An AI tool knows your company when it can answer using your definitions, your rules, and the institutional knowledge that makes your business specific. Most tools can't. They guess — confidently — from whatever they can infer from a prompt and a few rows of data. A guess that sounds right is still a guess.
The model was never your bottleneck
Everyone is shopping for a better model. The companies whose AI actually works this year won't be the ones with the best model. They'll be the ones whose company is written down somewhere the model can read.
The most cited evidence comes from MIT's Project NANDA, whose 2025 report The GenAI Divide: State of AI in Business 2025 reviewed roughly 300 public deployments alongside dozens of executive interviews and surveys. Its headline finding: about 95% of enterprise GenAI pilots delivered no measurable impact on the bottom line. It's worth being precise about the cause, because it's easy to misread — the report blames a "learning gap," not weak models. Most systems, it found, never retain feedback, adapt to context, or improve on the specifics of the business they're dropped into.
Gartner sees the same wall from a different angle: it predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate controls. Different study, same story — the projects don't die in the lab; they die on contact with the real organization.
That should change where you spend your attention. If the constraint were raw capability, a bigger model would fix it. The constraint is that the model has no access to how your business actually works.
See it happen: one term, three answers
This isn't abstract. Take a single question every revenue team asks — how many active customers do we have? — and put it to three tools already wired into your stack. Nobody ever wrote the definition down, so each one infers its own and reports it with total confidence:
- CRM sales copilot → 4,120. Inferred "active" as any account with a user login in the last 90 days. Counts trials and churned-but-still-poking-around accounts.
- Billing assistant → 3,540. Inferred "active" as any account with a paid invoice this quarter. Misses paid-annually customers who invoiced last quarter.
- Support bot → 2,980. Inferred "active" as any account with a recent or open ticket. Silent customers — often the healthiest — simply don't appear.
- Governed definition → 3,610. A paid-plan account with at least one active seat in the trailing 30 days — the number every tool now returns, because they all read the same entry. (Owner: VP Finance. Last confirmed: 2026-05-12. Visible to: revenue + CS tools.)
Three tools, three numbers, all plausible, none agreed on. Nobody chose this — each tool guessed, and they guessed differently. So nobody trusts any single output, the dashboard contradicts the board deck, and the AI project quietly stalls. The governed entry doesn't make the tools smarter; it makes them consistent and correct, which is the part that lets you ship.
Your company isn't written down anywhere AI can read it
Here's the uncomfortable part. Your company mostly lives in people's heads and in tools that don't agree with each other. It was never recorded in a form a machine can use. That single gap shows up in four ways — you'll recognize all of them.
1. Every tool invents its own definition. Exactly what you just saw with "active customer." Ask three tools the same question, get three answers. Nobody decided it; each one guessed. So nobody trusts the outputs, and nothing ships.
2. The rules that run the business aren't written down. Why that account gets the discount. How a renewal actually gets approved. Which customers are politically sensitive, and why. This is the knowledge that makes the company run — and it lives in DMs, in ex-employees' heads, and in Slack threads from two years ago. No model can infer it, and no model can reach it.
3. You teach every new tool from scratch. Six weeks getting one assistant to understand the business. Then the next tool needs the same. Then the one after that. Every tool starts at zero, and the explanations drift apart until no two agree. Your data team becomes the integration layer, by hand.
4. You can't audit what the AI knew. The answer sounds plausible. But you can't tell which definition it used, which data it pulled, or which policy it applied. You can't put AI in front of a customer or in a board deck when no one can trace the reasoning back to a source.
None of these are model problems. They're the same problem wearing four different coats: the company was never written down for the machine.
"Isn't this just a wiki?" — and three other fair questions
If you've been at this long enough, an objection is already forming. Good. Here are the four that come up, answered straight. Each of these tools is real and does a real job — none of them does this job: giving every AI tool the same governed understanding of your business at the moment it answers.
What a context layer actually is
The architecture name for the fix is a context layer: the governed layer between your data stack and your AI tools. It holds what your data means, how your company works, and the rules for who can see what — and serves it to any tool at the moment it answers. A semantic layer is one slice of it. A wiki is a rough draft of it. A context layer is the version a machine can actually use.
What it actually takes
AI that knows your company needs three things, and none of them is exotic.
- What your data means. Your definitions, metrics, and logic. Not "revenue" in the abstract, but how you define revenue — with an owner and a date it was last confirmed.
- How your company actually works. The renewal flow, the discount policy, the account sensitivities. The unwritten rules — written down once.
- Served with authority and audit. Every tool reads the same context at answer time, filtered by who's asking, and every answer traces back to its source.
That last part is the difference between a nice idea and something you can ship. AI that "knows" your company has to know it correctly. The promise is empty if it isn't trustworthy — which is why the context has to be governed and traceable, not confident guessing dressed up as knowledge.
That last part is the difference between a nice idea and something you can ship. AI that "knows" your company has to know it correctly. The promise is empty if it isn't trustworthy — which is why the context has to be governed and traceable, not confident guessing dressed up as knowledge.
What changes when you fix it
The picture shifts the moment your company is readable. Ask three tools what "active customer" means and you get one answer — the one your company actually uses. A new AI tool is useful on day one, because it inherits everything you already wrote down. Every output is traceable, which is what makes it safe to put in front of a customer. And a fifty-person company starts running its AI with the discipline of a five-hundred-person one, without hiring a team to babysit prompts.
The winners won't be the ones who found a better model. They'll be the ones who did the unglamorous work first.
How to start on Monday
You don't need a quarter-long rollout to begin. The work compounds, so start where the disagreement is loudest.
- Find the three terms your tools fight over. Active customer, churn, qualified lead, ARR — whichever ones produce different numbers in different places. Disagreement is a map of where the value is.
- Write each as a governed entry. One definition, an owner who's accountable for it, and a last-confirmed date. Capture the one or two unwritten rules behind it while the owner is in the room.
- Wire one tool to read from it. Point a single assistant at the entry and watch the answer become consistent. Then add the next tool. Each one you connect inherits everything already written down.
The work of this decade
Every company is wiring AI into how it operates. The ones that win won't be the ones that found a better model. They'll be the ones that did the unglamorous work first: writing down what their company means, so the machine can finally read it.
Your AI doesn't know your company. That's what we're building Sento to fix.
Frequently asked questions
Why do most enterprise AI projects fail in production? Usually not because the models are weak. MIT's Project NANDA found that 95% of enterprise GenAI pilots delivered no measurable bottom-line impact, and pinned the cause on a "learning gap" — the tools never adapt to or retain how the specific business works. Gartner, separately, expects more than 40% of agentic AI projects to be canceled by the end of 2027 for similar reasons. The gap is context, not the model.
What does it mean that "AI doesn't know your company"? The tool can't answer using your specific definitions, rules, and institutional knowledge. It infers from a prompt and a few rows of data — so the same question gets different answers in different tools, as the "active customer" example shows.
What is a context layer for AI? The governed layer between your data stack and your AI tools. It holds what your data means, how your company works, and who can see what, and serves it to any tool at the moment it answers — with every answer traceable. A semantic layer is one slice of it; a wiki is a rough draft of it.
How is a context layer different from a semantic layer or a wiki? A semantic layer defines metrics for BI tools — one slice. A wiki is written for humans who already have context, and goes stale. A context layer carries definitions, unwritten rules, and access policy together, and serves them to every AI tool at answer time with authority and audit.
How is a context layer different from RAG? RAG retrieves relevant documents so a model has something to read. A context layer governs meaning: which definition is authoritative, who's allowed to see it, when it was last confirmed — served consistently across tools. RAG can sit underneath; the context layer is what makes the retrieved answer correct and traceable.
Why not just put the definitions in the system prompt? A glossary in the system prompt doesn't travel across tools, doesn't update when a definition changes, has no authority behind it, and breaks the moment the model composes queries against live data instead of reciting a definition.
How do I start building one? Start small. Pick the three terms your tools disagree on most, write each as a governed entry with an owner and a last-confirmed date, capture the one or two unwritten rules behind each, and wire a single tool to read from it. Expand from there.
