- 95% of enterprise generative-AI pilots produce no measurable return: that’s an MIT report covered by Fortune.
- The problem is almost never the model. It’s the layer underneath: the company brain, the operating context that gives the work meaning, is missing.
- Agents fail because they don’t know the company: you ask them to reason over decisions, rules, and exceptions that live in heads, not in a readable place.
- The level can be measured: the L0-L4 scale shows why an advanced agent on an L0 company was never going to work.
Ninety-five out of a hundred leave no mark
An MIT report, covered by Fortune, puts a number on a widespread feeling: 95% of enterprise generative-AI projects produce no measurable return[1]. It’s not that the pilots don’t launch or don’t give a nice demo. It’s that, past the demo, almost no one manages to move a number that matters: time, cost, errors, revenue. The project ends, and the company stays where it was.
The easy read is "AI is overhyped". Ours is different: the model, in the vast majority of cases, isn’t the bottleneck. The bottleneck is everything underneath, the layer no one built before turning the pilot on.
95% of pilots don’t fail because the model is weak. They fail because there’s nothing underneath to reason over.
The missing piece is the layer underneath
Today’s models are more than good enough for most of a company’s processes. The missing piece isn’t more power: it’s the operating context, what we call the company brain. The decisions made and why, the rules and the exceptions, the real state of projects, who does what. A model, however capable, can’t guess all of this: either you give it in a readable place, or it starts from zero every time.
This is where almost every project skips a step. It goes straight to the agent, the tool, the integration, before putting anywhere the context the agent is supposed to work with. It’s like hiring a top performer and telling them neither what the company does, nor how, nor why. They do their first day on the job every day, forever.
An agent that doesn’t know the company can’t work
When an agent disappoints, the reflex is to switch models, add prompts, buy another tool. But the flaw is almost always upstream: you’re asking it to reason over knowledge that doesn’t exist anywhere it can read. The decisions are in a chat, the rules in someone’s head, the exceptions in an email from six months ago. The agent doesn’t see them, so it invents them or ignores them. Either way it gets it wrong.
The point isn’t philosophical, it’s operational: an agent is worth as much as the context it can read. On a company where that context lives only in heads and chats, even the most advanced agent produces plausible, useless answers. The brain comes first, then the agent. The reverse order is exactly why 95% leave no mark.
Why that pilot was never going to work
The good news is that this level isn’t an opinion: it can be measured. A company’s brain lives at a precise rung, from "everything in heads" (L0) to "agent-operated" (L4). Looking at the ladder, you see at a glance why an advanced agent, bolted onto an L0 or L1 company, was never going to work: the rung underneath was missing.
- L0In heads and chats
Knowledge lives in people and disconnected chats. Every AI chat starts from zero.
- L1Written but dead
A wiki exists, but it’s stale and nobody updates it. The AI ignores it: a false sense of safety.
- L2Owned and structured
Decisions, state and rules in files you own and govern, updated by habit. People and agents read them.
- L3Self-updating and queryable
It absorbs from email, chat and calls, answers with the source, and captures tacit knowledge through interviews.
- L4Agent-operated
Agents run the work on the brain and keep it fresh. The company operates from it.
Knowing your level is the first concrete step to not repeat the mistake. Take the 2-minute self-check and find out which rung to build before the agent.
The next step is to actually build that context, not buy it: decisions, rules, and state in files you own. We wrote the practical method here, how to build a company brain on files, and it’s the foundation on which agents finally make sense.
AI projects that don’t pay off, in practice
Why do 95% of AI projects fail to pay off?
According to an MIT report covered by Fortune, 95% of enterprise generative-AI pilots produce no measurable return. It’s almost never the model’s fault: what’s missing is the layer underneath, the operating context (decisions, rules, state) that gives the work meaning. Without that context, even a good model stays a toy.
If it’s not the model, what’s missing?
The company brain is missing: the company’s operating context in a place people and agents can read. Decisions made and why, rules and exceptions, the real state of projects. It’s the layer almost everyone skips by going straight to the agent, and it’s why the pilot doesn’t move a number.
Why do AI agents fail inside companies?
Because they don’t know the company. You ask them to reason over knowledge that lives in heads, in chats, and in old emails, not in a place they can read. So the agent invents or ignores the context, and either way it gets it wrong. An agent is worth as much as the context it can read.
How do I know if my company is ready for an agent?
By measuring what level your operating knowledge lives at. The company brain maturity scale goes from L0 (everything in heads) to L4 (agent-operated): an agent makes sense only when the context underneath is yours, structured, and alive. The self-check gives you the level in two minutes.
Sources
- [1]Fortune, “MIT report: 95% of generative AI pilots at companies are failing”, August 18, 2025. fortune.com
This page is written by Raffaele Zarrelli and Simone Bova, founders of Yempik, with editing done with Claude. The company brain and its maturity model are Yempik editorial models. Where we cite a number, you’ll find the source below.
Before you launch another pilot, build the layer underneath.
We start from your process, not the tool. We help you put the operating context on files you own, governed, so the agent has something real to reason over. Fixed price and timeline, the code is yours.