Almost every enterprise can now point to an AI pilot that "worked". Far fewer can point to a line in the management accounts that moved because of it. McKinsey's 2025 global survey found that 39% of respondents attribute any EBIT impact to AI, and that most of those put the figure at under 5% of EBIT. BCG, surveying 1,000 senior executives in late 2024, put it more bluntly: 74% of companies had yet to show tangible value from their use of AI. The problem is not only that value is scarce. It is that AI ROI measurement, as most enterprises practise it, is set up to find value where it is easiest to count and to miss it where it actually accrues.
This piece proposes a three-tier value ledger for AI ROI measurement: cost, revenue and risk, each with its own evidence standard, its own time horizon and its own point in the deployment lifecycle at which it becomes legitimate to count. The ledger is an analytical framework, not a survey finding. Its purpose is to stop two common errors at once: the pilot that claims savings it has not yet delivered, and the production system whose real value sits in a tier nobody is measuring.
The argument is aimed at CFOs and transformation leads who must sign off on AI budgets and defend them a year later. It draws on the McKinsey and BCG surveys, on EY's November 2025 survey of Indian enterprise leaders, and on the MIT NANDA report that generated the widely quoted claim that 95% of enterprise GenAI pilots deliver no measurable P&L impact, a claim this piece treats with care.
Why AI ROI measurement goes wrong at the pilot stage
A pilot is designed to answer a technical question: does the system do the task well enough? It is rarely designed to answer the financial one: does doing the task well enough change a number the enterprise already reports? The conflation of the two is the root of most inflated AI business cases.
Consider the standard pilot result: "the model classifies 92% of documents correctly and reduces handling time by 60%". Both numbers may be true. Neither is value. Handling time becomes value only if the freed hours are redeployed, released or avoided in hiring, and only if the 8% error rate does not create rework or risk that eats the saving. A pilot cannot show this because it does not run long enough, at enough volume, inside enough of the real process.
The evidence that this gap is systemic is consistent across sources. EY India found that 47% of surveyed Indian enterprises say more than 10% of their PoCs moved to production, which means a majority saw fewer than one in ten pilots graduate. Menlo Ventures' December 2025 survey of enterprise buyers reported 47% of AI deals reaching production, up from 25% a year earlier, an improvement but still roughly a coin toss. The MIT NANDA report of August 2025 claimed about 95% of enterprise GenAI pilots deliver no measurable P&L impact and only about 5% reach production with value. That report's methodology has been criticised, and the 95% figure should be read as a signal of direction rather than a precise estimate. But the direction is the same in every dataset: most pilots do not become measured value.
The ledger below is designed so that an enterprise never has to choose between "we cannot measure it yet" and "we will count the pilot number as if it were real".
The three-tier value ledger
The ledger separates value into three tiers that behave differently and should be counted differently.
Tier 1: Cost. Value that shows up as lower spend on a cost line the enterprise already reports: labour hours, vendor fees, error rework, infrastructure. Cost value is the most measurable and the most often overstated, because the saving is real only when the resource is actually released. The evidence standard is a before-and-after comparison on an existing cost centre, at production volume, over at least one full budget cycle.
Tier 2: Revenue. Value that shows up as higher sales, higher conversion, higher retention, faster time to market or new products. Revenue value is harder to attribute, because many things move revenue at once, and it is the tier where McKinsey finds high performers concentrate: revenue increases are most commonly reported in marketing and sales, strategy and corporate finance, and product and service development. The evidence standard is a controlled comparison, holdout group or A/B test where possible, or a documented attribution model where not.
Tier 3: Risk. Value that shows up as losses avoided: fraud not paid, fines not incurred, outages not suffered, compliance deadlines met. Risk value is the least measured and, in regulated Indian sectors, often the largest. It is also the easiest to inflate, because avoided losses are counterfactual. The evidence standard is a baseline loss rate from history, an observed change in incidents after deployment, and an explicit statement of what else changed.
The table sets out what to count in each tier, at which lifecycle stage counting becomes legitimate, and the standard of evidence required.
Two rules follow from the table. First, nothing is counted at pilot. A pilot produces a forecast of value, labelled as such, with its assumptions written down. Second, each tier is counted on its own evidence standard. Revenue is never counted on a cost-tier standard, which is what happens when a sales team's "productivity gain" is presented as a sales increase.
What the surveys say about where value actually lands
The ledger is more useful when read against what the large surveys find about where value shows up.
McKinsey's 2025 survey reports that cost benefits are most commonly reported in software engineering, manufacturing and IT, while revenue increases are most commonly reported in marketing and sales, strategy and corporate finance, and product and service development. High performers, about 6% of respondents, are those attributing 5% or more of EBIT to AI; they are nearly three times as likely to have fundamentally redesigned individual workflows, three-quarters are scaling AI compared with one-third of others, and over a third commit more than 20% of digital budgets to AI.
BCG's October 2024 study of 1,000 executives found that 26% of companies had built the capabilities to move beyond proofs of concept and generate tangible value, and that just 4% had developed cutting-edge capabilities across functions and consistently generate significant value. BCG's leaders reported 1.5 times higher revenue growth, 1.6 times greater shareholder returns and 1.4 times higher return on invested capital over the preceding three years, and BCG estimates that 62% of AI's value lies in core business functions, with operations at 23%, sales and marketing at 20% and research and development at 13%. BCG also offers a resource rule it calls 70-20-10: 70% of the challenge is people and process, 20% is technology, 10% is the algorithm.
Three implications for measurement follow. The first is that cost value and revenue value live in different functions, so a single enterprise-wide "AI ROI" figure is averaging two unlike things. The second is that the companies reporting the most value are the ones that changed the workflow, not the ones that added a model to an unchanged one, which means the measurement baseline must be the old workflow, not the old task. The third, from BCG's 70- 20-10, is that the cost side of the ledger must include the people and process spend, which is most of it, and not just licences and compute.
A worked example with stated assumptions
The following illustration applies the ledger to a single workflow, trade finance document checking at a mid-sized Indian bank. Every figure is an assumption chosen for arithmetic clarity, not an observation from any institution.
Assume the bank processes 20,000 trade finance document sets a month, each taking 45 minutes of checker time at a fully loaded cost of ₹600 per hour. Monthly checker cost is therefore 20,000 multiplied by 0.75 hours multiplied by ₹600, or ₹90 lakh. Assume a deployed document-intelligence system reduces handling time to 15 minutes for 80% of sets and leaves 20% as exceptions handled at the old 45 minutes. New monthly checker time is 16,000 multiplied by 0.25 plus 4,000 multiplied by 0.75, or 7,000 hours, against 15,000 before. The gross hour reduction is 8,000 hours a month.
On the cost tier, the saving is legitimate only to the extent the 8,000 hours are released. Assume half are redeployed to higher-value work that would otherwise have required hiring, and half simply become slack because the team cannot be reduced below a floor in the short run. Countable cost value is 4,000 hours multiplied by ₹600, or ₹24 lakh a month, less the system's running cost. Assume inference, integration maintenance and support at ₹9 lakh a month. Net cost-tier value: ₹15 lakh a month, or ₹1.8 crore a year.
On the revenue tier, faster turnaround may win trade finance business, but the bank cannot show this without a comparison. Assume it runs the new process for half its corporate clients for six months. If the treated group's trade finance volume grows 4 percentage points faster than the control group, and the margin on the incremental volume is known, that increment and only that increment is entered. If no control is possible, the revenue line is entered as zero with a note, not as a forecast dressed as a result.
On the risk tier, assume the bank's historical rate of documentary discrepancies missed at first check led to an average of ₹40 lakh a year in disputed charges and rework. If the observed rate after deployment falls by half over a comparable twelve months, and no other control changed, ₹20 lakh a year is entered. If the compliance team was also expanded in the same period, the entry is reduced or footnoted.
The point of the exercise is not the numbers. It is that the three tiers produce three different answers from the same deployment, on three different evidence standards, at three different times. The cost tier is countable after a budget cycle. The revenue tier needs a design decision at launch to be countable at all. The risk tier needs a twelve-month baseline. An enterprise that reports "ROI" as a single number six weeks after go-live has, almost by definition, counted only forecast cost savings.
Measurement decisions that must be made before launch
Three of the most consequential measurement choices cannot be made after deployment.
The first is the baseline. If the workflow is being redesigned, as McKinsey's high performers do, the baseline is the old workflow's full cost and outcomes, not the time taken for the one task the model performs. Capturing this requires measuring the old process before it disappears.
The second is the control. Revenue and some risk value can only be shown against a comparison group. Deciding to deploy to half the branches, half the client book or half the product lines is a measurement decision disguised as a rollout decision, and it has to be taken by the business owner, not the data science team.
The third is the cost perimeter. BCG's 70-20-10 implies that the model is a tenth of the spend. The ledger's cost side must include change management, process redesign, integration, data remediation, governance and the ongoing inference bill. A separate PanScience Research piece builds a fully loaded unit-cost model; here the rule is simply that the perimeter is fixed before launch so it cannot be narrowed afterwards to flatter the result.
Where PanScience ventures fit, as practice not proof
PanScience's portfolio offers two company-reported examples of the measurement discipline, offered as illustrations of the ledger rather than as evidence of outcomes. OnFinance AI's NeoGPT is described by the company as a BFSI compliance LLM with 70-plus agents across 40-plus regulators; a compliance system of that kind creates value almost entirely on the risk tier, deadlines met and findings closed, and so is a case where the twelve-month baseline rule and the "what else changed" footnote are the whole measurement problem. Indika AI, described as a data-centric AI and model operations venture with around 98% annotation accuracy, sits on the cost side of the perimeter: data remediation is one of the spend lines the 70-20-10 rule says enterprises forget to include. Neither description should be read as a claim about any client's results.
What a board should ask for
A board or investment committee reviewing an AI programme's return can ask five questions that map directly to the ledger.
Which tier is each claimed value in, and on what evidence standard was it counted? What was the baseline, and was it the old workflow or the old task? Where is the control group for revenue claims, and if there is none, why is revenue on the ledger at all? What is the full cost perimeter, and does it include people, process and inference as well as licences? And which claimed values are forecasts from pilots, labelled as such, rather than observed results from production?
An enterprise that can answer those five questions for each AI deployment will usually find that its total AI value is smaller than the pilots promised and larger than the sceptics assume, and that most of the gap between the two is in the risk tier, which nobody was counting.
Conclusion
AI ROI measurement is failing in most enterprises not because value is absent but because the measurement is done at the wrong stage, on the wrong baseline, with the wrong evidence standard for the tier in question. The three-tier ledger offers a discipline: count cost after a budget cycle on released resources, count revenue only against a control, count risk only against a historical baseline with other changes declared, and count nothing at pilot except a labelled forecast. The surveys that say most enterprises see little value from AI are describing enterprises that have not yet done this. McKinsey's high performers and BCG's 4% are, in part, simply the ones that redesigned the workflow and measured the whole of it.
Methodology and sources
The three-tier value ledger, its evidence standards and the worked trade finance example are original analytical contributions; every number in the worked example is an assumption chosen for clarity and is not drawn from any institution. Observed figures come from the following sources, opened and read for this piece unless marked as verified by the editor. McKinsey, The state of AI in 2025, survey of 1,993 participants in 105 nations fielded 25 June to 29 July 2025 (https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-state-of-ai-2025): 39% attribute EBIT impact, most under 5%; high performers about 6%; workflow redesign, scaling and budget figures; functions where cost and revenue benefits are most reported. BCG press release, 24 October 2024, "AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value" (https://bcg.com/ja-jp/press/24october2024-ai-adoption-in-2024- 74-of-companies-struggle-to-achieve-and-scale-value): 74%, 26%, 4%, 70-20-10, leaders' financial outperformance, 62% of value in core functions, survey of 1,000 executives. EY India, The AIdea of India: Outlook 2026, Chapter 2 survey findings, 16 November 2025 (https://www.ey.com/en_in/insights/ai/agentic-ai-india/aidea-survey-findings): 47% say over 10% of PoCs moved to production, editor-verified. Menlo Ventures, State of Generative AI in the Enterprise 2025, 9 December 2025: 47% of AI deals reach production versus 25% prior, editor-verified from the published report. MIT NANDA, The GenAI Divide: State of AI in Business 2025, August 2025: about 95% of pilots deliver no measurable P&L impact; editor-verified, and the report's methodology has been publicly criticised, so the figure is treated as directional. PanScience Innovations facts about OnFinance AI and Indika AI are company- reported from panscience.xyz, read 24 October 2026, and no client outcomes are claimed.