The true ROI of AI · PhD research
1 / 8
Your AI ROI
is missing
a term.
Dashboards count the time AI saves producing the work. Few count the time it takes to check it.
ROI=
G−C−?
C
The verification bill
review time
rework
what slips through
rework
what slips through
1.06M
merged pull requests, Oct 2024 to Feb 2026
470
organization-owned open-source repositories
1.7×
as long to merge when the change is delegated to AI
≈1.8×
as often sent back by reviewers
Aziz Camara · PhD research on AI coding tools
swipe →
Dashboard ROI
2 / 8
The version in the business case.
ROIdash=
G−C
C
G
Gain
Time saved producing the work, valued at loaded cost.
C
Cost
Licences, tokens, enablement, rollout.
What the deck shows · illustrative
G
C
It assumes the output is finished the moment it is generated.
An AI agent's change is not done when it is written. It is done when someone has checked it.
Aziz Camara · PhD research on AI coding tools
the missing term →
True ROI
3 / 8
Add the verification bill.
ROItrue=
G−C−[ V(φ)+r·R+(1−φ)·E ]the verification bill
C
V(φ)
Verification
The wait for a first human review, plus the review itself. Rises with φ.
r·R
Rework
Share of changes sent back, times the cost of each fix cycle.
(1−φ)·E
Escapes
Output nobody really examined, times what it costs when it fails downstream.
φ
Faithful appropriation
The share of AI output a human genuinely examines before it ships.
Lower φ and the visible bill shrinks. It does not disappear. It moves to E, where no dashboard looks.
Aziz Camara · PhD research on AI coding tools
my model · the record follows →
Where the bill is paid
4 / 8
Every delegated change passes one gate.
What happens there decides where the cost lands.
φ is faithful appropriation: the share of AI output a human genuinely examines before it ships.
Aziz Camara · PhD research on AI coding tools
what the record shows →
The cost at the gate · V
5 / 8
An AI-delegated change vs the same developer's own, in the same repository and month.
1.7×
as long to merge
+64%
longer wait for a first human review
Median wait for the first human review
Own changes, re-weighted to the delegated complexity mix
3.4 h
Delegated to an AI agent
6.4 h
0 h2468 h
Held fixed in the comparison
developer × repository × month
task type
complexity
weekday × hour
Not a night-shift effect. On a business-hours clock, most of the merge-time gap remains.
Calendar time, not reviewer hours. This is V, the first term of the bill.
Aziz Camara · PhD research on AI coding tools
1,064,292 merged PRs · 470 repos · Oct 2024 to Feb 2026
Are people still checking?
6 / 8
On average, yes.
Reviewers behaved as calibrated trust predicts, not as complacency predicts.
≈1.8×
as often sent back
with changes requested
with changes requested
Own changes3.0 in 100
Delegated to AI5.5 in 100
Each square = 1 in 200 pull requests · filled = sent back
−5.5 pts
silent approvals, same reviewer, same repository
No penalty
detected in reverts after merge (equivalence not shown)
φ · Faithful appropriation
Using a tool in line with its spirit (DeSanctis & Poole, 1994). For an AI agent, that spirit is extending a developer's judgment, not replacing it. A change sent back with a reason is faithful use. A silent approval is its absence.
Early warning: the scrutiny gap narrows as developers delegate more (an exploratory reading).
Aziz Camara · PhD research on AI coding tools
same developer · same repo · same month
Faithful vs silent
7 / 8
Same G. Same C.
Different bill.
Different bill.
Cutting the checks is the fastest way to improve AI ROI on paper.
Aziz Camara · PhD research on AI coding tools
illustrative, not estimated
Measure the bill
8 / 8
Scale delegation only as fast as you can measure it.
Compare delegated work with the same people's own work, in the same weeks.
01
Wait for the first human review
V
02
Time to merge
V
03
Share sent back
r·R
04
Share reverted or repaired after merge
E
Set no speed targets on delegated work.
A target on what is easy to observe rewards the reviewer who examines least.
Holmström & Milgrom, 1991
What my data cannot tell you
Whether the time saved producing (G) pays this bill. Production time is not in the record. So measure it. Do not assume it.
The equation
ROItrue=(G−C−[V(φ)+r·R+(1−φ)·E])/C
Aziz Camara · PhD research, Université Paris Dauphine | PSL
GitHub data · Oct 2024 to Feb 2026