Attribution Methodology

If you cannot say what would have happened without the AI, you have not measured its value — you have measured a coincidence.

Most organizations are stuck at the coincidence stage. Only 39% attribute any level of EBIT impact to AI, and most of those report under 5% of EBIT; roughly six in ten cannot tie any EBIT to AI at all, with only about 6% qualifying as high performers who attribute more than 5% of EBIT to it (McKinsey, 2025; McKinsey, 2025). The picture from other surveys is just as stark: 74% of companies have yet to show tangible value from AI, and only 26% have moved past proof-of-concept to generate value — of which a mere 4% have built cutting-edge cross-functional capability (BCG, 2024). Gartner projected that at least 30% of gen-AI projects would be abandoned after proof of concept by the end of 2025, citing unclear business value among the causes (Gartner, 2024). And in one of the bluntest findings, 95% of enterprise gen-AI pilots delivered no measurable P&L impact (MIT Project NANDA, 2025).

This page is the bridge from "we deployed AI" to "AI created $X, and here is how we know." It covers the attribution problem, the methods that actually isolate impact, the difference between direct and indirect value, how to convert measurement into a defensible investment case, and the errors that quietly invalidate the whole exercise.

The Attribution Problem¶

Attribution is hard for four structural reasons, and each one defeats a naive "we turned it on and revenue went up" story.

Confounding variables. Revenue, cycle time, and quality move for dozens of reasons at once — a pricing change, a new hire cohort, a seasonal swing, a competitor stumble. Any of these can masquerade as AI impact, or mask it.
The counterfactual is unobservable. You only ever see the world where the team had AI. You never directly see the same team, same quarter, without it. Every credible attribution method is, at heart, a way to estimate that missing counterfactual.
Hard-to-monetize outcomes. Much of AI's early value shows up as quality, speed, employee experience, or capability — things that are real but resist a clean dollar figure. Deloitte found that 41% of organizations have struggled to define and measure the exact impacts of their gen-AI efforts (Deloitte, 2024).
Long and lumpy payback. Productivity gains often precede financial results by quarters. Cutting the analysis off too early shows "no impact" when the value simply has not landed in the P&L yet.

The combined effect is the credibility gap above: the technology can be working at the desk while the boardroom sees nothing it trusts.

Methods for Isolating Impact¶

There is a hierarchy of attribution methods, ranked by how well each one reconstructs the counterfactual. Use the strongest method your context allows, and be explicit about which one you used — it determines how much confidence the number deserves.

Rank	Method	Counterfactual logic	Use when	Strength
1	RCT / A-B test	Randomly assign AI access; the control group is the counterfactual	You can randomize at the user or task level	Strongest — isolates causation
2	Holdout / control group	A comparable non-randomized group that does not get AI	Randomization is impractical but a similar cohort exists	Strong, if groups are truly comparable
3	Difference-in-differences (DiD)	Compare the change in treated vs. untreated groups over time	You have before/after data for both a treated and an untreated group	Strong, if the parallel-trends assumption holds
4	Before/after (pre-post)	The same group before vs. after	Nothing else is available	Weakest — no control for confounders

Why randomization wins. In a randomized controlled trial, the only systematic difference between groups is the AI itself, so any outcome gap is causal. The landmark GitHub Copilot RCT did exactly this: developers using Copilot completed the task 55.8% faster — 1 hour 11 minutes versus 2 hours 41 minutes — in a randomized trial of 95 developers (Peng et al., 2023). A pre-post comparison could never have ruled out "the developers just got better at the task"; randomization does.

Difference-in-differences and the parallel-trends assumption. When you cannot randomize, DiD compares the change in the treated group to the change in an untreated group over the same window. It subtracts out anything that affected both groups equally — seasonality, macro shifts. The critical condition is parallel trends: the two groups must have been moving in step before the intervention, so that the untreated group's post-period change is a fair estimate of what the treated group would have done absent the AI. The canonical example is the study of a customer-support gen-AI assistant, which raised productivity (issues resolved per hour) by 15% on average and 30% for the least-experienced workers, identified via difference-in-differences on a staggered rollout across roughly 5,172 agents (Brynjolfsson et al., 2025). The staggered rollout is what made it work — agents who had not yet received the tool served as the control for those who had.

Attribution must respect where AI helps. A pre-registered RCT of 758 BCG consultants found that those using GPT-4 completed 12.2% more tasks, 25.1% faster, with 40% higher quality — but only on tasks inside the AI's "jagged frontier." On tasks outside that frontier, AI users were 19 percentage points less likely to be correct (Dell'Acqua et al., 2023). The lesson for attribution: an aggregate average can hide both a strong positive and a real negative. Segment your measurement by task type, or you will attribute a blended number that is true on average and wrong everywhere.

Direct vs. Indirect Attribution¶

Measured value splits into two families, and mature programs report both — while being honest about which is which.

Direct attribution captures lagging financial outcomes that flow to the P&L:

Time saved (hours returned to higher-value work)
Cost reduced (lower cost-to-serve, fewer vendor hours, deflected tickets)
Revenue gained (faster sales cycles, higher conversion, new AI-enabled offerings)

These are what the CFO wants, because they are denominated in dollars and tie to the income statement.

Indirect attribution captures leading indicators that predict future financial value but do not yet appear as cash:

Quality (defect rates, rework, accuracy)
Satisfaction (customer CSAT/NPS, employee experience)
Capability (skills built, faster onboarding, broader adoption)
Option value (new capabilities that open future opportunities)

The CFO-comfort problem. Indirect value is real but uncomfortable for finance, because it is not yet money. This is why so few programs close the loop with the office of the CFO: only 16% of organizations produce regular reports to the CFO on the value created by gen AI (Deloitte, 2024). The fix is not to abandon indirect metrics — they are your early-warning system — but to (a) report them as leading indicators with an explicit hypothesis of how they convert to financial outcomes, and (b) never count them as dollars until they show up as direct value.

Building the Investment Case¶

The job here is to convert measured productivity into financial terms a CFO will sign, then deliberately discount it.

The productivity-to-dollars formula (standard practice):

Annual $ value =
    hours saved per user
  × fully-loaded hourly labor cost
  × number of active users
  × adoption / utilization rate
  × realization rate (haircut)

Two terms carry most of the rigor:

Fully-loaded hourly labor cost is not base salary. It is salary + benefits + payroll taxes + overhead (facilities, equipment, management), converted to an hourly rate. Using base salary alone understates the value; using a market bill rate often overstates it. Pick one definition and document it.
Active users, not licenses. A purchased license is not a user, and a user is not a daily user. The single most common way these models inflate is by multiplying by seats sold rather than people actively getting value. Drive the model off observed adoption and utilization, not procurement.

The realization rate (haircut) is the deliberate discount that turns a theoretical maximum into a defensible claim. Not every saved hour is reinvested in productive work; some leaks to slack. Conservative haircutting is what makes finance trust the number. Forrester's Total Economic Impact (TEI) methodology formalizes this with four components — Benefits, Costs, Flexibility, and Risk — and systematically adjusts benefits down and costs up (in one study, benefits were cut 10% and costs raised 10%), then applies a discount rate typically in the 8–16% range to account for the time value of money (Forrester, TEI methodology). Adopt the same posture: present a conservative case, not a ceiling.

For external benchmarking, organizations realized an average return of $3.70 for every $1 invested in generative AI (IDC, 2024) — useful as a sanity check on your own modeled return, not as a substitute for measuring it.

A worked sketch: 100 hours saved per user per year × $75 fully-loaded hourly cost × 400 active users × 70% utilization × 60% realization = $1.26M per year. Drop the 60% haircut, and the same model reads $2.1M; swap "active users" for "licenses sold" and it inflates further still — which is exactly how investment cases lose credibility.

Common Attribution Errors¶

Correlation ≠ causation. "Revenue rose after we deployed AI" is not attribution; it is a timeline. Without a counterfactual, it proves nothing.
No baseline. If you did not capture the pre-AI metric, you cannot compute a delta. Baseline before you deploy, not after.
No control. A pre-post number with no comparison group cannot separate AI from everything else that changed that quarter.
Ignoring value leakage. Time saved that is not redeployed to valuable work — a saved hour that gets absorbed by meetings — is not realized value. Haircut for it.
Conflating share-of-code (or share-of-output) with acceptance, or acceptance with value. "X% of code is AI-generated" is an activity metric, not an outcome. Accepted suggestions are not the same as faster delivery, which is not the same as business value.
Double-counting. Counting the same hour saved in two teams' cases, or counting indirect value as both a leading indicator and as direct dollars, inflates the total. Each unit of value belongs to exactly one line.

Key Takeaways¶

Attribution is counterfactual estimation. Every defensible number answers "what would have happened without the AI?" — pick the strongest method (RCT > holdout > DiD > pre-post) your context allows.
The landmark studies prove the methods work — Copilot's 55.8% RCT speedup (Peng et al., 2023) and the 15%/30% DiD support-productivity result (Brynjolfsson et al., 2025) are templates, not just headlines.
Segment by where AI helps. A blended average can hide a real negative outside the "jagged frontier" (Dell'Acqua et al., 2023).
Report direct and indirect value separately, and never count leading indicators as dollars — that is what builds CFO trust, which today is rare (Deloitte, 2024).
Build the case conservatively. Drive the model off active users, use fully-loaded labor cost, apply a realization-rate haircut, and adjust benefits down / costs up in the spirit of TEI (Forrester, TEI methodology).
Avoid the silent invalidators — no baseline, no control, value leakage, share-of-output vanity metrics, and double-counting.

Companion pages: start with the measurement framework for the metric model this attribution sits on top of, then use the standing up AI measurement practitioner guide to instrument the data, and the measurement maturity scoring assessment to gauge readiness. For where the productivity actually originates, see the workflow optimization framework; for how attributed value feeds prioritization, see the AI strategy framework.

Sources¶

McKinsey — The State of AI, 2025. About 39% of organizations report any enterprise-level EBIT impact from AI; most of those say less than 5% of EBIT is attributable to AI use. View source · verified 2026-06-20 · primary
McKinsey — The State of AI, 2025. Respondents who attribute EBIT impact of 5 percent or more to AI use and say their organization has seen significant value from AI — about 6 percent of respondents — are defined as AI high performers. View source · verified 2026-06-20 · primary
BCG — Where's the Value in AI?, 2024. Only 26% of companies have developed the necessary set of capabilities to move beyond proofs of concept and generate tangible value (4% cutting-edge plus 22% advanced). View source · verified 2026-06-20 · primary
Gartner — Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025, 2024. At least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs or unclear business value. View source · verified 2026-06-20 · primary
MIT Project NANDA — The GenAI Divide: State of AI in Business 2025, 2025. Just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact. View source · verified 2026-06-20 · ⚠ secondary mirror
Deloitte — The State of Generative AI in the Enterprise: Now decides Next (Q3), 2024. 41% have struggled to define and measure the exact impacts of their GenAI efforts. View source · verified 2026-06-20 · primary
Peng et al. — The Impact of AI on Developer Productivity: Evidence from GitHub Copilot, 2023. The treatment group with access to the AI pair programmer completed the task 55.8% faster than the control group; developers with GitHub Copilot took on average 1 hour 11 minutes versus 2 hours 41 minutes, across 95 professional programmers. View source · verified 2026-06-20 · primary
Brynjolfsson, Li & Raymond — Generative AI at Work (Quarterly Journal of Economics), 2025. Access to AI assistance increases issues resolved per hour by 15% on average, with a 30% increase for less-skilled and less-experienced workers, using data from 5,172 customer-support agents. View source · verified 2026-06-20 · primary
Dell'Acqua et al. — Navigating the Jagged Technological Frontier (HBS Working Paper 24-013), 2023. In a pre-registered experiment with 758 consultants, those using AI inside the frontier completed 12.2% more tasks, 25.1% more quickly, with more than 40% higher quality; on a task outside the frontier AI users were 19 percentage points less likely to be correct. View source · verified 2026-06-20 · primary
Deloitte — The State of Generative AI in the Enterprise: Now decides Next (Q3), 2024. Only 16% have produced regular reports for the CFO about the value being created with GenAI. View source · verified 2026-06-20 · primary
Forrester — Total Economic Impact (TEI) Methodology. Forrester's Total Economic Impact methodology evaluates investment value across four components (Benefits, Costs, Flexibility, Risk), risk-adjusts the model (reducing benefits and increasing costs for uncertainty), and discounts future cash flows. View source · verified 2026-06-20 · primary
IDC — The Business Opportunity of AI (IDC InfoBrief, sponsored by Microsoft), 2024. For every $1 a company invests in generative AI, the average ROI is 3.7x. View source · verified 2026-06-20 · ⚠ secondary mirror