AIReturn

AI ROI — the fundamentals

Why 95% of Enterprise AI Pilots Never Reach the P&L (and What the 5% Do)

≈95% of enterprise AI pilots never hit the P&L. The 5% that do measure output net of rework, not adoption. Here's the difference.

Rodrigo Paredes BassiPublished 10 min read

TL;DR

  • The pilot-to-production gap is a measurement gap, not a model gap. Pilots die because no one baselined the work before AI, so no one can prove the work got better after — and boards no longer fund what can't be proven.
  • MIT NANDA's 2025 GenAI Divide report found only 5% of integrated AI systems created significant value — 95% of pilots showed no measurable P&L impact. Directional, but it rhymes with the hard data: S&P found the average organization scraps 46% of proofs-of-concept before production.
  • Independent numbers agree the return is rare: 28% of AI use cases fully meet ROI expectations (Gartner, Apr 2026); roughly 5% of organizations report substantial ROI (IBM); 20% are seeing AI actually increase revenue (Deloitte).
  • The 5% share one habit: they pick an outcome work item, baseline the team before AI, integrate into the real workflow, and measure output net of rework per team — then kill or scale on evidence.
  • The one question that separates them: not "did people adopt it," "does the model work," or "is it compliant," but the 4th question — did the work actually improve, by team, net of rework?

Why do most AI pilots fail to reach the P&L?

Most enterprise AI pilots fail to reach production because they were never designed to be measured. They prove a model can produce plausible output in a demo, but they skip the baseline that would show whether the real work improved — so when the board asks for the return, there is no number. The gap between pilot and P&L is a measurement gap, not a capability gap. That is why the spend keeps rising while confidence falls. The models are good enough. The pilots still stall — not because the AI can't do the task, but because no one instrumented the task well enough to prove it did, net of the cleanup.

Define it: an outcome work item, a per-team baseline, and the 4th question

Three terms carry this argument. Define them once, use them precisely. An outcome work item is a specific, countable unit of work a team actually ships — a resolved support ticket, a merged pull request, a closed deal, a published brief — chosen so that "done" is unambiguous and success ties to a business outcome, not to activity. It is the opposite of a demo task. You can baseline it, count it, and check whether it needed redoing. A per-team baseline is the record of how a specific team performed on its outcome work item before AI — throughput and rework, measured against that team's own history, never against another team's. Output isn't comparable across functions, so a baseline is always a team compared to its earlier self. Without it, there is no "after" to compare against, and no proof. The 4th question is the one nobody answers continuously: did AI actually improve the work — by team, tied to business outcomes, net of rework — and what do I fix? The other three (what did it cost, does it work technically, is it compliant) are answered by FinOps, observability, and governance. Pilots that only answer those three never reach the P&L, because none of them measures whether the work got better.

How bad is the pilot-to-production gap, really?

Bad, and consistent across independent sources. No single figure carries the point — but every serious data source points the same direction, and the most-cited one needs a caveat.

FindingNumberSourceDate
GenAI pilots showing no measurable P&L impact (directional)95% (only 5% create significant value)MIT NANDA, The GenAI DivideAug 2025
Proofs-of-concept scrapped before production46% (avg per org)S&P Global Market Intelligencepub. Oct 2025
Organizations that abandoned most AI initiatives42%, up from 17% YoYS&P Global Market Intelligencepub. Oct 2025
AI use cases that fully meet ROI expectations28% (20% fail outright)Gartner (782 I&O leaders)Apr 2026
Organizations reporting substantial ROI from AI~5%IBM (via WRITER)2026
Organizations seeing AI increase revenue today20% (74% aspire to)Deloitteearly 2026
A necessary caveat on the 95%: MIT NANDA's figure measures pilots with no measurable P&L impact — not "95% of AI failed." It rests on a small qualitative base (52 interviews, 153 survey leaders, 300 public deployments), and the number has been widely contested. We cite it as directional and attributed. Its value is that it is corroborated from three independent directions: IBM's roughly 5% reporting substantial ROI, Deloitte's 20% actually seeing revenue, and Gartner's 28% meeting ROI expectations. Different studies, same signal — value is real but rarely proven.
Notice what the S&P numbers add that MIT's cannot. They are hard survey data, not qualitative interviews, and they measure the exact transition this post is about: the average organization scraps 46% of its POCs before production, and the share abandoning most AI initiatives nearly tripled from 17% to 42% in a single year. The pilots are being killed. The question is whether they are being killed on evidence or on exhaustion.

Why pilots die: four failure patterns

Failed pilots tend to fail the same handful of ways. Each is a variation on measuring the wrong thing.

  • No baseline. The pilot never recorded how the team performed before AI, so "after" has nothing to compare against. Any result — good, bad, flat — is unfalsifiable, and unfalsifiable results don't survive a budget review.
  • Demo tasks, not real work. The pilot proves the model on a curated task in a sandbox, not on the messy outcome work item the team actually ships. It clears the demo and dies on contact with the real workflow.
  • Adoption counted as success. Seats activated, prompts sent, "hours saved" self-reported. These measure activity, not return. High adoption of a high-rework tool is a faster way to lose money, not a win. This is why hours saved isn't ROI.
  • Rework ignored. The pilot books the first-order gain — the draft produced in seconds — and never subtracts the second-order cost of correcting it. In one study, 41% of workers received "workslop" (plausible-looking but substandard AI output), costing an average 1h 56m to fix each time and an estimated ~$186 per employee per month (BetterUp Labs + Stanford, HBR, Sep 2025; self-reported). A pilot that ignores that cost reports a gain that isn't there. The common thread: each pattern lets a pilot look successful without proving the work improved. That is exactly the gap the board has learned to distrust — and it is the whole of the 2026 AI ROI reckoning already underway.

What the 5% do: a five-step playbook

The organizations whose pilots reach the P&L are not using better models. They run a tighter measurement loop. Five steps, in order.

  1. Pick an outcome work item. Choose one countable unit of real work with an unambiguous "done" — resolved tickets, merged PRs, closed deals, shipped briefs — tied to a business outcome. Not a demo task. This is what you will measure.
  2. Baseline the team before AI. Record throughput and rework on that work item for the specific team, against its own recent history. This is the "before" that makes any later claim provable rather than anecdotal.
  3. Integrate into the real workflow. Put the AI where the work actually happens — in the ticketing queue, the IDE, the CRM — not in a side sandbox. A pilot that lives outside the workflow measures a demo, not the job.
  4. Measure output net of rework, per team. Track the useful output the team ships minus the cost of redoing flawed AI work, compared to that team's baseline. Rework is function-specific: code churn and reverts in engineering; reopened tickets in support; stage regression in sales. The number that matters is output-per-dollar-net-of-rework, not raw volume.
  5. Kill or scale on evidence. If the team's rework-adjusted output rose against its baseline, scale and fund it. If it didn't, kill it — fast and without apology. The point of the loop is a defensible decision, not a permanent pilot. The difference between the 95% and the 5% is not step 1 or step 3 — plenty of failed pilots pick real work and wire it in. It is steps 2 and 4: a baseline to compare against, and rework subtracted from the output. Answer the 4th question, per team, and the pilot has a number the board can fund. Skip it, and the pilot joins the 46% that get scrapped.

From pilot to production: the shift in one line

The move from pilot to production is the move from "can the model do this?" to "did the work get better for this team, net of rework, against its own baseline?" The first question is answered in a demo. The second is answered continuously, in the real workflow, per team — and it is the only one a CFO can put in a budget. That is the framework CFOs use to measure and prove AI ROI, and our step-by-step framework for measuring AI ROI walks the loop end to end. If you want to see the output of that loop, here is a sample per-team AI ROI report.

FAQ

Why do most AI pilots fail? Most AI pilots fail to reach production because they were never built to be measured. They prove a model works in a demo but skip the pre-AI baseline, so no one can show the real work improved — and boards won't fund an unprovable return. It's a measurement gap, not a model gap: MIT NANDA found only 5% of integrated AI systems created significant value (Aug 2025, directional). Is the "95% of AI pilots fail" statistic true? It's real but needs care. MIT NANDA's 2025 GenAI Divide report found 95% of GenAI pilots showed no measurable P&L impact — not that 95% of AI "failed." It rests on a small qualitative base and is contested, so treat it as directional. It's corroborated from three sides: IBM's roughly 5% reporting substantial ROI, Deloitte's 20% seeing revenue, and Gartner's 28% meeting ROI expectations. How do you move an AI pilot to production? Pick one countable outcome work item, baseline the team's throughput and rework before AI, integrate the tool into the real workflow, then measure output net of rework against that team's own baseline. Scale what beats the baseline; kill what doesn't. The transition is from "can the model do this?" to "did the work get better, net of rework?" — the only version a CFO can fund. What's the difference between a failed AI pilot and a successful one? Not the model — the measurement loop. Failed pilots count adoption or demo results and ignore rework. Successful ones baseline the team first and subtract the cost of redoing flawed AI output. S&P found the average organization scraps 46% of POCs before production; the survivors are the ones that produced a defensible per-team number, not a better demo. Why does adoption of an AI tool not prove it worked? Because adoption is activity, not return. Seats activated and "hours saved" measure usage, not whether the work improved. A high-adoption tool that generates heavy rework loses money faster, not slower. In one study only 28% of workers said AI improved decision quality even as output volume rose (BetterUp + Stanford, HBR, 2025) — proof that volume and value are different questions.

Sources

See the return on your own numbers.

Get a clear, per-team return-on-AI picture in weeks — output net of rework, reconciled to your invoice.