Our AI already knows all this. Why write it down?

It knows it for the length of a session, for one person, in one tool. It cannot be checked, versioned or corrected, and the next agent or the next hire cannot read it. The fact layer is that same knowledge, verified once, then readable by every tool you use, including the next one you adopt.

Is this another system of record to maintain?

It replaces maintenance rather than adding it. Today the same claim about your company lives in a deck, a website, a proposal template and three people's heads. The test is whether the number of places to update goes down by month six.

Will my team keep it up to date?

They will not maintain a wiki nobody reads. They will maintain the thing their own assistant answers from, because a wrong fact produces a wrong answer in front of them the same day.

Our information is sensitive.

Keys are narrow and scoped: a key can be limited to named stacks, and read only. Nothing needs a key wider than its job. Start with what is already shared internally.

Won't the AI make things up?

That is the problem the fact layer exists for. Answers are grounded in facts a person verified, and an unanswered question comes back as a gap rather than a guess. It is also why the second stage ends in verification, not drafting.

We tried a wiki and it died.

A wiki is written for people who might read it later. A fact layer is read every day by assistants and agents, so being wrong has an immediate, visible cost.

We are a small startup. Is this for us?

Yes. A twenty-person company and a two-thousand-person company do the same five stages; what changes is the content. Start on the Free plan, and if you are an agentic B2B startup, apply for twelve months of the Team plan through NOAN for Startups.

Rolling out NOAN

Start with one team

NOAN is headless, so the first day of a rollout is a connection and a draft, not a workshop or a migration. Connect the AI your team already uses, let it draft your first facts, and have a person verify them. A twenty-person startup and a two-thousand-person enterprise do the same five stages.

Start free See the five stages

Watch a rollout

From the first team to every team

How to think about it

Connect first, write second

Your team already uses AI, and that AI holds a working picture of your business that it loses every time a session ends. A rollout gives it somewhere to write that picture down, checked by a person, and read again tomorrow by every other tool.

Your AI writes the first draft

The assistant you already use has your documents, your repositories and your threads. Asked properly, it drafts the first version of your facts in an afternoon. Correcting a draft is work people will actually do. Writing from a blank page is not.

A person verifies it

A draft is not a fact until a person has read it. An AI is confidently out of date about pricing, headcount and anything decided last quarter. Mark what is unknown rather than filling it: a gap is the next task.

One team, one win, four weeks

Adoption is won in the first four weeks, in one team, on one visible outcome. Not a launch announcement, and not a training program.

Never more stacks than owners

Every stack has one steward, never a committee. Add an area only when someone owns it. An unowned stack decays into the wiki everybody stopped trusting.

The method

Five stages, in order

Each stage opens when the one before it is done. Size changes what goes into each stage, never the order.

01
Connect the AI you already use
Point Claude, ChatGPT, Cursor or Claude Code at NOAN's MCP server on day one, before anything is written. Start with one narrow key for one person, then widen. It takes a morning, not a project.
In Claude Code:
claude mcp add --transport http noan https://mcp.getnoan.com/mcp
Other clients are on the MCP page.
Done when
Your own assistant answers from your workspace, and one fact exists that it wrote.
02
Let it draft the spine, then verify
Who you are, what you sell, to whom, how you win, how you price, how you operate. Drafted from the deck, the site and the repository your AI already has, then corrected and verified by the people who know.
Done when
Six to ten verified blocks a new joiner could read in twenty minutes, and one real decision this month that cited them.
03
Put a name against every stack
Each stack gets one steward who keeps it true: twenty minutes a week to read it, fix what changed and verify what they checked. Name the stewards in a fact, so ownership outlives the project plan.
Done when
Every stack has one owner, and each has run the review at least twice.
04
Reach everyone without an AI open
People who will never install an MCP client get the same answers from the assistant in the NOAN app and from Verity in Slack, in the channel where the work happens. Pick the team that already feels the pain, give them one faster workflow, and let them tell the story.
Done when
A team outside the sponsor's own function has used NOAN weekly for a month without being asked.
05
Ground your systems, then build
Point one real system at the facts first, ideally where stale copy already causes work: a pricing page, an onboarding email, a proposal template. Mint narrow keys, one per use. Then build agents, apps and sites on the same facts, starting from agents NOAN already runs rather than a blank repository.
Done when
A production system reads NOAN, keys are governed rather than shared, and you ship something you designed on top of it.

Same ladder, any size

A startup or an enterprise

What changes with size is who owns what, and how many teams you bring in. The stages stay the same.

A startup

One founder, one afternoon, one spine

1The founder is often the sponsor and the connector. Connecting usually takes the first hour and does not need engineering.
2Your AI drafts the spine from the deck, the site and the repository it already works in. You verify it.
3Keep the spine small and sharp: one stack per area someone owns, and a new stack each time someone takes on an area.
4A good first win: a coding agent that stops being told the same context every morning, or a pricing claim that is finally identical everywhere.
5When you build, start from agents NOAN already runs, including the open-source agent pack, rather than a blank repository.

Start on the Free plan. Agentic B2B startups can apply for twelve months of Team through NOAN for Startups.

An enterprise

One team at a time, on the same facts

1Run the same five stages per business unit, with a sponsor each, and the spine at group level.
2Start with the team that owns its facts, often marketing. Their AI drafts from the deck and the site, marketing verifies, and the first win is theirs.
3Then sales, starting from the positioning and proof marketing already verified. Anyone can post a fact; only the stack's steward verifies it.
4Bring each team in at a moment it already has: a launch, an ops pack, results day. Each Slack channel sees only the stacks you allow, and keys can be limited to named stacks, read only.
5Then build: your site and product read the same facts over the API, so a price finance verifies once shows everywhere on the next read, with no deploy.

Owners first, then breadth: never more stacks than people who own them. NOAN for Enterprise

What to expect

Month one to month nine

Month 1

Connected, written, one win

Your AI writes to NOAN in week one, the spine is verified by week three, and one visible win has landed.

Measure
Verified facts on the spine, and one decision that cited them.

Month 3

Habit

Stewards review weekly, one or two teams use the app assistant and Slack, and the first agent is adopted.

Measure
People outside the sponsor's function using it weekly, and facts edited in the last thirty days.

Month 6

Grounded

One or two production systems read the API, key scope is governed, and two or three agents do real work.

Measure
Systems reading the API, and stale facts found by review rather than by a customer.

Month 9

Yours

NOAN is the system of record. Your own agents, apps or sites run on it, and decisions that cost money cite it.

Measure
The number of places a claim about your company has to be maintained. It should be one.

Who does what

Four roles, one name each

Sponsor

The senior leader accountable for the change. Picks the first win, names the stewards, clears blockers. One per company, or one per business unit.

Steward

Owns one stack and keeps it true, twenty minutes a week. The person to ask when something in that area is disputed. One name, never a committee.

Connector

Gets the MCP server and keys working, usually in the first hour. Often the sponsor; it does not need engineering.

Verity

NOAN's agent. Drafts, surveys and reminds. She proposes changes to facts and a person applies them, so your record never changes without someone choosing it.

The rhythm that keeps it alive: each steward spends twenty minutes a week on their stack; the sponsor sends the exec team one short note on what moved, what is at risk and what is needed; and once a month, if the horizon has not landed, shrink the next month rather than repeating it.

Questions

What people worry about

Avoid

What stalls a rollout

Mapping the whole company before anything is used.
Waiting for a formal integration project before anyone connects their own AI.
Publishing AI drafts nobody read.
One person as the single writer.
Outsourcing the first facts, so your record arrives in someone else's voice.
Agents and production integrations before ownership.
A launch announcement instead of a first win.

Get your rollout plan

Subscribers get the implementation workstream, included. Verity builds a 1, 3, 6 and 9 month roadmap from your own workspace, writes it there as a fact you own, puts month one on your board, and checks in each weekday. Start it on the page, or by emailing verity@getnoan.com.

Start your plan Start free