The 95% number,
and the word carrying it

Rung 1, and what it argues against is the default assumption of the period: that model capability is the constraint on enterprise value, so a better model produces a better outcome. The claim — usually quoted as roughly 95% of enterprise AI pilots showing no measurable impact on profit and loss, though the report's headline and funnel count organisations — is an argument that the constraint is somewhere else entirely.

The whole sentence turns on one word, and reading it carelessly is how the number gets misused. "Measurable" is doing an enormous amount of work. A pilot can change how people work, reduce time on a task, or produce a capability the organisation keeps, and still not show up in a P&L line — because the effect was not instrumented, because no baseline was taken before it started, because the benefit landed as capacity rather than cost, or because the attribution was never agreed. "No measurable impact" is a joint statement about the intervention and the measurement, and the failure can sit on either side.

Which is what makes it the commercial argument for forward-deployed engineering rather than an argument against the technology. If capability is abundant and outcomes are rare, then the scarce work is the part between them — integration, workflow change, instrumentation, and agreeing in advance what would count as evidence. That is a description of a job, and the number is the clearest available case for why someone pays for it.

Two words decide what this supports. Measurable: the finding is about attributable, demonstrated effect on a financial statement — not about whether users liked the tool, whether the model was accurate, or whether anyone felt more productive. And pilot: a bounded trial, not a production deployment, which selects for exactly the projects most likely to end before an accounting period closes. Note too what the report counts: its funnel is of organisations, not pilots — 20% piloted a task-specific tool and 5% got one into production, so about one in four of those that piloted succeeded.

The first two are the reason a forward deployed engineer exists as a role. Both are solved by work that happens before and after the model — discovery that finds a workflow with a number attached, and instrumentation that makes the change attributable. Neither is model work, and neither gets done by shipping an API.

Read alongside a16z’s services-led-growth argument, the picture is coherent rather than contradictory. a16z’s thesis is that companies should deliberately accept lower gross margin on deployment work because the resulting customer outcome is defensible.⁸ The NANDA-style finding is what the market looks like when nobody does that work: capability is available, outcomes are not, and the gap between the two is a services gap rather than a model gap.

The figure comes from The GenAI Divide: State of AI in Business 2025, by Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari of MIT's Project NANDA. It is dated July 2025, covers research from January to June 2025, and describes itself as preliminary findings. It is a report, not a peer-reviewed paper.

1. Collect three kinds of evidence. A systematic review of more than 300 publicly disclosed AI initiatives and announcements; structured interviews with representatives of 52 organizations; and survey responses from 153 senior leaders gathered at four industry conferences. Company data and quotes are anonymized.
2. Split the tools into two classes. General-purpose LLM tools such as ChatGPT and Copilot, and embedded or task-specific GenAI: enterprise systems, custom-built or bought from a vendor, aimed at a particular workflow. The headline failure rate is about the second class.
3. Follow organizations through a funnel. For each class, the share of organizations that investigated a tool, piloted one and successfully implemented one.
4. Define success. A task-specific tool counts as successfully implemented when users or executives describe a marked and sustained effect on productivity or P&L. The methods note adds deployment beyond the pilot phase with measurable KPIs, and judges ROI six months after the pilot, adjusted for department size.
5. Score industries. A disruption index combines five publicly observable indicators to rank sectors; the authors test other weightings and report that the top and bottom of the ranking hold.
6. Compare how deployments were staffed. Interviewees' projects are grouped as strategic partnerships (buy), internal development (build) or a hybrid of the two, and their success rates compared.

The cause it proposes: a learning gap

The report's own explanation is what it calls the learning gap. Its interviews describe enterprise tools that do not retain feedback, adapt to context or improve over time, and users who happily draft with ChatGPT but will not trust it, or a brittle internal tool, with work that depends on remembering past corrections. On this reading the barrier is neither model quality nor regulation nor infrastructure but whether the system learns inside the workflow, and the authors present agentic systems with persistent memory as the way across. Their secondary findings fit that story:

• Buying beats building. In the interview sample, external partnerships reached deployment about 67% of the time and internal builds about 33%.
• A shadow AI economy. Only 40% of companies said they had bought an official LLM subscription, but workers at more than 90% of surveyed companies reported regular use of personal AI tools for work.
• Budgets favor the front office. Asked to split a hypothetical $100 of GenAI budget, executives put about half into sales and marketing, while the clearest savings the report documents came from back-office work, such as cutting outsourced processing and agency spend.
• Disruption is narrow. Only technology, and media and telecom, show clear structural change.
• Mid-market moves faster. Top mid-market firms reported about 90 days from pilot to full implementation; enterprises, defined as firms with over $100 million in revenue, took nine months or more.

That is a different cut from the three failures named on this page, measurement, attribution and genuine failure, and the two fit together: the learning gap is one cause of genuine failure, while the other two are how a system that worked can still land in the 95%.

The headline figures, with what each rests on:

Read with its own caveats, the report is narrower than its headline. The 95% appears three ways, as organizations getting zero return, as integrated pilots with no measurable P&L impact, and as custom tools that never reach production, and the funnel behind it counts organizations: the 5% that reached production is a quarter of the 20% that piloted. Success rests on what users and executives said, inside a six-month window after the pilot. The authors flag that their figures are directional, that organizations willing to discuss AI problems may differ from those that declined, that success definitions vary across organizations, and that the buy-versus-build gap may reflect the capabilities of the organizations that bought rather than the approach itself. The text is also not fully consistent: it gives sales and marketing about half of the hypothetical budget in one place and about 70% in another, and counts the sectors as eight in one place and nine in another.

This is arithmetic on the report's own funnel figures, which are shares of organizations, not of pilots or dollars.

Per 100 organizations, 20 piloted a task-specific tool and 5 reached production, so 15 pilots stalled. That one funnel supports three different failure rates: 75 percent of piloting organizations did not reach production, 91.7 percent of investigating organizations did not, and 100 − 5 = 95 percent of all organizations had no task-specific tool in production, a denominator that includes the 40 that never looked. The headline 95 matches only the last reading, which is why the same report gets quoted for very different claims.

Two more divisions keep the figure honest. Because 95 percent is a share of organizations, it does not by itself say what fraction of the $30–40 billion went to the organizations that failed. And the buy-versus-build result, about 67 against 33 percent reaching deployment, is a ratio of about 2.0 measured on 52 interviewed organizations, small enough that a few cases move it. The general-purpose column cannot be pooled with the other: its 80 percent pilot-to-implementation rate measures adoption, mostly individual productivity, not a measurable effect on P&L.

Acting on the report means spending effort before and around the model rather than on it. Each choice that makes an outcome provable also gives something up, and it is better to name the cost to the customer up front than to discover it at renewal.

• A metric that cannot move inside the window

  The report judged return about six months after the pilot. A number whose movement arrives a year later, such as annual churn, will read as zero however well the system works. Pair it with a leading indicator that responds within weeks and that the owner agrees is tied to it.

• Counting freed hours as savings

  Time saved lands as capacity, and capacity shows up on no financial statement until something changes: overtime falls, a contractor is not renewed, the backlog clears. Agree in advance what the freed time will turn into, or the pilot works and still joins the 95%.

• Comparing against no AI at all

  Workers at more than 90% of surveyed companies already used personal AI tools. The real baseline is often people pasting work into a consumer chatbot, so measure the official tool against that, and study the shadow use first; it shows which tasks people actually want help with.

• A baseline that does not match the after-picture

  Measure a quiet month before and a busy month after, or count a case differently once the system routes them, and the comparison is noise. Fix the definition, the sample and the season of both measurements when the baseline is taken.

• Reading the learning gap as a model problem

  The report's explanation is that tools do not retain corrections or context inside the workflow, not that models are too weak. A larger model leaves that untouched; feedback capture, persistent context and a route for corrections address it. The argument rests on interviews, so test it on your own deployment.

Start with what kind of document it is: a research report, and the most-cited statistic from it travelled far beyond the context that produced it. As with any survey figure, the answer depends on who was asked, how a "pilot" was defined, what window was allowed for impact to appear, and who decided whether impact was measurable. Those definitions decide the number, and a figure this round and this quotable should be treated as directional rather than precise.

It does not support "AI does not work" — which is the reading it is most often recruited for. A pilot showing no measured P&L movement is equally consistent with the technology working and the organisation being unable to capture or demonstrate it, and that ambiguity is intrinsic rather than a flaw in the reporting. It also does not support any claim about a specific deployment, yours included.

What it does support is a change in how you run one, and this is the part worth keeping whatever happens to the statistic. Agree the measurement before the build. Name the metric, take the baseline while the old process is still running, and get the person who owns the number to agree in advance that a movement in it would count. A pilot with no pre-agreed measurement cannot succeed — not because it will not work, but because there is no state of the world in which it can be shown to have worked. Most of the 95% is likely to be that, and it is avoidable for free.

Work each one out before you open it — a number, a prediction or a reason, not a definition. Producing the answer is what makes it stick. Your progress is kept on this device.

Not research papers: a consultancy census, a VC essay and a job-posting analysis. Included because their figures get quoted as if they were, and read here for their method.

1. 1The 95% Number, Examined2026
2. 2Trading Margin for Moat2026
3. 31,000 FDE Jobs, Analysed2026

All 110 papers, in 19 ladders →