOps, Jou nal o Inno a ion & Knowledge (JIK), ISSN 2444-569X, Else ie ,
Ams e dam, Vol. 10, Iss. 1, pp. 1-16,
h ps://doi.o g/10.1016/j.jik.2024.100637
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/327539
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 ps://c ea i ecommons.o g/licenses/by-nc-nd/4.0/
An analysis o he challenges in he adop ion o MLOps
Chin an Am i
*
, Ashwini Kola Na ayanappa
Ams e dam Business School, Uni e si y o Ams e dam, P.O. Box 15953 1001 NL Ams e dam, he Ne he lands
ARTICLE INFO
JEL classi ica ion:
C88
D83
L86
L17
O32
Keywo ds:
Machine lea ning ope a ions (MLOps)
G ounded heo y
Da a science
ABSTRACT
The ield o MLOps (Machine Lea ning Ope a ions), which ocuses on e ec i ely managing and ope a ionalizing
ML wo k lows, has g own because o he ad ancemen s in machine lea ning (ML). The goal o his s udy is o
examine and con as he di icul ies encoun e ed in he implemen a ion o MLOps in en e p ises wi h hose
encoun e ed in De Ops. An SLR (Sys ema ic Li e a u e Re iew) is he i s s ep in he esea ch p ocess o ind he
issues aised in he li e a u e. The esul s o his s udy a e based on quali a i e con en analysis using g ounded
heo y and semi-s uc u ed in e iews wi h 12 ML p ac i ione s om di e en sec o s. O ganisa ional, echnical,
ope a ional, and business p oblems a e he ou dis inc aspec s o challenges o MLOps ha ou s udy high-
ligh s. These challenges a e u he de ined by ele en di e en hemes. Ou esea ch indica es ha while some
issues, such as da a and model complexi y, a e unique o MLOps, o he s a e sha ed by De Ops and MLOps as
well. The epo o e s sugges ions o u he esea ch and summa ises he di icul ies.
In oduc ion
The inc easing use o machine lea ning (ML)-based echniques has
made i mo e di icul o in eg a e hem in o p oduc ion sys ems while
main aining he e iciency and dependabili y o cons an ly changing ML
p ojec s (Rzig e al., 2022). Acco ding o Ven u eBea esea ch (VB S a ,
2019), jus 13% o machine lea ning p ojec s in he business can each
p oduc ion. P o o yping and p oduc ion deploymen can ake up o 90%
o he p ojec ’s ime, despi e i seeming like he inal 10% (Flaounas,
2017). In esponse o hese challenges, he no ion o MLOps (Machine
Lea ning Ope a ions) has eme ged as a comp ehensi e collec ion o
p ocedu es in ended o gua an ee he dependable and e ec i e imple-
men a ion and upkeep o machine lea ning (ML) models in ope a ional
se ings (Alla & Ada i, 2020). MLOps is an adap a ion o he De Ops
discipline, which was c ea ed o add ess compa able con inuous
deploymen p oblems o s anda d so wa e and has been in exis ence
o mo e han en yea s. MLOps, as opposed o De Ops, ies o add ess
p oblems speci ic o machine lea ning, like con inuous aining, model
moni o ing, es ing, and e sioning o da a and models.
Al hough MLOps a e becoming mo e and mo e popula , he e isn’
much esea ch on how au oma ion is adop ed and how ha a ec s
changes in ML-enabled sys ems (Cale a o e al., 2022). Nowadays, all we
know abou MLOps comes om a agmen ed ield o whi e pape s,
s o ies, and opinion a icles (Shanka e al., 2022). Acco ding o John
e al. (2021), a signi ican amoun o he li e a u e on so wa e
enginee ing (SE) bes p ac ices o machine lea ning applica ions is
non-pee - e iewed, o g ey li e a u e. This ca ego y includes p esen a-
ion slides, blog pos s, and whi e pape s (Se ban e al., 2020). Mos
w i e s ag ee ha MLOps is challenging. 90% o ML models a e ne e
pu in o p oduc ion, and 85% o ML ini ia i es a e ine ec i e (Shanka
e al., 2022).
Ve y ew p oduc ion-g ade machine lea ning p ojec s we e disco -
e ed by Cale a o e al. (2022) du ing hei examina ion o MLOps p o-
jec s on Gi Hub. They also d ew a en ion o he dea h o MLOps
ool-using open-sou ce machine lea ning sys ems (Cale a o e al.,
2022). The e a e e y ew a icles on case s udies om o ganisa ions.
Some pape s examine he echnical obs acles o applying MLOps and
o e solu ions (Ca doso Sil a e al., 2020; Symeonidis e al., 2022). We
see he need o a comp ehensi e amewo k ha discusses he chal-
lenges o ganiza ions ace in implemen ing MLOps. While exis ing
esea ch has explo ed a ious challenges in he implemen a ion o
MLOps, he e is a no able gap in p o iding a comp ehensi e amewo k
o he di e en challenges. P e ious s udies ha e o en ocused on spe-
ci ic challenges in deploying and ope a ing machine lea ning in p ac ice
o su eyed indi idual case s udies o ML deploymen ( o e.g.: Baie
e al. (2019), Diaz-de-A caya e al. (2023), Paleyes e al. (2022)) bu a
uni ied amewo k ha in eg a es all essen ial elemen s is s ill missing.
This lack o a comp ehensi e o e iew has led o inconsis en un-
de s andings and implemen a ions o MLOps ac oss di e en o ganiza-
ions and p ojec s. Addi ionally, he apid e olu ion o ML echnologies
* Co esponding au ho .
E-mail add esses: [email p o ec ed] (C. Am i ), [email p o ec ed] (A.K. Na ayanappa).
Con en s lis s a ailable a ScienceDi ec
Jou nal o Inno a ion & Knowledge
jou nal homepage: www.else ie .com/loca e/jik
h ps://doi.o g/10.1016/j.jik.2024.100637
Recei ed 16 Ap il 2024; Accep ed 28 No embe 2024
Jou nal o Inno a ion & Knowledge 10 (2025) 100637
A ailable online 4 Decembe 2024
2444-569X/© 2024 The Au ho s. Published by Else ie España, S.L.U. on behal o Jou nal o Inno a ion & Knowledge. This is an open access a icle unde he CC
BY-NC-ND license (
h p://c ea i ecommons.o g/licenses/by-nc-nd/4.0/ ).
and p ac ices has ou paced he academic li e a u e, esul ing in a
disconnec be ween heo e ical concep s and p ac ical applica ions. This
esea ch aims o b idge his gap by p o iding a ho ough, up- o-da e,
and p ac ice-o ien ed concep ualiza ion o he challenges in MLOps
implemen a ion, syn hesizing insigh s om bo h academic li e a u e
and indus y expe ise.
Since he e isn’ much esea ch ha con ex ualises he li e a u e on
MLOps, exposes he challenges ela ed o MLOps, and o e s a ho ough
o e iew o ecen ma e ial, we ocused on e alua ing he challenges
ha companies ace while implemen ing MLOps in his s udy. We
he e o e ask he ollowing esea ch ques ion:
Wha a e he key challenges o ganisa ions ace while implemen ing
MLOps?
The esul s om ou s uc u ed li e a u e e iew (SLR) show many
implemen a ion p oblems in addi ion o a gene al lack o empi ical da a
on MLOps. The esul s o ou in e iews show ha he e a e ou majo
ypes o obs acles: echnical, ope a ional, business, and o ganisa ional.
These challenges a e u he di ided in o ele en opics o explana ion.
Also, one o he p ima y con ibu ions o ou wo k is a i ing a a Ty-
pology o he challenges ha ep esen Type 1 heo y (G ego , 2006),
which is he cen al heo e ical con ibu ion o his pape .
The emainde o his pape is s uc u ed as ollows. Sec ion 2 gi es a
backg ound o he exis ing li e a u e on MLOps, sec ion 3 desc ibes ou
esea ch design, and sec ion 4 p o ides esul s om he Sys ema ic
Li e a u e Re iew. Sec ion 5 p esen s and analyses he esul s o he
In e iews. We discuss hese esul s in Sec ion 6 and we conclude he
pape in Sec ion 7.
Li e a u e backg ound
We in oduce and e iew exis ing li e a u e on MLOps, which is a
modi ica ion o De Ops. The ollowing subsec ion highligh s hei
di e ences.
De Ops s MLOps
The de ini ion o De Ops is he subjec o deba e in esea ch (K ey
e al., 2022). Di e en iews and s ances exis in he scien i ic com-
muni y. Maca hy and Bass (2020) highligh wo con lic ing pe spec-
i es: one iews De Ops as a cul u al mo emen o apid so wa e
de elopmen , while he o he iews i as a job i le equi ing bo h
de elopmen and IT ope a ion expe ise. In his pape , we will conside
he ollowing de ini ion: De Ops is a pa adigm ha aims o b ing
inno a i e p oduc s and ea u es o he ma ke as e (Ebe e al. 2016)
by in eg a ing he de elopmen , es ing, and ope a ional so wa e
de elopmen eams h ough au oma ion, ools, (Maca hy & Bass, 2020)
and cul u al philosophy ha emphasizes eam empowe men ,
c oss- eam communica ion, collabo a ion, and echnology au oma ion.
Academic s udies ocusing on he p inciples, echniques, and ad-
an ages o De Ops a e well-documen ed. Acco ding o esea ch s udies,
he undamen al componen s o De Ops include con inuous deli e y,
so wa e de elopmen p ocess au oma ion, and communica ion and
collabo a ion be ween di e en ac o s deli e ing he so wa e
(Sub amanya e al. 2022; Da ies & Daniels, 2016). MLOps is no he
same as De Ops as i has mo e componen s in ol ed han jus he
so wa e applica ion. In De Ops, he e needs o be a close collabo a ion
wi h he de elopmen and ope a ions eams, whe eas in MLOps, i also
ex ends o he da a science eam. Makinen e al. (2021) also con i med
in hei su ey esea ch ha , like De Ops o adi ional so wa e,
MLOps, o con inuous deli e y o machine lea ning so wa e, is
becoming a p e equisi e o businesses using ML in p oduc ion, espe-
cially o mo e ML-ma u e o ganiza ions. Since cloud-based De Ops and
con aine ised mic ose ices ha e shown success in p oduc ion de-
ploymen s (Hui Kang e al., 2016), o ganisa ions a e adop ing
con inuous p ac ices in ML sys em de elopmen h ough he adop ion o
De Ops concep s o end- o-end au oma ion (John e al., Sep 2021).
MLOps is p esen ly in he "Peak o In la ed Expec a ions" phase, pe
Ga ne ’s (2022) hype cycle o da a science (DS) and machine lea ning
(ML) (Choudha y & K ensky, 2022).
MLOps, like De Ops, has no o mal de ini ion bu can be seen as he
mee ing poin o ML and De Ops p ac ices (Ma sui & Goya, 2022). An
MLOps p ojec encompasses in as uc u e managemen , in eg a ion,
es ing, elease, and deploymen , as well as au oma ion, in eg a ion, and
moni o ing h oughou he en i e p ocess o cons uc ing an ML sys em
(Tes i e al. 2022a). The goal o MLOps is o c ea e a se o p ocedu es o
quickly and e icien ly c ea ing machine lea ning models using ools,
deploymen lows, and wo k p ocesses.
While De Ops p ima ily deals wi h so wa e ac ions, MLOps in ol es
cons uc ing, aining, and uning machine lea ning models using
hype pa ame e s and da ase s (Liu e al. 2020). Ano he signi ican
dis inc ion be ween MLOps and De Ops is ha while he o me em-
phasizes op imizing deli e y p ocedu es and s anda dizing de elopmen
en i onmen s, he la e places a s ong emphasis on da a om which
applica ions a e de i ed. As a esul , da a scien is s and so wa e engi-
nee s collabo a e closely in MLOps (Mboweni e al., 2022).
MLOps b ings in newe challenges han adi ional so wa e wi h i s
complexi y. Unde s anding he challenges o MLOps in ol es g asping
he s eps equi ed o aining and deploying ML modules. This includes
da a p epa a ion, di iding da a in o aining, es ing, and c oss-
alida ion se s, selec ing an ML model and hype pa ame e s, aining
he model wi h i e a i e adjus men s, alida ing he model, and
deploying i . Once deployed, ML ea u es equi e moni o ing, consid-
e ing ML-speci ic ac o s like biases and d i . Addi ionally, echniques
o imp o ing he model in eal- ime while in use mus be inco po a ed
in o he moni o ing sys em. In summa y, he p e-model comple ion
phases in ML esemble a wa e all app oach while ope a ionalizing he
model aligns wi h con en ional so wa e p ac ices (Makinen e al.,
2021; G anlund e al., 2021).
MLOps, like De Ops, aims o au oma e so wa e deli e y, ensu ing
con inuous deli e y and eedback loops. Howe e , in eg a ing ML ap-
plica ions in o a De Ops-based CI/CD en i onmen equi es addi ional
s eps due o a ia ions in ML me hodologies. The adap a ion o MLOps is
s ill in i s ea ly s ages, wi h limi ed esea ch compa ed o De Ops. Da a
scien is s and ope a ions eams a e wo king on au oma ing he end- o-
end ML li ecycle using De Ops concep s, acing challenges in gene al-
izing MLOps componen s. Di e en designs and p ocesses, such as
I e a i e/Inc emen al and Con inuous deli e y o ML(CD4ML), ha e
been p oposed (Sub amanya e al., 2022). I is c ucial o include da a,
in as uc u e, and co e ML code in he MLOps li ecycle, as highligh ed
by Sculley e al. (2015).
Challenges in ML sys ems, like model complexi y and ep oduc-
ibili y, a e di e en om De Ops componen s. Howe e , MLOps sha es
ypical De Ops limi a ions, including he o e sigh o human ac o s in
echnology de elopmen and adop ion (Mucha e al., 2022). Add essing
e sioned da a, ML models, and dependencies is a challenge, as exis ing
ools and p ocesses in De Ops a e inadequa e. Expe ise in da a science,
AI, and so wa e enginee ing is equi ed o success ul MLOps imple-
men a ion (Lwaka a e e al., 2020). The li e a u e backg ound con i ms
ha he e is a g owing need o MLOps in businesses ha use ML in
p oduc ion.
MLOps ools
Many o ganiza ions ha e he op ion o ei he build hei ools and
pla o ms o choose om he a ailable ools in he ma ke . MLOps
p ocess and i s componen s can be buil using a combina ion o open-
sou ce ools and en e p ise solu ions, which allows o lexibili y and
cus omiza ion. I is possible o le e age bo h en e p ise and open-sou ce
ools oge he o achie e MLOps objec i es e ec i ely (K euzbe ge
e al., 2023).
C. Am i and A.K. Na ayanappa
Jou nal o Inno a ion & Knowledge 10 (2025) 100637
2
Se e al MLOps pla o ms ha e eme ged in ecen yea s o o e
guidelines o building en e p ise-le el AI/ML applica ions (Ga g e al.,
2022). App oxima ely 50% o IT businesses a e using hese ools, which
is e iden om he s agge ing numbe s o p ojec s, de elope s, and
companies ac i ely engaged in pla o ms like Gi Hub. As o Janua y 26,
2022, 200 million p ojec s a e being wo ked on by 65 million de elope s
and 3 million companies (Symeonidis e al., 2022).
Kube low, MLFlow, and Ked o a e some o he open-sou ce MLOps
echnologies. These pla o ms a e open-sou ce p ojec s ha con ain a
cu a ed selec ion o compa ible machine-lea ning ools and amewo ks
o simpli y he au oma ed de elopmen p ocess (Ga g e al., 2022). The
mos well-known MLaaS include Azu e’s Machine Lea ning, Google’s AI
Pla o m, and Amazon’s SageMake (Maya & Felipe, 2021).
In he nex sec ion we discuss ou esea ch design o unde s anding
he key challenges o ganiza ions ace in hei implemen a ion o MLOps.
Resea ch design
When he e is minimal p e ious esea ch o when he empi ical
en i onmen is new o unde s udied, induc i e easoning based on
quali a i e da a is pa icula ly applicable (Bansal e al., 2018). The e-
o e, o ob ain a comp ehensi e o e iew o pe inen esea ch, we i s
pe o med a sys ema ic li e a u e e iew (SLR). Concu en ly, we con-
duc ed semi-s uc u ed in e iews wi h ML specialis s om a ange o
businesses. To each pe inen conclusions, he esul s o he SLR we e
hen con as ed wi h indings om expe in e iews (in he discussion
po ion).
Indi idual in e iews wi h pa icipan s who a e knowledgeable
abou MLOps and can o e an opinion on i we e used o ga he da a o
he s udy. This is a ypical app oach o ga he ing da a (Bei in, 2012).
To disco e indi iduals wi h expe ience wo king wi h machine
lea ning and/o i s ope a ions, we used a pu pose ul sampling echnique
o ind pa icipan s wi h expe ience wo king wi h machine lea ning
and/o i s ope a ions wi h job i les such as Da a scien is , ML Manage ,
MLOps enginee , o compa able. The esponden s we e ound h ough
mul iple channels like in he pe sonal ne wo k o he i s au ho ,
LinkedIn, and online ML communi ies. We go 12 esponden s who
ag eed o he in e iew. We used an in-dep h, semi-s uc u ed in e iew
me hod which allowed us o change he o de ing o ques ions o ask
abou he ele an inding o one in e iew o ano he (Qu and Dumay,
2011). All he in e iews we e pe o med online using Mic oso Teams,
and hey we e all eco ded a e ge ing he pa icipan ’s consen . In-
e iews ypically las ed be ween 25 and 60 minu es. To ensu e
au hen ic esponses, in e iew ques ions we e no e ealed be o e he
in e iews, and he in e iew s a ed o wi h a sho in oduc ion o he
pu pose o his esea ch. The i s d a was c ea ed using he an-
sc ibing ea u e o Mic oso Teams, and all adjus men s we e ca e ully
e iewed and modi ied. The i s au ho also made no es du ing he
in e iew, which made he pa icipan s hink and add o he answe
du ing he silence (Qu & Dumay, 2011).
Fig. 1 shows ou o e all esea ch design. As a esul , we conduc ed
he semi-s uc u ed in e iews and he s uc u ed li e a u e e iew
simul aneously. The insigh s om he SLR and he in e iews we e
combined o c ea e he ypology o challenges ha we a i e a in Fig. 2.
We s a wi h he de ails o he SLR in he nex sec ion.
S uc u ed li e a u e e iew
To be e unde s and he obs acles ha o ganisa ions encoun e
while deploying MLOps, we s a ed by pe o ming a ho ough analysis o
exis ing li e a u e. Se ing a esea ch opic, choosing pe inen s udies,
ex ac ing da a, and syn hesising he da a we e all s eps in he me h-
odology we ollowed o pe o m he SLR (Ki chenham, 2004). We
employed a manual sea ch s a egy using elec onic esea ch esou ces
o his SLR. To ind pe inen ma e ial, we sea ched bo h Scopus and
ScienceDi ec , wo impo an da abases. We hen complemen ed his
sea ch wi h a sea ch using Google Schola and he snowballing ech-
nique (Wohlin, 2014).
The elease o "Hidden Technical Deb in Machine Lea ning Sys ems"
in 2015 (Sculley e al., 2015) helped o popula ise he ela i ely young
a ea o MLOps. We c ea ed inclusion and exclusion c i e ia, which a e
ou lined in Table 3 and Table 4 (in Appendix A), and limi ed ou sea ch
o publica ions published be ween 2015 and 2023 o make su e ha we
concen a e on pe inen in o ma ion. When i came o he challenges
associa ed wi h in oducing MLOps, we p io i ised o ganisa ional
esea ch. The ini ial esul s we e limi ed o pape s published be ween
2018 and 2023 and we e p oduced using he sea ch ph ases "MLOps OR
machine lea ning in p oduc ion OR machine lea ning ope a ions". Ou o
175 i ems, we only included esea ch a icles and con e ence pape s
wi h lib a y access. The i s au ho manually sc eened he abs ac s o
hese pape s and selec ed en a icles o da a ex ac ion and syn hesis.
The p ocess is depic ed in Fig. 3 (in Appendix B).
Al hough no all he publica ions we ound add essed MLOps spe-
ci ically, nine o hem add essed MLOps’ challenges. We only included
sec ions ha we e di ec ly ela ed o he di icul ies aced by MLOps o
ensu e hei ele ancy. Table 1 displays he concep ma ix o a ious
asks.
When deploying ML models in p oduc ion, o ganisa ions ha e
dis inc p oblems ela ed o da a quali y and quan i y, as no ed by Baie
e al. (2019) in hei analysis o MLOps in p ac ice. Long- e m da a
alidi y is especially challenging o add ess. The da a u ilised o
aining mus be o a high calib e and adequa e h oughou ime o
gua an ee co ec p edic ions when he ML model is p esen ed wi h
esh da a o es ing in p oduc ion. The au ho s also disco e ed ha
managing concep d i is an addi ional special di icul y ela ed o
MLOps. Concep d i is he p ocess ia which he model becomes
obsole e as unde lying da a, use beha iou , and/o s a is ical cha ac-
e is ics o he da a change o e ime. Baie e al. (2019) iden i ied
non- echnical challenges as well as echnical ones. These challenges
included adhe ing o o ganisa ional and na ional s anda ds o e hical
and legal compliance, in eg a ing and s eamlining he ML model in o
exis ing business ope a ions, and a lack o collabo a ion and commu-
nica ion among echnical and non- echnical eams.
Acco ding o Tambu i (2020), he e a e se e al in e nal and
ex e nal hu dles ha o ganisa ions ha e when implemen ing MLOps.
These include in eg a ing MLOps wi h he cu en IT in as uc u e,
which may lead o p oblems wi h scalabili y and sus ainabili y, and
making su e MLOps a e in line wi h he o ganiza ion’s objec i es and
co e alues. To ensu e adhe ence o pe inen laws and s anda ds, he
MLOps pipeline’s sa e y and secu i y a e also majo conce ns. Finding
quali ied candida es wi h he MLOps skill se who a e mo i a ed,
Fig. 1. The esea ch design adop ed in his pape .
C. Am i and A.K. Na ayanappa
Jou nal o Inno a ion & Knowledge 10 (2025) 100637
3
expe ienced, and ele an is ano he di icul y.
Di e en o ganisa ions ha e di e en obs acles when pu ing MLOps
in o p ac ice. G anlund e al. (2021) poin ed ou ha he gene a ion,
in eg a ion, and disco e y o new ea u es ha a e con ibu ed o he
ea u e se ed o he ML model all esul in he cons an upda ing o ML
models. Consequen ly, MLOps ha e an u gen issue in de eloping he
capaci y o deploy an upda ed ML model ha enables e icien and
ins an aneous deploymen . The In eg a ion Challenge and he Scaling
Challenge a e wo mo e signi ican MLOPs di icul ies ha a e co e ed
in his s udy. The In eg a ion Challenge a ises when he e a e dispa a e
con ac ual du ies, da a o ma s, and machine lea ning ea u es among
se e al o ganisa ions, esul ing in incompa ible APIs.
In o ganisa ional amewo ks, Lima e al. (2022) examine he
di e en MLOps me hods and di icul ies. One o hese di icul ies is ML
model e sioning, which has o do wi h selec ing he igh model among
a a ie y o e sions o he ML models c ea ed along he ML pipeline.
The au ho s also no e ha a majo sou ce o esis ance o MLOps,
especially among in e nal eams wi hin businesses, is he absence o
s anda disa ion. Gi en ha a mul i ude o ools ac oss a ious pla o ms
can be used o c ea e ML models, da ase s, and ea u e se s (Lima e al.,
2022).
Challenges ela ed o he explainabili y, and anspa ency o ma-
chine lea ning models and decisions we e discussed by Tes i e al.
(2022). In he con ex o MLOps o o ganisa ions, a co ec knowledge
o he ML model h ough anspa ency and explainabili y is essen ial
since he elemen s in luencing he ML model’s a ious decisions mus be
p ecisely ecognised. In addi ion o a dea h o s anda disa ion and bes
p ac ices, businesses implemen ing MLOps ha e a se ious issue wi h
limi ed and isola ed esea ch. Fu he mo e, Tes i e al. (2022) disco -
e ed ha companies we e s uggling o deal wi h he con inuously
changing da a, which occasionally necessi a es e aining o e en a o al
ebuild o he model in he machine lea ning pipeline.
Painoli and Da ike (2021) concen a e on gene al AI p oblems in
businesses. The au ho s d aw a en ion o businesses ha a e ha ing
ouble selec ing which quali ies o include in hei model, how o clean
hei da a, and wha algo i hm is app op ia e. The au ho s also poin ou
ha success ul deploymen equi es o e coming challenges wi h in a-
s uc u e, da a p i acy, and secu i y. Lack o quali ied pe sonnel also
makes i di icul o o ganisa ions o unde s and and sha e he model’s
esul s. The au ho s conclude ha making su e he model is ai , accu-
a e, and anspa en when making judgemen s is one o he la ges
p oblems in applying AI/ML (Painoli & Da ika, 2021).
Paleyes e al. (2022) su ey case s udies o iden i y challenges
associa ed wi h implemen ing machine lea ning models. They no e ha
ea u e enginee ing, model selec ion, and da a managemen a e issues
ha en e p ises equen ly deal wi h. The au ho s also emphasise he
impo ance o he models’ in e p e abili y and explainabili y as well as
he need o igo ous es ing and alida ion p ocedu es. They also add
ha a lack o in as uc u e and skilled pe sonnel may make i di icul
o use machine lea ning models. When applying machine lea ning
models, he au ho s poin ou he g owing signi icance o da a p i acy
and e hical conside a ions including jus ice and anspa ency (Paleyes
Fig. 2. The ypology o challenges in MLOps implemen a ion.
C. Am i and A.K. Na ayanappa
Jou nal o Inno a ion & Knowledge 10 (2025) 100637
4
e al., 2022).
O ganisa ional challenges in moni o ing machine lea ning models
a e explained by Sch ¨
ode and Schulz (2022) while emphasising ope -
a ions. Issues pe aining o da a d i , mislabelling, and quali y a e
emphasised. The au ho s also emphasise how impo an i is o keep an
eye on he model’s pe o mance and add ess issues wi h ai ness, bias,
and in e p e abili y. Addi ionally, hey no e ha companies may ind i
challenging o implemen and in eg a e moni o ing sys ems in o
al eady-exis ing in as uc u es (Sch ¨
ode & Schulz, 2022).
Con e sely, Zhang e al. (2020) con end ha he de elopmen o
machine lea ning a i icial in elligence sys ems is pa icula ly chal-
lenging in knowledge-in ensi e ields. Ine ec i e adop ion and esis-
ance o change may a ise om a lack o unde s anding ega ding he
in eg a ion o AI wi h exis ing o ganisa ional sys ems and p o ocols. A
u he challenge hey iden i y is choosing which models and algo i hms
o u ilise om he many accessible models and algo i hms while main-
aining a balance be ween accu acy, accessibili y, and complexi y. Da a
quali y issues could nega i ely impac AI sys em pe o mance because
hey equi e la ge amoun s o da a o lea n. This p esen s ano he
challenge. Fu he mo e, while mul idisciplina y eams wi h a ange o
skill se s a e necessa y o de eloping AI sys ems, hey can be chal-
lenging o loca e and manage (Zhang e al., 2020).
K euzbe ge e al. (2023) iden i y se e al key challenges o imple-
men ing MLOps, ca ego ized in o o ganiza ional, ML sys em, and
ope a ional challenges. O ganiza ional challenges include he need o a
cul u e shi owa ds p oduc -o ien ed machine lea ning, a lack o skil-
led expe s in oles like ML enginee s and De Ops enginee s, and he
necessi y o mul idisciplina y eamwo k. ML sys em challenges in ol e
designing o luc ua ing demand, especially in ML aining p ocesses,
and managing po en ially oluminous and a ying da a. Ope a ional
challenges include he di icul y o manually ope a ing ML due o
complex so wa e and ha dwa e s acks, he need o obus au oma ion
o handle cons an da a s eams and e aining, go e nance o nume ous
a i ac s, e sioning o da a, models, and code, and he complexi y o
esol ing suppo eques s due o he in ol emen o mul iple compo-
nen s and pa ies.
The sys ema ic e iew by Diaz-de-A caya e al. (2023) examines
challenges, oppo uni ies, and ends in MLOps and AIOps. Whe e hey
de ine AIOps as a solu ion o handling g owing da a and IT in-
as uc u es. The key challenges hey desc ibe include c oss-domain
expe ise equi emen s, da a managemen complexi ies, and o ganiza-
ional cul u e ba ie s. They highligh oppo uni ies, such as, con inuous
deli e y in MLOps, AI applica ions ac oss indus ies, and edge
compu ing u iliza ion. They discuss amewo ks acili a ing MLOps and
AIOps adop ion, emphasizing li ecycle managemen ools, and s a e ha
while MLOps is mo e p e alen in adi ional indus ies, AIOps is gain-
ing ac ion in eme ging echnologies like 5G The au ho s conclude ha
bo h me hodologies a e c ucial o success ul AI implemen a ion in
p oduc ion en i onmen s, equi ing collabo a i e cul u e and
c oss- unc ional skills o o e come iden i ied challenges.
The li e a u e e iew, as summa ized in Table 1, e eals a ange o
challenges o ganiza ions ace when implemen ing MLOps. The mos
p e alen issues include model- ela ed challenges (such as scalabili y,
accu acy, e sioning, and moni o ing), da a issues (a ailabili y, quali y,
p i acy), and in eg a ion and in as uc u e conce ns ( he i s h ee
columns o Table 1). Regula o y compliance, s anda diza ion, and lack
o ool suppo o MLOps we e also iden i ied as signi ican hu dles.
Fu he mo e, he e iew highligh s human-cen ic challenges like lack
o alen , collabo a ion di icul ies, and knowledge gaps. Howe e , he
impac o ool suppo and es ing has been gi en li le a en ion in he
li e a u e. We ound hey we e men ioned only once in he li e a u e we
e iewed (Table 1). In gene al, while s udies p o ide aluable insigh s
in o speci ic aspec s o MLOps implemen a ion, hey o en ocus on
speci ic componen s o p ac ices. This agmen ed app oach unde sco es
he need o a mo e comp ehensi e amewo k ha in eg a es all
essen ial elemen s o MLOps. In he nex sec ion, we add ess his gap by
using semi s uc u ed in e iews and induc i ely c ea ing a ypology (a
Type 1 heo y) o he challenges in implemen ing MLOps.
In he nex sec ion, we discuss he esul s om he semi-s uc u ed
in e iews ha we compa e wi h he SLR indings.
Table 1
Concep ma ix o challenges aced in MLOps.
A icles MLOps Challenges – Concep ma ix
Model Issues
(Scalabili y,
accu acy,
Ve sioning,
Moni o ing)
Da a Issue
(A ailabili y,
quali y,
p i acy)
In eg a ion
and
In as uc u e
Regula o y
Compliance
S anda diza ion Tools
Suppo
o
MLOps
Tes ing Lack
o
Talen
Collabo a ion &
communica ion
Lack o
Knowledge
Paleyes e al.
(2022)
✔ ✔ ✔ ✔ ✖ ✖ ✖ ✖ ✖ ✖
G anlund e al.,
(2021)
✔ ✔ ✔ ✖ ✖ ✖ ✖ ✖ ✔ ✔
Lima e al.,
(2022)
✔ ✔ ✔ ✖ ✔ ✖ ✖ ✖ ✖ ✖
Baie e al.,
(2019)
✔ ✔ ✔ ✔ ✔ ✔ ✖ ✖ ✔ ✔
Tambu i
(2020)
✖ ✔ ✖ ✔ ✖ ✖ ✖ ✔ ✖ ✔
Painoli &
Da ike
(2021)
✔ ✔ ✔ ✖ ✖ ✖ ✖ ✖ ✖ ✖
Sch ¨
ode and
Schulz (2022)
✖ ✔ ✔ ✖ ✖ ✖ ✔ ✖ ✖ ✖
Tes i e al.,
(2022)
✔ ✔ ✖ ✖ ✔ ✖ ✖ ✖ ✖ ✔
Zhang e al.
(2020)
✔ ✔ ✖ ✖ ✖ ✖ ✖ ✔ ✖ ✔
K euzbe ge
e al. (2023)
✔ ✖ ✔ ✖ ✖ ✖ ✖ ✔ ✔ ✔
Diaz-De-Z caya
e al. (2023)
✔ ✔ ✔ ✖ ✖ ✖ ✖ ✔ ✔ ✔
To al 9 10 8 3 3 1 1 4 4 7
C. Am i and A.K. Na ayanappa
Jou nal o Inno a ion & Knowledge 10 (2025) 100637
5
Resul s om he semi s uc u ed in e iews
We used an in-dep h, semi-s uc u ed in e iew me hod ha allowed
us o change he o de o ques ions o ask abou he ele an indings o
one in e iew o ano he . To o ganize, analyze, and isualize semi-
s uc u ed da a, we used ATLAS. i, a compu e -aided quali a i e da a
analysis so wa e. Table 5 in Appendix C p o ides an o e iew o he
in e iew pa icipan s. The da a we e analyzed induc i ely using a
g ounded heo y app oach while e lec ing on he esea ch ques ion.
A e ansc ibing he i s wo in e iews, we s a ed coding he quo-
a ions, which is also known as open coding. Coding in ol es gi ing
sho labels o g oups o da a. We added codes and compa ed hem o he
p io da a a e inishing he ansc ip ion o each in e iew, ensu ing
ha he da a coding and he analy ic p ocess we e consis en . We
ini ially had 102 open codes, bu a e mul iple i e a ions and ocusing
on he esea ch ques ion, we eached sa u a ion and ended up wi h 51
i s -o de codes. The 11 h and 12 h in e iew ansc ip ions did no
yield any new codes, bu we u ilized he quo es om hem o b oaden
he s udy’s applicabili y.
In Fig. 4 (Appendix D), we see a Sankey diag am ep esen ing he
codes gene a ed om he di e en in e iew pa icipan s. Pa icipan s 1
and 7 had mo e codes, and he commonali y be ween hem was ha hey
had implemen ed MLOps in mul iple o ganiza ions. Mo eo e , i clea ly
shows ha he las wo in e iews did no gene a e any new codes. In he
nex s ep, we combined ela ed i s -o de concep s in o ele en ocus
codes o second-o de hemes. The second-o de hemes we e hen
ca ego ized in o ou agg ega ed dimensions (Fig. 5, Appendix D). Fig. 6
(Appendix D) p o ides an imp ession o ou codebook. The s udy’s
indings will comp ise a ho ough na a i e suppo ed by da a ha uses
2nd-o de hemes and agg ega ed dimensions, equen ly quo ing he
i s -o de s a emen s o he in o man s.
Following analysis, we c ea ed a ypology in Fig. 2 ha e lec ed he
pe inen hemes (blue) benea h he agg ega e dimensions (yellow).
Al hough hey a e no displayed he e, he coding sys em (Fig. 5 in Ap-
pendix D) con ains he i s -o de no ions.
O ganiza ional challenges
As shown in Table 2, he da a was dis illed in o high-densi y codes
o o ganisa ional p oblems, which comp ise ou opics ha a e elab-
o a ed upon below.
Human Resou ces and Skillse : The a ea o machine lea ning is
ela i ely new o he co po a e wo ld, posing unique esou ce chal-
lenges. The ac ha esou ces employed o machine lea ning engi-
nee ing equen ly lack ele an expe ience is one o he main p oblems.
I ’s possible ha wha hey lea ned in school o college won’ apply in
p ac ical si ua ions. This misma ch be ween academic aining and in-
dus ial demands is a p oblem since i akes a lo o ime o hese
esou ces o become up o speed and acqui e he knowledge and
expe ise needed o accep esponsibili y o da a.
“We ha e been acing he p oblem o he las 5–6 yea s ega ding he
esou ces no being eady, and some imes wha hey ha e lea ned in
college o school is no p ac ical enough. So, he onboa ding ime o
machine lea ning enginee s is 6 o 7 mon hs. Only a e six o se en
mon hs can you us ha hey can ake da a owne ship,”- Pa icipan 1
The lack o ele an skills needed o build MLOps pipelines in one
eam o some imes e en wi hin he o ganiza ion can lead o knowledge
gaps.
Use Engagemen and Resis ance: E en i some cu en eam
membe s may esis lea ning new skills because hey lack compe ence, i
is ne e heless c ucial o ain hem.
“Because along wi h ou pipeline came all he new lea ning o do, which
was challenging o some eam membe s o adap o hese new ools.” –
Pa icipan 11. Teams wo king on machine lea ning applica ions migh
no see he bene i s o unused MLOps pipelines. As Pa icipan 7 said, “I
would say he bigges challenge was keeping he eams ha we wo ked wi h,
in ol ed and in e es ed. We spend a lo o ime in ol ing di e en eams and
making hem eel engaged in he p ocess because ou bigges challenge was no
making he model o making he pipeline, bu ge ing people o jus hink abou
using i and making people see he bene i s.”
Resis ance o adop ion can be educed h ough he O ganiza ional
cul u e, as Pa icipan 1 explains, “Cul u e is good he e because i is a da a-
based company, and ha is why he si ua ion is be e ,” demons a ing ha
eams did no oppose hei use o he newes MLOps ools.
Slow P ocesses: One ecu ing opic in alks abou MLOps is he
impo ance o o ganisa ional and eam s uc u e. The success ul
implemen a ion o MLOps ypically equi es he pa icipa ion o se e al
eams. These eams migh , howe e , ha e di e en p io i ies, which
could lead o dependencies ha con inue longe han expec ed. Pa ic-
ipan 1 sha ed, “Ve y o en, i ’s no one eam ha manages his whole hing,
and ha means ha i you, o example, ha e a da a enginee who wo ks wi h
you om a di e en eam, hey’ e also wo king on o he p ojec s. They need
o manage all o hose, and ha means ha hey don’ always ha e ime o
help you ou wi h wha e e you need a ha ime, causing delays.”
Collabo a ion & Communica ion: Ou esul s show ha poo un-
de s anding o machine lea ning models and da a wi hin an o ganisa ion
migh cause p oblems wi h collabo a ion. E ec i e coo dina ion is
di icul since MLOps in ol es a ied eams including da a scien is s,
de elope s, and ML enginee s who ha e di e en p io i ies and objec-
i es. P oblems wi h eamwo k and miscommunica ion can occu when
a eam lacks he essen ial compe encies; Pa icipan 7 sha ed, “Ve y
o en i ’s no one eam ha manages e e y hing which means ha hey don’
always ha e ime o help you ou wi h wha e e you need making collabo-
a ion di icul .”
I can be ha d o communica e changes o a model o non-ML clien s
and o nume ous eams a once.
“You ge a lo o miscommunica ions be ween hose eams and a lack o
accoun abili y o esponsibili y. Because i ’s a lo easie o say. Yeah,
he e’s a p oblem. Tha ’s no my p oblem. Tha ’s hei p oblem. And you
jus s a going in ci cles.” -Pa icipan 7
Technical challenges
O ganisa ions looking o pu MLOps in o p ac ice usually ace a a-
ie y o echnological di icul ies. These di icul ies can be caused by a
a ie y o ac o s, including he in icacy o machine lea ning models, a
lack o da a, o he usion o a ious ools and echnologies. A e dis-
cussing he o ganisa ional p oblems, his has been explained unde he
ollowing h ee hemes. Mo e codes a e p oduced in he da a as a esul ,
c ea ing mo e echnological di icul ies.
In as uc u e and Da a Managemen : The unde lying in as uc-
u e is essen ial o he complica ed ask o building pipelines o machine
Table 2
Code densi y.
Ca ego y Themes Codes To al
densi y
O ganiza ional
Challenges
Human Resou ces and Skillse 7 23
Use Engagemen and
Resis ance
4
Slow P ocesses 8
Collabo a ion &
Communica ion
4
Technical Challenges In as uc u e and Da a
Managemen
7 14
S anda ds and F amewo k 4
Technical ools 3
Ope a ional
Challenges
Di icul ies in Pipeline 4 7
Implemen a ion T ade-o s 3
Business Challenges Business Value 2 7
Cos and Budge Cons ain s 5
C. Am i and A.K. Na ayanappa
Jou nal o Inno a ion & Knowledge 10 (2025) 100637
6
lea ning models. Dependable in as uc u e is equi ed o he p oduc-
ion implemen a ion o machine lea ning models. To un models a
scale, o ganisa ions need o make su e ha he deploymen en i onmen
can suppo he ex a p ocessing and s o age equi emen s. I migh be
di icul o manage dispe sed sys ems, maximise ha dwa e e iciency,
and scale up esou ces. In eg a ion p oblems a ise i he in as uc u e
con igu a ion does no ma ch he echnical s ack needed o machine
lea ning applica ions. In MLOps, e icien da a go e nance and man-
agemen a e essen ial. The con iden iali y, quali y, and accessibili y o
he da a used o in e ence and aining mus be gua an eed by o ga-
nisa ions. As Pa icipan 6 says, “I hink he big challenge we ha e is he
eno mous a ia ion in da a. As I men ioned, we ha e hal a million models, so
i ’s e y ha d. I he da a se s a e e y di e en , hen he da a quali y is e y
di e en . Cus ome s a e no in e es ed in low-quali y esul s”. The quali y o
MLops depends on he da a supplied by he pipeline; Pa icipan 3
elabo a es, “The en i e MLOps ou pu depends on how good you da a
quali y is and a he no on MLOps.”
S anda ds and F amewo k: The e seems o be a “gap in he bes
p ac ices p o ided by indus y leade s and he de elopmen o he echnol-
ogy”, says Pa icipan 10. S anda ds and guidelines should be es ab-
lished by leade s in he indus y and solu ion supplie s, bu sadly, his
isn’ he case e e ywhe e a he momen . Teams ind i challenging o
comp ehend and use open-sou ce ools e icien ly when he e is a dea h
o clea and easily accessible documen a ion, which can esul in
inconsis en esul s and p oblems du ing implemen a ion.
W i ing uni es s is s anda d p ocedu e in adi ional so wa e
de elopmen o gua an ee code quali y. Howe e , because machine
lea ning models a e non-de e minis ic, es ing becomes mo e di icul in
his en i onmen . The s anda disa ion o es ing me hodologies is made
mo e di icul by he need o unique me hods and echniques o es ing
machine lea ning sys ems, which go beyond simple so wa e es ing.
“Fo egula so wa e, I’m used o w i ing uni es s. Fo machine lea ning,
hese es s a e a bi mo e complex because i ’s no de e minis ic.” – Pa ici-
pan 3
Technical Tools: Ha ing he app op ia e echnological ools is one
o he mos impo an needs o implemen ing MLOps, acco ding o
se e al in e iewees. Two in e iewees claim ha al hough he e a e
many ools accessible, i can be di icul o selec he bes one depending
on business equi emen s. Fu he mo e, he cos o loca ing a quali y
MLOps pla o m can be high.
Technical ool challenges d aw a en ion o he challenges ha o -
ganisa ions encoun e . The e icacy and e iciency o MLOps p ocedu es
migh be hampe ed by p oblems including ins abili y, poo in eg a ion,
es ic ed capabili ies, and he equi emen o cus omisa ion. This
demons a es how impo an i is o ca e ully assess and es ools, keep
ab eas o new de elopmen s, and ake he o ganiza ion’s MLOps-
speci ic objec i es and limi s in o accoun .
Ope a ional challenges
Two hemes o ope a ional challenges ha a e mo e ocused on day-
o-day ope a ions and closely ela e o echnological challenges su aced
du ing he in e iew.
Deploymen P ocess: As was al eady es ablished, he e a e ce ain
echnical di icul ies wi h da a, ools, and in as uc u e in he deploy-
men p ocess. Bu in e iewees ha e also poin ed up ce ain ope a ional
di icul ies. One such ope a ional di icul y is he inc eased complexi y
o con inuous in eg a ion in MLOps as opposed o De Ops. The in-
e iewees s a ed ha MLOps necessi a es a mo e in ica e and ho ough
me hod o con inuous in eg a ion.
Au oma ing and gene alising he deploymen p ocess p esen s
ano he challenge. I is di icul o c ea e a deploymen s a egy ha
unc ions well o se e al da ase s wi h di e en s a is ical p ope ies.
Ca e ul planning and esea ch a e needed o au oma e he aining, al-
go i hm selec ion, ine- uning, and o he p ocesses o achie e accep able
esul s ac oss a a ie y o da ase s.
In MLOps, elying on o he eams can p o ide di icul ies. One
in e iewee s a ed ha hei eam depends on o he eams’ deploymen
pipelines and API hos ing p o ide s. Any p oblems hese ex e nal de-
pendencies may expe ience could ha e a big e ec on how he MLOps
eam ope a es.
Implemen a ion ade-o s: One o he mos impo an hings ha
he in e iewees men ioned is ha ML de elope s o en ha e o make
sac i ices while implemen ing MLOps. When hey ha e o choose be-
ween con lic ing p io i ies, like cos and o ecas speed, hey encoun e
di icul ies.
Businesses need o weigh he expenses o o - he-shel p oduc s
agains hei dependabili y and bene i s o in eg a ion. O e -
enginee ing solu ions is a common end in MLOps, and i can lead o
was e ul use o ime and esou ces. As a esul , i is impe a i e o
con inuously op imise and enhance; ne e heless, he e should be
easonable limi a ions o p e en o e -enginee ing, which can lowe
e iciency.
Business challenges
Acco ding o some esponden s, a common obs acle o adop ing
MLOps p ojec s is esis ance om en e p ises. Ge ing suppo and buy-
in a he highes le el is c ucial, pa icula ly because he p ocesses
in ol ed a e ge ing mo e complica ed. Explaining he e u n on in-
es men o businesses can be di icul , hough.
Howe e , some companies end o hink o AI as a panacea. This
iewpoin is equen ly s a ed by non- echnical s akeholde s who don’
ge he con ex and jus use buzzwo ds. This kind o hinking can make i
ha de o con ol expec a ions and in o m he company abou he con-
s ain s and p ac ical esul s o MLOps e o s. Rega ding a "simple
desc ip i e analy ics p ojec ," one o he in e iewees ad ises agains he
o ganisa ion sea ching o AI/ML solu ions. C ea ing a dashboa d
should be su icien ins ead.
Consume s ypically ha e high expec a ions o he sys em, coun ing
on comple e p oblem-sol ing and 100% accu acy. I can be di icul o
con ol hese expec a ions, he e o e i ’s impo an o in o m clien s
abou he dange s and es ic ions associa ed wi h machine lea ning
models.
Cos and Budge Cons ain s: Companies equen ly unde alue
MLOps, which leads o inadequa e budge alloca ion o his c ucial ole.
An esponden said ha occasionally, money is spen mo e on p oduc
ma ke ing han on in as uc u e imp o emen s. This makes i di icul
o upda e ools o unde lying legacy sys ems since he company can
adop a sunk cos allacy.
Pa icipan 7 explains ha Managemen and Business may suppo
MLOps un il a budge is eques ed., “When we s a ed, i was open and ee
o us o expe imen . I didn’ eel any es ic ions. La e , when we posed he
ques ion: could we ha e a managed Kube low en i onmen ? And hen we did
eel he es ic ions in e ms o budge .”
MLOps s De Ops
MLOps is a De Ops ex ension ha concen a es on machine lea ning
model deploymen . The inse ion o componen s speci ic o he model
deploymen is he p ima y dis inc ion be ween he wo. The di icul y is
in using cu en De Ops p ocedu es while comp ehending and adjus ing
o hese no el componen s. Simila o De Ops, MLOps has challenges in
building a s ong a mosphe e and p omo ing eamwo k. I ’s challenging
o make su e ha di e en pa ies coo dina e hei ac ions, in eg a e
seamlessly, and communica e e ec i ely. Howe e , because machine
lea ning models in oduce un amilia and sophis ica ed me hodologies,
MLOps p esen s special challenges. Planning and ca e ul conside a ion
a e necessa y o handling and accoun ing o hese isks.
C. Am i and A.K. Na ayanappa
Jou nal o Inno a ion & Knowledge 10 (2025) 100637
7
“MLOps is De Ops ha is applied o machine lea ning. Bu he challenge
is he e a e a lo o unknown p ocesses ha make i a bi mo e di icul .”-
Pa icipan 9
In he nex sec ion we discuss he abo e indings in he con ex o
published li e a u e in mo e de ail.
Discussion
As we s a ou discussion, we conside he p ima y cha ac e is ics
and hemes indica ed in Fig. 2 as we analyse ou indings. We hen go
o e he implica ions, es ic ions, and sugges ions o addi ional s udy.
We also go o e how he indings connec o he e iew o he li e a u e.
O ganisa ional challenges
O ganisa ional con ex has been widely employed in IS/IT esea ch
pape s (Gaskin e al., 2018) and echnology adop ion models (Gangwa
e al., 2015). Resea ch shows ha many IS/IT implemen a ions ail due
o use esis ance (Kim & Kankanhalli, 2009), ools, skills o o ganisa-
ional cul u e (Bunke e al., 2008). The o ganisa ional con ex ocuses
on desc ip i e measu es, which include, among o he hings, esou ce
a ailabili y and skill in use, i m scope, i m size, slack esou ces, social
in luences, cul u e, s uc u al con igu a ions, and manage ial belie s
(Awa e al., 2017).
In his s udy, o ganisa ional challenges consis o ou hemes, as
shown in Fig. 2 o he p e ious sec ion. The issue o ha ing ewe human
esou ces who know ML and da a enginee ing is a ecognised challenge
bo h in he expe in e iew as well as esea ch pape s. To ully ha ness
he ad an ages o MLOps, businesses ace challenges in accessing
pe sonnel expe ienced in a i icial in elligence and machine lea ning
(Zhang e al., 2020) and s opping he u no e o IT p o essionals who
ha e o he pe cei ed job oppo uni ies (Joseph e al., 2007).
This lack o expe ise makes i di icul o companies o align hei
MLOps s a egy wi h o e all goals (Painoli & Da ika, 2021), and he
sho age o ully specialised da a enginee ing alen wi hin he human
esou ces depa men u he exace ba es he si ua ion. The limi ed
educa ional ou pu and he insu icien expe ise le el o g adua es ail
o mee indus y expec a ions, pa icula ly in e ms o quali y o skills,
wi h a s ong emphasis on he enginee ing aspec (Tambu i, 2020). This
con i ms he lack o da a and ML enginee ing alen and he long
onboa ding ime in e iewees men ioned.
Pa icipan s in ou in e iews explained he ime aken o keep he
use s engaged, as some o hem did no pe cei e he use ulness o
MLOps. Employee esis ance o echnology implemen a ion is ecog-
nised in he li e a u e (Lapoin e & Ri a d, 2005) which may be due o
indi idual issues, o ganisa ional issues, sys em issues, p ocess- ela ed
ac o s (Klaus & Blan on, 2010) o he complex na u e o wo k
(Aube e al., 2008). The applica ion o ML models o people wi h li le
knowledge o da a science is qui e challenging, and hence employees
may be esis an o change when i comes o es ablished co po a e
p ocedu es (Baie e al., 2019; Painoli & Da ika, 2021).
I migh be di icul o c ea e machine lea ning AI sys ems o
knowledge-in ensi e wo kplaces due o he need o change how an
o ganisa ion alues inno a ion and welcomes new ways o doing hings.
This shi in mindse is c ucial o MLOps implemen a ion because i
helps businesses o e come eluc ance o change and success ully
inco po a e AI in o hei ope a ions (Zhang e al., 2020).
In e iew pa icipan s explained he delays in implemen ing MLOps
due o eams ha ing di e en p io i ies, o ganisa ions wi h longe
app o al chains, and people’s conse a i e mindse s. Baie e al. (2019)
alk abou how digi alisa ion, in gene al, is slowe in he heal hca e
indus y as some da a is no e en a ailable in digi al o ma . Hence he
p ocess o da a collec ion o build MLOps becomes e y slow. Usually,
da a scien is s and enginee s a e no pa o he same eam, which b ings
challenges o dependencies and wai imes. Paleyes e al. (2022) sugges
including bo h oles in he same eam o a oid such a dependency.
Communica ion and collabo a ion issues we e ecognised by mul i-
ple in e iew pa icipan s, which a e e lec ed in he gene ic da a sci-
ence challenge (Cao, 2017) as well as in De Ops adop ion challenges
(Lassenius e al., 2015). G anlund e al. (2021) b ie ly discuss he need
o seamless communica ion be ween eams and collabo a ions ac oss
he o ganisa ion, which is a he a di icul challenge like ha o
De Ops. In e iew pa icipan s ela e his issue o ha ing less knowl-
edge abou models o machine lea ning among eam membe s hey mus
collabo a e wi h. Baie e al. (2019) ecognise he challenges in
cus ome communica ion along wi h expec a ion managemen as cus-
ome s wan anspa ency in he models, which a e complex o explain.
The oo cause o he ou O ganiza ional challenges (in Fig. 2) can
be aced back o a lack o o ganiza ional eadiness and s a egic
alignmen o MLOps implemen a ion. To add ess hese challenges, o -
ganiza ions should adop a comp ehensi e app oach ha includes
in es ing in alen de elopmen h ough aining p og ams and collab-
o a ions wi h educa ional ins i u ions (Tambu i, 2020), implemen ing
change managemen s a egies wi h clea communica ion o MLOps
bene i s (Kim & Kankanhalli, 2009), c ea ing c oss- unc ional eams ha
in eg a e da a scien is s and enginee s (Paleyes e al., 2022), and
es ablishing pla o ms o knowledge sha ing ac oss he o ganiza ion
(G anlund e al., 2021) Addi ionally, p omo ing a cul u e o inno a ion
and con inuous lea ning is c ucial o o e coming esis ance o change
and os e ing be e collabo a ion (Zhang e al., 2020). By ocusing on
hese solu ions, o ganiza ions can c ea e an en i onmen conduci e o
MLOps adop ion, leading o mo e success ul implemen a ions and be e
u iliza ion o machine lea ning echnologies (Baie e al., 2019).
Technical challenges
The e is a ce ain deg ee o complexi y in e e y echnology, whe e
complexi y is de ined as he pe cei ed di icul y in lea ning and imple-
men ing a sys em (Sonnenwald e al., 2001). Implemen ing MLOps is
di icul because o da a quali y and a ailabili y issues, as s a ed by
Paleyes e al. (2022). Pa icipan s con i med ha MLOps implemen a-
ion should begin only a e making su e ha he e is enough da a
a ailable o c ea e models.
E en he mos ad anced ML echnology canno be e ec i ely
le e aged wi hou an es ablished da a in as uc u e and he high-
quali y da a i deli e s (Shollo e al., 2022). Acco ding o he case
s udies examined by Paleyes e al. (2022), da a- ela ed wo ies a e a
majo impedimen o implemen ing machine lea ning models. Incom-
ple e, skewed, o alse in o ma ion may lowe he quali y o machine
lea ning models, and his is a common p oblem o Machine lea ning
and no speci ic o MLOps implemen a ion. One o he key issues wi h
ML model deploymen is in as uc u e. Shollo e al. (2022) sugges ha
based on he a ailabili y o he echnical in as uc u e and he p ocess
ma u i y, businesses should adap hei machine lea ning app oach.
Se ing up pe inen da a in as uc u es is a challenge in addi ion o
deploying in as uc u es o unning ML models (Baie e al., 2019).
Fu he mo e, as Painoli e al. (2021) and Zhang e al. (2020) poin ou ,
ackling hese da a and model di icul ies o en in ol es conside able
echnical skills, which may be di icul o en e p ises o acqui e.
The exis ing li e a u e ocuses mo e on echnical challenges such as
da a a ailabili y, da a d i , model e sioning, scalabili y, and model
moni o ing (Baie e al., 2019; Lima e al., 2022; Tes i e al., 2022).
Howe e , in in e iews, i was ound o be mo e abou he in eg a ion o
ools in o exis ing in as uc u e, managing da a p i acy, and no ha ing
enough s anda disa ion o MLOps ools. Da a science, acco ding o Cao
(2017), could be enhanced o inco po a e social issues, including p i-
acy, secu i y, and us . The e needs o be mo e s anda disa ion o he
applica ion o MLOps, i s ools, and documen a ion wi hin o ganisa ions
and in he indus y. Since he esea ch is sca e ed and isola ed and ML
models, da ase s, and ea u e se s can be p oduced using nume ous ools
on a wide ange o pla o ms, lack o s anda disa ion is one o he mos
C. Am i and A.K. Na ayanappa
Jou nal o Inno a ion & Knowledge 10 (2025) 100637
8
Fig. 6. An imp ession o he Codebook.
Re e ences
Ab ahamsson, P., Jedli schka, A., Nguyen Duc, A., Felde e , M., Amasaki, S., &
Mikkonen, T. (2016). De Ops adop ion bene i s and challenges in p ac ice: A case
s udy. P oduc -Focused so wa e p ocess imp o emen (pp. 590–597). Sp inge
In e na ional Publishing AG. h ps://doi.o g/10.1007/978-3-319-49094-6_44
Alla, S., & Ada i, S. K. (2020). Wha is MLOps?. Beginning MLOps wi h MLFlow (pp.
79–124). Ap ess L. P.. h ps://doi.o g/10.1007/978-1-4842-6549-9_3
Aube , B. A., Ba ki, H., Pa y, M., & Roy, V. (2008). A mul i-le el, mul i- heo y
pe spec i e o in o ma ion echnology implemen a ion. In o ma ion Sys ems Jou nal,
18(1), 45–72. h ps://doi.o g/10.1111/j.1365-2575.2007.00279.x
Awa, H. O., Ukoha, O., & Igwe, S. R. (2017). Re isi ing echnology-o ganiza ion-
en i onmen (T-O-E) heo y o en iched applicabili y. The Bo om Line, 30(1), 2–22.
h ps://doi.o g/10.1108/BL-12-2016-0044
Baie , L., J¨
oh en, F., & Seebache , S. (2019). Challenges in he deploymen and ope a ion
o machine lea ning in p ac ice. In , 1. ECIS. h ps://doi.o g/10.5445/i /
1000095028
Bansal, P., Smi h, W. K., & Vaa a, E. (2018). New ways o seeing h ough quali a i e
esea ch. Academy o Managemen Jou nal, 61(4). h ps://hal.a chi es-ou e es.
/hal-02312197.
Bei in, B. (2012). In e iew and sampling: How many and whom The SAGE handbook o
in e iew esea ch: The complexi y o he c a , 243–253.
Bunke , D., Kau z, K., & Anh uan, A. (2008). An explo a ion o in o ma ion sys ems
adop ion: Tools and skills as cul u al a e ac s - he case o a managemen
in o ma ion sys em. Jou nal o In o ma ion Technology, 23(2), 71–78. h ps://doi.
o g/10.1057/palg a e.ji .2000134
Cale a o, F., Lanubile, F., & Qua an a, L. (2022). A p elimina y in es iga ion o MLOps
p ac ices in gi hub. I haca: Co nell Uni e si y Lib a y. h ps://doi.o g/10.1145/
3544902.3546636. a Xi .o g.Re ie ed om Publicly A ailable Con en Da abase
h ps://sea ch.p oques .com/doc iew/2718002010.
Cao, L. (2017). Da a science: Challenges and di ec ions. Communica ions o he ACM, 60
(8), 59–68. h ps://doi.o g/10.1145/3015456
Ca doso Sil a, L., Rezende Zaga i, F., Sil a Se e, B., Nildaimon dos San os Sil a, L.,
Luc edio, D., Fu ado Sil a, D., & de Medei os Caseli, H (2020). Benchma king
machine lea ning solu ions in p oduc ion. In 2020 19 h IEEE In e na ional Con e ence
on Machine Lea ning and Applica ions (ICMLA) (pp. 626–633). IEEE. h ps://ieeexplo
e.ieee.o g/documen /9356298.
Choudha y, F., & K ensky, P. (2022). Hype cycle o da a science and machine lea ning
2022. June 29. Ga ne h ps://www.ga ne .com/.
Conway, M. E. (1968). How do commi ees in en . Da ama ion, 14(4), 28–31.
Ebe , C., Galla do, G., He nan es, J., & Se ano, N. (2016). De Ops. IEEE So wa e, 33(3),
94–100. h ps://doi.o g/10.1109/MS.2016.68
Diaz-De-A caya, J., To e-Bas ida, A. I., Z´
a a e, G., Mi˜
n´
on, R., & Almeida, A. (2023).
A join s udy o he challenges, oppo uni ies, and oadmap o MLOps and AIOps: A
sys ema ic su ey. ACM Compu ing Su eys, 56(4), 1–30.
E ich, F. M. A., Am i , C., & Dane a, M. (2017). A quali a i e s udy o De Ops usage in
p ac ice. Jou nal o So wa e: E olu ion and P ocess, 29(6). h ps://doi.o g/10.1002/
sm .1885. n/a.
Flaounas, I. (2017). Beyond he echnical challenges o deploying machine lea ning
solu ions in a so wa e company.10.48550/a xi .1708.02363.
Gangwa , H., Da e, H., & Ramaswamy, R. (2015). Unde s anding de e minan s o cloud
compu ing adop ion using an in eg a ed TAM-TOE model. Jou nal o En e p ise
In o ma ion Managemen , 28(1), 107–130. h ps://doi.o g/10.1108/JEIM-08-2013-
0065
Ga g, S., Pundi , P., Ra hee, G., Gup a, P. K., Ga g, S., & Ahlawa , S. (2022). On
con inuous in eg a ion /con inuous deli e y o au oma ed deploymen o machine
lea ning models using MLOps. In 2021 IEEE ou h in e na ional con e ence on a i icial
in elligence and knowledge enginee ing (AIKE) (pp. 25–28). IEEE.
Gaskin, J., Be en e, N., Lyy inen, K., & Rose, G. (2018). Inno a ion among di e en
classes o so wa e de elopmen o ganiza ions. In o ma ion Sys ems Jou nal, 28(5),
849–878. h ps://doi.o g/10.1111/isj.12171
Gioia, D. A., Co ley, K. G., & Hamil on, A. L. (2013). Seeking quali a i e igo in
induc i e esea ch. O ganiza ional Resea ch Me hods, 16(1), 15–31. h ps://doi.o g/
10.1177/1094428112452151
G anlund, T., Kopponen, A., S i bu, V., Myllyaho, L., & Mikkonen, T. (2021). MLOps
challenges in mul i-o ganiza ion se up: expe iences om wo eal-wo ld cases.1
0.48550/a xi .2103.08937.
G ego , S. (2006). The na u e o heo y in in o ma ion sys ems. MIS Qua e ly, 611–642.
He manowicz, J. C. (2013). The longi udinal quali a i e in e iew. Quali a i e Sociology,
36, 189–208.
Hui Kang, Le, M., & Shu, Tao (2016). Con aine and mic ose ice d i en design o cloud
in as uc u e De Ops. In 2016 IEEE In e na ional Con e ence on Cloud Enginee ing
(IC2E) (pp. 202–211). IEEE. h ps://ieeexplo e.ieee.o g/documen /7484185.
John, M. M., Olsson, H. H., & Bosch, J. (2021). Towa ds MLOps: A amewo k and
ma u i y model. In 2021 47 h Eu omic o Con e ence on So wa e Enginee ing and
Ad anced Applica ions (SEAA) (pp. 1–8). IEEE. h ps://ieeexplo e.ieee.o g/documen
/9582569.
Jo dan, M. I., & Mi chell, T. M. (2015). Machine lea ning: T ends, pe spec i es, and
p ospec s. Science (Ame ican Associa ion o he Ad ancemen o Science), 349(6245),
255–260. h ps://doi.o g/10.1126/science.aaa8415
Joseph, D., Ng, K., Koh, C., & Ang, S. (2007). Tu no e o in o ma ion echnology
p o essionals: A na a i e e iew, me a-analy ic s uc u al equa ion modeling, and
model de elopmen . MIS Qua e ly, 31(3), 547–577. h ps://doi.o g/10.2307/
25148807
C. Am i and A.K. Na ayanappa
Jou nal o Inno a ion & Knowledge 10 (2025) 100637
15
Khan, M. S., Khan, A. W., Khan, F., Khan, M. A., & Whangbo, T. K. (2022). C i ical
challenges o adop De Ops cul u e in so wa e o ganiza ions: A sys ema ic e iew.
IEEE Access, 10, 14339–14349. h ps://doi.o g/10.1109/ACCESS.2022.3145970
Kim, H., & Kankanhalli, A. (2009). In es iga ing use esis ance o in o ma ion sys ems
implemen a ion: A S a us Quo Bias pe spec i e. MIS Qua e ly, 33(3), 567–582.
h ps://doi.o g/10.2307/20650309
Ki chenham, B. A. (2004). P ocedu es o pe o ming sys ema ic e iews, 33. Keele, UK:
Keele Uni e si y, 2004.
Klaus, T., & Blan on, J. E. (2010). Use esis ance de e minan s and he psychological
con ac in en e p ise sys em implemen a ions. Eu opean Jou nal o In o ma ion
Sys ems, 19(6), 625–636. h ps://doi.o g/10.1057/ejis.2010.39
K euzbe ge , D., Kühl, N., & Hi schl, S. (2023). Machine Lea ning Ope a ions (MLOps):
O e iew, de ini ion, and a chi ec u e. IEEE Access, 11, 31866–31879. h ps://doi.
o g/10.1109/ACCESS.2023.3262138
K ey, M., Kabbou , A., Osmani, L., & Saliji, A. (2022). De ops adop ion: Challenges &
ba ie s. In Pape p esen ed a he 55 h Hawaii In e na ional Con e ence on Sys em
Sciences (HICSS), Vi ual, 3-7 Janua y 2022 (pp. 7297–7309).
Lapoin e, L., & Ri a d, S. (2005). A mul ile el model o esis ance o in o ma ion
echnology implemen a ion. MIS Qua e ly, 29(3), 461–491. h ps://doi.o g/
10.2307/25148692
Lassenius, C., Dingsøy , T., & Paasi aa a, M. (2015). De Ops: A de ini ion and pe cei ed
adop ion impedimen s. Agile p ocesses in so wa e enginee ing and ex eme
p og amming (pp. 166–177). Sp inge In e na ional Publishing AG. h ps://doi.o g/
10.1007/978-3-319-18612-2_14
Lei e, L., Rocha, C., Kon, F., Milojicic, D., & Mei elles, P. (2020). A su ey o De Ops
concep s and challenges. ACM Compu ing Su eys, 52(6), 1–35. h ps://doi.o g/
10.1145/3359981
Lima, A., Mon ei o, L., & Fu ado, A. (2022). MLOps: P ac ices, ma u i y models, oles,
ools, and challenges – A sys ema ic li e a u e e iew. In Pape p esen ed a he
P oceedings o he 24 h In e na ional Con e ence on En e p ise In o ma ion Sys ems -
Volume 2: ICEIS (pp. 308–320). h ps://doi.o g/10.5220/0010997300003179
Liu, Y., Ling, Z., Huo, B., Wang, B., Chen, T., & Mouine, E. (2020). Building a pla o m o
machine lea ning ope a ions om open sou ce amewo ks. IFAC Pape sOnLine, 53
(5), 704–709. h ps://doi.o g/10.1016/j.i acol.2021.04.161
Lwaka a e, L. E., C nko ic, I., & Bosch, J. (2020). De Ops o AI - Challenges in
de elopmen o AI-enabled applica ions. In 2020 in e na ional con e ence on so wa e,
elecommunica ions and compu e ne wo ks (So COM) (pp. 1–6). IEEE. h ps
://ieeexplo e.ieee.o g/documen /9238323.
Lwaka a e, L. E., Kilamo, T., Ka onen, T., Sau ola, T., Heikkil¨
a, V., I konen, J.,
Ku aja, P., Mikkonen, T., Oi o, M., & Lassenius, C. (2019). De Ops in p ac ice: A
mul iple case s udy o i e companies. In o ma ion and So wa e Technology, 114,
217–230. h ps://doi.o g/10.1016/j.in so .2019.06.010
Maca hy, R. W., & Bass, J. M. (2020). An empi ical axonomy o De Ops in p ac ice. In
Pape p esen ed a he 2020 46 h Eu omic o Con e ence on So wa e Enginee ing and
Ad anced Applica ions (SEAA) (pp. 221–228).
Makinen, S., Skogs om, H., Laaksonen, E., & Mikkonen, T. (2021). Who needs MLOps:
Wha da a scien is s seek o accomplish and how can MLOps help?. In 2021 IEEE/
ACM 1s Wo kshop on AI Enginee ing-So wa e Enginee ing o AI (WAIN) (pp.
109–112). h ps://ieeexplo e.ieee.o g/documen /9474355.
Ma sui, B. M. A., & Goya, D. H. (2022). MLOps: Fi e s eps o guide i s e ec i e
implemen a ion. In P oceedings o he 1s In e na ional Con e ence on AI Enginee ing:
So wa e Enginee ing o AI (pp. 33–34). h ps://ieeexplo e.ieee.o g/documen
/9796456.
Maya, V., & Felipe, A. (2021). The S a e o MLOps h p://hdl.handle.ne /1992/51495.
Mucha, T., Ma, S., & Abha i, K. (2022). Beyond MLOps: The li ecycle o machine
lea ning-based solu ions. In AMCIS 2022 P oceedings. h ps://amcis2022.aiscon e e
nces.o g/.
Painoli, G., & Da ika, V. (2021). A i icial in ellegence in business-bene i s and
challenges. Tu kish Online Jou nal o Quali a i e Inqui y, 12, 1377–1388.
Paleyes, A., U ma, R., & Law ence, N. D. (2022). Challenges in deploying machine
lea ning: A su ey o case s udies. ACM Compu ing Su eys, 55(6), 1–29. h ps://doi.
o g/10.1145/3533378
Qu, S. Q., & Dumay, J. (2011). The quali a i e esea ch in e iew. Quali a i e Resea ch in
Accoun ing and Managemen , 8(3), 238–264. h ps://doi.o g/10.1108/
11766091111162070
Ru , P., Madan, M., Reich, C., & Ould-Abdeslam, D. (2021). Demys i ying MLOps and
p esen ing a ecipe o he selec ion o open-sou ce ools. Applied Sciences, 11(19),
8861. h ps://doi.o g/10.3390/app11198861
Rzig, D. E., Hassan, F., & Kessen ini, M. (2022). An empi ical s udy on ML De Ops
adop ion ends, e o s, and bene i s analysis. In o ma ion and So wa e Technology,
152, A icle 107037. h ps://doi.o g/10.1016/j.in so .2022.107037
Sch ¨
ode , T., & Schulz, M. (2022). Moni o ing machine lea ning models: A ca ego iza ion
o challenges and me hods. Da a Science and Managemen , 5(3), 105–116. h ps://
doi.o g/10.1016/j.dsm.2022.07.004
Sculley, D., Hol , G., Golo in, D., Da ydo , E., Phillips, T., Ebne , D., Chaudha y, V.,
Young, M., C espo, J., & Dennison, D. (2015). Hidden echnical deb in machine
lea ning sys ems. In Pape p esen ed a he P oceedings o he 28 h In e na ional
Con e ence on Neu al In o ma ion P ocessing Sys ems - Volume 2 (pp. 2503–2511).
Se ban, A., Blom, K. .d., Hoos, H. H., & Visse , J. M. W. (2020). Adop ion and e ec s o
so wa e enginee ing bes p ac ices in machine lea ning. In P oceedings o he 14 h
ACM/IEEE In e na ional Symposium on Empi ical So wa e Enginee ing and
Measu emen (ESEM) (pp. 1–12). h ps://hdl.handle.ne /1887/3307600.
Shanka , S., Ga cia, R., Helle s ein, J.M., & Pa ameswa an, A.G. (2022).
Ope a ionalizing machine lea ning: An in e iew s udy. a Xi p ep in 10.4855
0/a xi .2209.09125.
Shollo, A., Hop , K., Thiess, T., & Mülle , O. (2022). Shi ing ML alue c ea ion
mechanisms: A p ocess model o ML alue c ea ion. The Jou nal o S a egic
In o ma ion Sys ems, 31(3), A icle 101734. h ps://doi.o g/10.1016/j.
jsis.2022.101734
Sonnenwald, D. H., Maglaughlin, K. L., & Whi on, M. C. (2001). Using inno a ion
di usion heo y o guide collabo a ion echnology e alua ion: Wo k in p og ess. In
P oceedings Ten h IEEE In e na ional Wo kshop on Enabling Technologies: In as uc u e
o Collabo a i e En e p ises. WET ICE (pp. 114–119). IEEE. h ps://ieeexplo e.ieeFe.
o g/documen /953399.
Sub amanya, R., Sie la, S., & Vya kin, V. (2022). F om De Ops o MLOps: O e iew and
applica ion o elec ici y ma ke o ecas ing. Applied Sciences, 12(19), 9851. h ps://
doi.o g/10.3390/app12199851
Symeonidis, G., Ne an zis, E., Kazakis, A., & Papakos as, G. A. (2022). MLOps -
De ini ions, ools and challenges. In 2022 IEEE 12 h Annual Compu ing and
Communica ion Wo kshop and Con e ence (CCWC) (pp. 0453–0460). IEEE.
Mboweni, T., Masombuka, T., & Dongmo, C. (2022). A sys ema ic e iew o machine
lea ning De Ops. In Pape p esen ed a he - 2022 In e na ional Con e ence on Elec ical,
Compu e and Ene gy Technologies (ICECET) (pp. 1–6). h ps://doi.o g/10.1109/
ICECET55527.2022.9872968
Tambu i, D. (2020). A. Sus ainable MLOps: T ends and challenges. In 2020 22nd
in e na ional symposium on symbolic and nume ic algo i hms o scien i ic compu ing
(SYNASC) (pp. 17–23). IEEE. h ps://ieeexplo e.ieee.o g/documen /9356947.
Tanzil, M. H., Sa ke , M., Uddin, G., & Iqbal, A. (2023). A mixed me hod s udy o De Ops
challenges. In o ma ion and So wa e Technology, 161, A icle 107244. h ps://doi.
o g/10.1016/j.in so .2023.107244
Tes i, M., Ballabio, M., F on oni, E., Iannello, G., Moccia, S., Soda, P., & Vessio, G.
(2022). MLOps: A axonomy and a me hodology. IEEE Access, 10, 63606–63618.
h ps://doi.o g/10.1109/ACCESS.2022.3181730
To na zky, L. G., & Klein, K. J. (1982). Inno a ion cha ac e is ics and inno a ion
adop ion-implemen a ion: A me a-analysis o indings. IEEE T ansac ions on
Enginee ing Managemen , EM-29(1), 28–45. h ps://doi.o g/10.1109/
TEM.1982.6447463
VB S a . (2019). Why do 87% o da a science p ojec s ne e make i in o p oduc ion?.
Wohlin, C. (2014). Guidelines o snowballing in sys ema ic li e a u e s udies and a
eplica ion in so wa e enginee ing. In P oceedings o he 18 h in e na ional con e ence
on e alua ion and assessmen in so wa e enginee ing (pp. 1–10).
Zhang, Z., Nandhakuma , J., Hummel, J., & Waa denbu g, L. (2020). Add essing he key
challenges o de eloping machine lea ning AI sys ems o knowledge-in ensi e wo k.
MIS Qua e ly Execu i e, 19(4), 221–238. h ps://doi.o g/10.17705/2msqe.00035
C. Am i and A.K. Na ayanappa
Jou nal o Inno a ion & Knowledge 10 (2025) 100637
16
