The forty-billion-dollar silence: why most enterprise AI never ships
The world spent a fortune teaching machines to work, and 95% of it returned nothing. A field report on where enterprise AI goes to die, the vendors faking the cure, and the discipline that ships it.
The world spent a fortune teaching machines to work. Then almost nothing happened. This is a short field report on where enterprise AI actually goes to die, the impostors selling the cure, and the unglamorous discipline that brings it back to life.
The silence
In 2025, enterprises poured tens of billions into generative AI. By the end of the year, 95% of it had returned nothing measurable. Not low returns. Zero (MIT, 2025). The industry even coined a name for the place these projects go: pilot purgatory. Funded, demoed, applauded, and then quietly never heard from again.
The autopsy
The autopsy held a surprise. The model was rarely the cause of death. MIT found the failure was approach and operating discipline: messy data, no owner, no path from a demo to a workflow that actually runs. And buried in the same research, a tell. Teams that bought from specialists reached production about twice as often as teams that built in-house (MIT, 2025). By 2025, even Gartner was telling CIOs the quiet part out loud: buy, don't build.
A demo is the floor. Production is the moat.
The impostors
Which sounds like good news, until you go shopping. Of the thousands of vendors now selling agentic AI, Gartner reckons only around 130 are the real thing (Gartner, 2025). The rest are agent washing: a chatbot in a trenchcoat, an RPA script with a new logo and a confident deck. The question was never whether to bring in an operator. It was which one actually ships.
The commerce twist
This bites hardest for companies selling into commerce. The buyers — brands, retailers, marketplace operators — already run fragmented stacks, and the vendors selling them platforms, OMS, PIM and channel technology inherit that mess in every deal cycle. Ambition is high. Production discipline is rare. The same constraint logic applies whether you are shipping an agent inside a retailer or installing GTM for the software company that sells to them: find the choke, clear it against a number, then the next.
The method
So here is the unglamorous truth we built a company around. You don't fix an organisation by pouring more AI — or more leads — into it. Output is never governed by how busy every part looks. It's governed by one constraint: the single place the work actually chokes. You find it. You clear it with one install, shipped to a number. Then you find the next one. It's the oldest idea in operations, and the AI scramble forgot it entirely. Motion you can measure, not motion that looks busy.
There's a stalled constraint in your commercial system right now. Give us thirty minutes and we'll tell you, honestly, whether it's worth clearing — and whether we're the right people for it.
Sources & further reading
Frequently asked questions
- Why do most enterprise AI pilots never reach production?
- MIT's 2025 research found the failure is rarely the model. It's approach and operating discipline: messy data, no clear owner, and no path from a demo to a workflow that runs. 95% of pilots returned nothing measurable.
- Is it better to build AI in-house or buy from a specialist?
- The same MIT research found teams that bought from specialists reached production about twice as often as teams that built in-house. By 2025, Gartner was advising CIOs to move toward commercial solutions for more predictable value: buy, not build.
- What is agent washing?
- Vendors rebranding chatbots and RPA as 'AI agents' without the capability to ship them into production. Gartner estimates only around 130 of the thousands of agentic-AI vendors are genuine. The question isn't whether to bring in an operator, but which one actually ships.
We’ll name what’s capping revenue.
Thirty minutes. If we shouldn’t be the ones to clear it, we say so.