95% of enterprise GenAI pilots hit zero P&L impact. Here's what separates the 5%.

In July 2025, MIT's Project NANDA published findings from 150 executive interviews and an analysis of 300 enterprise AI deployments. The number that circulated widely was this: 95% of GenAI pilots delivered zero measurable P&L impact. Enterprise organisations had spent $30–40 billion on AI, and only 5% of those pilots reached production.

The response from some quarters was instructive. Researchers and consultants argued that P&L was the "wrong metric" for AI. The real value, they said, was in organisational capability, strategic positioning, and workforce readiness. A piece in Digitalisation World framed 2026 as "the year AI impact replaces AI inputs" as a benchmark.

When the argument becomes that ROI is the wrong way to measure AI investment, you have left analysis and entered rationalisation. Which is unfortunate, because the MIT data points at something specific and actionable. Not a reason to abandon AI investment, but a reason to understand why the 95% failed.

The actual variable that separates the 5% from the 95% is task structure.

The task-specificity rule

Across every AI deployment category where independent evidence shows measurable ROI, the same pattern holds: the task has a right answer, the answer is verifiable, and the inputs are structured. Categories where AI consistently fails to deliver measurable value have the opposite: open-ended outputs, hard-to-verify quality, and messy or ambiguous inputs.

Call this the task-specificity rule.

A support bot resolving password resets and order status queries is solving a lookup problem. The answer is either correct or it isn't. Across transactional support tasks, AI handles these at 98.2% accuracy. Compare that to an AI bot managing escalated complaints from enterprise customers approaching churn, where success means de-escalation, relationship repair, and account retention. Accuracy on emotionally complex conversations drops to around 61%. The category label is still "customer support." The underlying task structure is entirely different.

The same rule applies to code generation. Boilerplate, test scaffolding, and CRUD functions are high-structure tasks with verifiable outputs: either the tests pass or they don't. Complex algorithm design, security-critical code review, and architecture decisions are lower-structure tasks where the quality of the output is hard to evaluate automatically. AI code assistants deliver clear productivity gains on the first set of tasks; the gains on the second set are much harder to demonstrate.

What the task-specificity rule gives you, practically, is a pre-investment diagnostic. If you can write an evaluation rubric for the task's output before you start — a clear definition of what "working" looks like, measurable without human judgement — AI is likely worth investing in. If you can't, you are probably heading for the 95%.

Where AI genuinely moves numbers

Three categories have meaningful independent evidence.

Customer support deflection. This is the clearest category, with the strongest evidence. Intercom's Fin AI had resolved 40 million conversations as of December 2025, with a 67% autonomous resolution rate across its install base. B2B SaaS platforms report around 60% ticket deflection from human agents. Salesforce's internal Agentforce deployment handled 3 million support conversations, producing an 8% year-over-year caseload reduction (170,000 fewer cases), with an annualised cost saving Salesforce puts at $100 million. These figures come from the vendors themselves, which matters (more below), but independent benchmarks on AI support ROI are consistent in direction: $3.50 returned per dollar invested in AI customer service, with leading implementations reaching 148–200% ROI within 12 to 18 months.

Code velocity. The most rigorously studied AI application in SaaS. A controlled experiment involving 4,800 developers and GitHub Copilot, published on ResearchGate, found a 55% faster task-completion rate on common coding tasks, 46% of final commit code AI-generated, and an 84% improvement in successful build rates. The gains concentrate in boilerplate generation, test scaffolding, and repetitive CRUD work. Security-critical code and novel algorithms show more modest improvement. But the productivity signal is real and replicated across multiple independent studies. The ROI case for AI code assistance is probably the most solidly established of any AI application in B2B SaaS.

Churn prediction and product analytics. This is earlier-stage, but the signal is there. Work published in 2025 showed gradient boosting models with usage and sentiment signals predicting churn at 92.5% accuracy; the interventions based on those predictions reduced realised churn by around 10% while increasing feature adoption by 15%. Pendo attributes a 25–40% improvement in trial-to-paid conversion to AI-guided activation. These numbers come from smaller samples and need more replication. But they follow from the task-specificity rule: churn risk is a classification problem over structured behavioural signals. It fits.

Where the budget goes instead

Here is the inversion that rarely gets discussed.

A Harlem Capital analysis of enterprise AI spending found that more than half of the budget flowed to sales and marketing AI tools, the category with the lowest measured ROI in their dataset. Back-office automation and developer tooling, which showed the highest ROI, received a much smaller share of spend.

Why? The straightforward answer is visibility. An AI SDR writing personalised cold emails is a thing leadership can demo on a Tuesday afternoon. An internal routing tool that classifies support tickets at 80% accuracy doesn't make it into the board deck. The incentive structure inside most organisations tilts AI spend toward what can be shown, not toward what works.

Marketing content generation is the clearest example. Raw AI-generated B2B content produces around 42% lower engagement than human-written equivalents, according to Semrush's content study. The refined AI-plus-human workflow (AI drafts, humans edit to quality) achieves around 89% of human content performance. But the human editing layer required to get to 89% is almost exactly the editorial effort you'd spend on a human-authored piece. The marginal gain from the AI step is real, but smaller than the "scale to 10× volume" premise requires.

Sales AI tools show a similar pattern. 83% of sales teams using AI tools report revenue growth. So do 66% of teams not using AI. The 17-percentage-point difference is self-reported and comes with substantial selection bias. The specific productivity gains (saved hours per week from meeting transcription, cleaner CRM data from automated entry) are real but don't require the AI-native sales stack that enterprise SaaS vendors are pricing at enterprise levels.

The vendor data problem

The case studies that circulate in AI sales decks have selection bias baked in. A vendor's marquee deployment is the deployment that worked best, implemented by the customer who cared most. That's not dishonest — but it's not a representative sample either.

The MIT 95% failure rate isn't evidence that AI is broken. It's a description of what happens when the median enterprise deployment looks nothing like the best-case reference customer. The median enterprise has messier data, less implementation capacity, weaker change management, and a task portfolio that skews toward the unstructured end.

The practical implication before any AI tool purchase: ask for three things. First, the measurement methodology: how the ROI number was calculated, not just what it is. Second, the comparison baseline: before/after the same team, not a generic industry benchmark. Third, at least one reference from a customer whose implementation underperformed expectations. If the vendor can't provide the third, the headline figure reflects the best case, not the expected case.

The "wrong metric" defence

When the MIT NANDA findings circulated, some of the response from AI-optimistic quarters was worth noting. The argument was that P&L is an "industrial-era metric" that doesn't capture the full value of AI investments. The real value, the argument went, was in positioning, optionality, and building internal capability over time. These things take years to show up in revenue.

This argument has a defensible version. Capability investments do sometimes precede measurable returns, and measuring short-term cost reduction when the actual value is in a changed workflow can give a misleading picture.

But the argument also has an unfalsifiable version, which is the one that tends to circulate when the ROI question is uncomfortable. If no amount of negative P&L evidence counts against an AI investment, because ROI is always the wrong metric in any time frame, you have a framework for avoiding accountability rather than a framework for making good decisions.

The practical test: can you articulate, in a single sentence, how this investment eventually shows up in revenue, retention, or cost reduction, even if the time horizon is 18 to 24 months? If you can't, the investment is closer to positioning spend than operational investment.

A diagnostic you can run in under an hour

The task-specificity framework translates into a practical pre-investment audit. Five steps:

1. Write the success criterion before you start. If you cannot state what "working" looks like in measurable terms before the implementation begins, the task is probably too unstructured for AI to reliably deliver value. This step alone eliminates a significant share of bad AI investments.
2. Identify the comparison baseline. What is the current cost and quality of the process you are improving? AI's value is relative to the status quo. A 60% resolution rate sounds strong until you learn the existing team resolves 72% of the same queries.
3. Audit where the vendor's evidence comes from. Is the headline ROI figure from an independent study or from the vendor's own case study programme? Is the sample size large enough to be informative? Is the implementation context comparable to yours?
4. Pilot on the most structured end of the task first. Start with the highest-structure, most-verifiable use case for any AI product. If the results are poor there, they will be worse on the harder end.
5. Watch your net revenue retention alongside any AI feature launch. This is the tell that gets missed most often. Shipping AI features to improve retention works only if the features address the actual reasons customers leave. If your NRR is declining for reasons AI doesn't address, adding AI features may accelerate rather than slow the decline — customers pay more for a product that still doesn't solve their problem.

The gap that matters

The 95% figure isn't an argument against AI. The 5% is. That 5% is real, measurable, and reproducible. The gap between them comes down to whether the deployment matches what AI is actually good at.

Customer support on transactional queries. Code generation for structured, verifiable tasks. Churn classification over behavioural signals. These are the categories where the evidence is strong enough to act on. They are not the categories that dominate enterprise AI budgets.

The budget allocation inversion (most spend on lowest-ROI categories) will correct, probably slowly. In the meantime, the companies that identify which of their specific problems are structured enough for AI to solve, measure honestly against a real baseline, and resist the pressure to ship visible AI over effective AI, are the ones who will end up in the 5%.