The MIT AI Report: 95% of AI Pilots Show Zero P&L Impact. The 5% That Pay Do This Instead.

Via MIT Project NANDA: The GenAI Divide: State of AI in Business 2025

What the MIT AI report found: 95% of AI pilots are accounting-neutral

This is the MIT AI study behind the "95% of AI projects fail" headlines. Your firm probably has a ChatGPT subscription, a Copilot trial, and a partner pitching an "AI transformation." Here is the uncomfortable scoreboard: according to The GenAI Divide: State of AI in Business 2025, a working paper from MIT's Project NANDA (July 2025, still widely cited in 2026), roughly 95% of organisations in the sample saw no measurable P&L impact within about six months of a pilot. That is not the same as “95% of pilots failed.” Only about 5% of the custom, task-specific tools MIT tracked reached production with marked and sustained productivity or P&L impact, while the rest stalled earlier or never connected to a number anyone trusted.

The research reviewed 300 publicly disclosed AI initiatives, interviewed representatives from 52 organizations, and surveyed 153 senior leaders. Despite an estimated $30–40 billion in enterprise GenAI investment, the authors describe a stark "GenAI Divide": high adoption on one side, low transformation on the other.

For owner-led SMEs, this is not an excuse to wait. It is a diagnostic. Most firms are not failing because AI is useless. They are failing because they deploy it the way the 95% do — and then wonder why the P&L stays flat.

The funnel underneath the 95%: 60, 20, 5

The headline hides a funnel, and the funnel is more useful than the headline. For custom or vendor-sold, task-specific GenAI tools, MIT reports that about 60% of organisations evaluated such tools, about 20% reached pilot stage, and only about 5% reached production — where MIT defines successful implementation as a marked and sustained productivity and/or P&L impact.

Generic tools look different. MIT puts pilot-to-implementation for general-purpose LLMs such as ChatGPT and Copilot at about 83% — easy to try, flexible, familiar. That does not mean they moved the P&L. It usually means individuals got faster at pieces of work while the firm never turned those minutes into a measured saving. Our 90-day pilot test is built around closing exactly that gap: one finished task, a baseline first, and a kill-or-scale call on a date you set before you start.

Read the funnel plainly: failure concentrates in custom tools that do not learn, do not remember, and do not fit the workflow they were bought for. MIT calls that the learning gap. It is a workflow problem before it is a model problem.

Why pilots stall: four mechanisms MIT actually measured

The headline number is blunt. The why is where the report gets useful — and where most vendor decks go quiet.

1. The learning gap. MIT's lead finding is that the core barrier is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve with use. Employees love ChatGPT for drafting an email; they abandon it for mission-critical work because it forgets yesterday's conversation. Generic chatbots show high pilot rates (~83% implementation for general-purpose LLMs) but mask a deeper split: individual productivity gains that never convert to enterprise P&L.

2. Shadow AI without governance. While roughly 40% of organizations report licensed AI deployment, MIT found that workers at more than 90% of the companies surveyed reported regular use of personal AI tools for work, while only about 40% of companies said they had bought an official LLM subscription. A thriving "shadow AI economy" runs on personal accounts — fast for the individual, invisible on the balance sheet, and impossible to scale safely. Productivity happens in pockets; ROI does not aggregate.

3. Budget aimed at the wrong workflows. MIT estimates roughly half of AI budgets flow to sales and marketing — visible, board-friendly, easy to demo. But the report's successful buyers report higher returns from back-office automation: cutting BPO contracts, reducing external agency spend, streamlining operations, and automating document-heavy processes. Front-office tools get applause; back-office tools get margin.

4. Build fever and missing baselines. In MIT's interview sample, strategic external partnerships reached deployment about 67% of the time versus about 33% for internal builds — a self-reported pattern the report itself warns may reflect buyer capability as much as the approach. Meanwhile, many pilots never establish a pre-deployment P&L baseline — so "no measurable impact" becomes the default finding even when hours were saved. You cannot prove ROI if nobody measured the workflow before the pilot started.

Put simply: firms buy visibility instead of margin, deploy static tools instead of learning systems, and skip the accounting homework. That is why 95% stay on the wrong side of the divide.

What the 5% do differently

MIT's successful cohort — the integrated pilots that did move P&L — share a pattern owner-led firms can adopt without a central AI lab:

• Pick one painful process, not a platform. Winners narrow scope: a specific intake workflow, a BPO replacement, a document pipeline — not "AI everywhere." Lead author Aditya Challapally has noted that high performers "pick one pain point, execute well, and partner smartly."
• Demand workflow fit, not feature lists. Buyers who cross the divide evaluate tools on business outcomes — reduced external spend, faster cycle time, fewer errors — not benchmark scores. Integration with existing systems (CRM, ERP, file stores) is non-negotiable.
• Buy learning-capable systems; don't DIY fragility. Strategic partnerships reached deployment about 67% of the time in MIT's sample, versus about 33% for pure internal builds, with the report's own caveat that capable buyers may choose partnerships more often. Hybrid approaches work when internal teams own the process knowledge and vendors own the learning infrastructure.
• Measure before and after. MIT defines success as "marked and sustained productivity and/or P&L impact" — which requires a baseline. Hours saved, agency fees eliminated, error rates dropped: pick a number and track it.
• Target back-office ROI first. Reported wins include $1M+ annual savings on outsourced risk checks, 10% customer-retention lifts from automated follow-ups, and six-figure BPO reductions — often without headcount cuts. The money shows up as reduced external spend, not layoff theater.

The 5% are not smarter about models. They are smarter about where AI touches money.

Why AI isn't saving your team time (yet)

A common complaint from owners is that everyone uses AI but nobody seems to have more time. Researchers call this the AI productivity paradox, and recent work suggests the productivity stall in many workplaces isn't a mystery.

In an eight-month study at a US technology company of about 200 people, UC Berkeley Haas researchers Xingqi Maggie Ye and Aruna Ranganathan found that generative AI intensified work rather than reducing it. Employees worked faster, took on a broader scope of tasks and let work spill into breaks and evenings, often without being asked. Fortune's coverage of the same theme called it the AI productivity paradox: more work, not less.

For an owner-led firm, the lesson matches the MIT findings above. Does AI increase productivity? It can, but only when the time saved on a specific, mapped task is measured and deliberately redirected. Otherwise it's absorbed by extra work, extra checking or polished output that creates rework for colleagues, which Harvard Business Review has called workslop. How does AI increase productivity in practice? Pick one task, price one finished task before and after (our AI ROI calculator does the arithmetic), and decide in advance what the freed hours are for.

Why do AI projects fail?

The MIT GenAI Divide report, often searched as the "MIT AI study 95" or "MIT 95% of AI projects fail" study, points to four mechanisms covered above: tools that don't learn, shadow AI on personal accounts, budgets aimed at visible front-office projects, and missing baselines. In small firms the same AI implementation challenges show up as no named owner, an unmapped process and no before number. Our guide to AI mistakes businesses make and our AI hype vs reality page cover the owner-side version.

Is the 95% real? What MIT counted, what critics say, and what we conclude anyway

Short answer: the 95% is real as MIT defined it, and narrower than almost every headline that quoted it. It counts organisations, not pilots: about 95% of organisations in the sample saw no measurable P&L impact within about six months of a pilot. It does not say 95% of pilots failed, and it does not say the tools produced nothing. Work can get faster while profit stays unmeasured — that gap is the report's point.

What MIT actually counted

• Sample: a systematic review of more than 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organisations, and survey responses from 153 senior leaders, gathered January to June 2025.
• Status: The GenAI Divide is a July 2025 working paper from MIT's Project NANDA. Its own notes say the views are the authors', not those of affiliated employers. Treat it as NANDA research carrying the MIT name, not as a peer-reviewed, MIT-wide census.
• Success bar: deployment beyond pilot with measurable KPIs — for task-specific tools, a marked and sustained productivity and/or P&L impact, with ROI assessed about six months after the pilot.
• Its own limits: the appendix warns the sample may not represent every segment, organisations willing to talk may differ from those that declined, success definitions varied, and six months may be too short for slower enterprise rollouts, potentially understating success.

Where the criticism lands

Critics have made three fair points, and a careful reader should keep all three. First, the funnel maths: if about 20% piloted custom tools and about 5% reached production, then roughly one in four actual pilots succeeded under a high, short-window bar — 80,000 Hours and NewMR both argue “95% of pilots fail” is therefore the wrong sentence. Second, launch coverage inflated the method: early reporting described far larger interview and employee counts than the PDF's 52 organisations and 153 survey responses, and those larger numbers are still repeated. Third, Wharton legal scholar Kevin Werbach and other reviewers have called the 95% derivation hard to trace in the body of the report and the line between “unsuccessful” and “zero return” too vague. Those are criticisms of precision and packaging. They are not evidence that pilots usually pay.

One phrase to retire: “pilot purgatory” never appears in the MIT report. Vendors and later commentary coined it. We use it on our pilot purgatory page as plain search language for a real condition — a pilot with no baseline, no owner and no decision date — never as a quote from MIT.

What we conclude anyway

Discount the headline and the pattern underneath still holds in MIT's own data: custom tools stall between evaluation and production (60% to 20% to 5%), generic tools spread without aggregating into profit, staff cross the divide on personal accounts while official projects stall (the shadow AI gap below), and the clearest savings sit in back-office and external-spend work rather than visible front-office demos. That is enough to act on. It is not enough to quote “95% fail” in a board deck.

For how the 95% compares with the 80% and 40% figures people quote alongside it, see 95%, 80%, 40%: what each AI failure number actually measures. For two firms that counted cost before quality, see Klarna and Commonwealth Bank rehiring people after AI cuts and Ford rehiring veteran engineers.

How many AI projects fail?

Different figures circulate. The MIT working paper's headline is that roughly 95% of organisations in MIT's sample saw no measurable P&L impact within about six months of a pilot. You will also see "Gartner: 85% of AI projects fail" quoted widely; we could not trace that line to a current primary Gartner source, so treat it with caution. The useful point isn't the exact percentage. It's that most AI projects fail for the same fixable reasons, and a firm can check itself against them with our 20-minute AI readiness assessment.

What to do at your size: the test the funnel suggests

You do not need a central AI lab to run MIT's bar honestly. You need one task, one baseline and one decision date.

1. Score the task before you buy. Run candidates through the AI use-case scorer: time taken, repetitiveness, clean inputs, how easily errors are caught, and what happens when the tool is wrong.
2. Give the pilot 90 days, not an open ending. Our 90-day pilot-to-production test sets the baseline in week one, counts finished work weekly, and forces a kill, scale or redesign call on day 90. MIT's six-month window is the outer bar; a small firm should know sooner.
3. Price the finished task, not the seat. The AI ROI calculator works in cost per finished task — minutes today, minutes with AI including review, plus tool cost — which is the number the 5% can point to.
4. Close the shadow gap before you scale. If staff are already getting value from personal tools, sanction that work properly: the shadow AI risk audit for small firms, the data-exposure side in what staff may be feeding free AI, and the controls on the AI governance checklist.
5. Check readiness honestly. If there is no owner, no mapped process and no before number, the AI readiness assessment will say so in about 20 minutes — cheaper than learning it from an invoice.

Not sure which of your tasks would survive a six-month test? Talk it through with Ada — free, no signup. Tell her the task and how long it takes today; you will get a straight answer on where AI fits, where it should not touch, and the next step, on the way to a free report. No paid audit is pitched before that report.

Frequently asked questions

What is the MIT GenAI Divide study?

A July 2025 working paper from MIT's Project NANDA, The GenAI Divide: State of AI in Business 2025, which found that roughly 95% of organisations in its sample saw no measurable P&L impact within about six months of a pilot.

Why do most AI projects fail?

According to the MIT report: tools that don't learn from feedback, shadow AI use, budgets aimed at the wrong workflows and no pre-pilot baseline.

Is the 95% in the MIT AI report real?

As MIT defined it, yes — and narrower than the headlines. It counts organisations with no measurable P&L impact within about six months of a pilot, from 52 organisation interviews, a review of more than 300 public initiatives and 153 survey responses. It does not mean 95% of pilots failed: about one in four actual custom-tool pilots reached production under that bar. Treat it as a strong signal from a modest, self-selected sample, not a census.

Does the MIT report say 95% of AI pilots failed?

No. About 60% of organisations evaluated custom tools, about 20% piloted them and about 5% reached production with sustained impact. That is roughly one in four pilots clearing a high bar in about six months — a funnel, not a 95% pilot-failure rate.

Why isn't AI saving my team time?

Time saved on one task is often absorbed by extra work, checking or rework. UC Berkeley Haas research found AI intensified work at one company. Measure a specific task and redirect the hours deliberately.

Stop piloting. Start measuring.

The MIT GenAI Divide is not a verdict on AI. It is a verdict on how most firms deploy it: static tools, shadow usage, sales-and-marketing budgets, no baseline, no learning loop. The 5% that extract value treat AI like a BPO negotiation — specific process, clear metric, vendor accountability — not like a SaaS free trial.

If you already have pilots running, the first question is not "which model?" It is "show me the P&L line you expected to move — and what you measured before you started." If the answer is silence, you are statistically in the 95%. That is fixable.

Start with the free version of that question. Ask Ada which of your tasks would survive MIT's six-month test — free, no signup, and a free report follows before anything paid is discussed. Or run the numbers yourself first in the AI ROI calculator.

Why pilots stall after the demo

The follow-on reporting around stalled pilots keeps naming the same causes, and none of them is mainly the model. Pilots stall when nobody set a measure before the trial, so no one can prove whether it paid. They stall at the last mile, where the tool meets the real workflow, the integrations and the person who has to trust the output. They stall when tools are fragmented across teams, so information never arrives in one place. The practical sequence is the one this page recommends: wire the measure and the checks in before you scale, fix the handoffs the pilot exposed, and only then widen it. A pilot that cannot say which number it moved has not failed — it has not finished.

This article summarizes publicly reported findings from The GenAI Divide: State of AI in Business 2025 (MIT Project NANDA, Aditya Challapally et al., July 2025). It is for general informational purposes only and does not constitute legal, tax, financial, investment, security, or compliance advice. BuildBrain is not a law firm, accounting firm, or registered investment adviser. Statistics and interpretations cited here reflect the source document at the time of writing and may change as follow-on research is published. Readers should verify current information independently and consult qualified professionals regarding obligations specific to their industry, jurisdiction, and circumstances — including applicable federal, state and local requirements. Nothing in this article is an endorsement of any specific AI product, model, or provider.