MIT's Project NANDA spent a year interviewing 150 executives, surveying 350 employees, and analysing 300 public AI deployments. Their conclusion was stark: 95% of enterprise generative AI pilots are failing to deliver meaningful results. Only 5% achieve what the researchers call "rapid revenue acceleration." I've been sitting with that number for weeks. Not because it's surprising — but because of what it tells us about how organisations are approaching AI adoption.
The failure isn't in the technology. The technology works. The failure is in where companies are deploying it, how they're building it, and what they're measuring. And those are all strategy and change management problems, not engineering ones.
Companies are investing in the wrong places
Here's the finding that stopped me: over 50% of enterprise AI budgets are being allocated to sales and marketing tools. Customer-facing, revenue-generating, front-of-house applications. Which sounds logical — until you look at where AI adoption actually delivers the strongest return on investment.
Back-office automation consistently outperforms customer-facing deployments. Operational efficiency, process streamlining, internal workflow redesign — these are the areas where AI compounds fastest and fails least often. But they're unglamorous. They don't make for a compelling demo at a board meeting. So companies underfund them and overfund the flashy stuff, and then wonder why the pilot didn't move the needle.
Companies are investing heavily where AI adoption returns the least, and lightly where it compounds fastest.
Build vs buy is not a neutral decision
The MIT research also found a significant split in adoption outcomes based on how organisations were deploying AI. Specialised vendor solutions succeeded 67% of the time. Internal builds succeeded 33% of the time — exactly half the rate.
This matters because most large organisations have a strong bias toward building internally. It feels more controllable, more customisable, more "ours." But the data doesn't support that instinct, at least not at this stage of AI maturity. Vendors who specialise in a specific problem domain have usually already iterated through the failure modes that internal teams are about to discover for the first time.
That doesn't mean never build. It means the build decision needs to be justified on its merits — not just on a preference for internal ownership or a distrust of vendors.
The 5% who are getting it right
The MIT report identifies what they call the "GenAI Divide" — a widening gap between a small group of organisations compounding AI advantage and a large majority running stalled pilots. The 5% who are succeeding share a few things in common.
They're deploying AI where operational complexity is high and the cost of human error is also high — areas where AI doesn't just save time but actively reduces risk. They're measuring adoption outcomes, not just deployment counts. And crucially, they've treated AI adoption as a capability transformation, not a technology rollout. The people side of the change — training, communication, role redesign — is treated as primary, not as an afterthought.
The 95% failure rate for enterprise AI solutions represents the clearest manifestation of the GenAI Divide. — Aditya Challapally, MIT NANDA Project
The RAND Corporation, which has analysed over 80% failure rates across AI projects more broadly, reaches similar conclusions: the technical capability exists. The organisational capability to absorb and sustain it often doesn't.
What this means for your adoption strategy
If your organisation is currently running AI pilots that haven't yet delivered results, it's worth asking a few honest questions. Where in the business did you start? Was it where the opportunity for AI adoption was greatest, or where it was most visible? Did you build when you could have bought — or buy something generic when a specialist would have served you better?
And most importantly: did you treat this as a technology deployment or a capability change? Because the evidence is clear that the organisations landing in the 5% are the ones who did the harder work — the people work — before the tools even went live.
AI adoption doesn't fail at the algorithm. It fails at the implementation strategy, the communication plan, and the change readiness of the teams expected to use it.
