MIT's Project NANDA spent 2025 studying how businesses were actually faring with their AI initiatives, not the demos, the real deployments. The headline finding, published in their report The GenAI Divide: State of AI in Business 2025, was stark: 95% of organisations saw no measurable return.

  • 67% Success rate for initiatives built through an external partnership
  • 33% Success rate for initiatives built entirely in-house
  • 2x Partnering roughly doubles the odds the thing actually works

Four sources, four methods, one conclusion

That finding doesn't sit alone. Three other independent bodies of research reach the same conclusion from different angles.

RAND Corporation's analysis of enterprise AI initiatives found that more than 80% fail to deliver their intended value, twice the failure rate of ordinary IT projects. The more useful number sits underneath that one: 84% of those failures trace back to leadership and organisational decisions, not technical limitations. The technology, in other words, usually isn't the problem.

Gartner's research points at a related cause. Their forecast: 60% of AI projects lacking properly prepared data will be abandoned before 2026 is out. Not because the model was wrong. Because nobody had mapped what the business actually had to work with before building on top of it.

And Databricks' 2026 State of AI Agents report, drawn from data across more than 20,000 organisations, quantifies what happens once governance and evaluation are actually in place. Companies that evaluate their systems properly get nearly six times more projects into production. Companies with real governance get twelve times more.

AI failure is almost never a technology problem. It's a diagnosis problem: skipping the work of understanding the business, the data, and the accountability before building anything.

Where the returns actually showed up

There's one more detail from the MIT NANDA research worth sitting with. More than half of AI budgets in 2025 went toward sales and marketing pilots, the visible, exciting use cases. The strongest actual returns showed up somewhere quieter: back-office automation, process efficiency, the unglamorous operational work. Businesses chasing the flashy win, without the operational foundation underneath it, are disproportionately represented in that 95% failure column.

This is, in a fairly direct way, the case for building growth capability and operational capability together rather than one without the other, and for treating diagnosis as the first step, not a formality on the way to a build.

Before you hire for it

This is playing out in how businesses are actually approaching the problem right now, not just in the research. A growing number of businesses looking to bring AI capability in-house are engaging outside expertise first, to properly map what a permanent role would need to own, before committing to a lengthy, expensive search to fill it. Know what you're actually building, and what it requires, before you commit resources to it. That logic applies as much to a headcount decision as it does to a technology one.

It's also, frankly, exactly why we work the way we do. Every engagement starts with a proper assessment before a single system gets built, not as a sales step, but because the data above is unambiguous about what happens when that step gets skipped. We'd rather be slower out of the gate and actually get you into the successful minority than move fast and join the 95%.

Where this fits

If you're currently scoping a role like an AI Strategy Lead, an AI Enablement Manager, or a Head of AI, we can help map out exactly what the role needs to own, what capability already exists in your business, and what the first 90 days should actually look like, before you commit to the hire. That's often a faster, lower-risk way to get moving than a six-to-twelve-month search for someone to write that scope themselves once they start.