Why Your AI Agent Keeps Getting Your Business Wrong — and How Open Knowledge Format (OKF) Fixes It

You've connected an AI agent to your business. It writes fluent answers, drafts reports, and queries data in seconds. And then it confidently tells a customer the wrong refund policy, or calculates "active customers" in a way nobody on your team would recognise.

The model isn't the problem. The agent simply doesn't know how your business works — and nobody ever wrote it down in a form it can reliably read.

That gap is exactly what Google Cloud's Open Knowledge Format (OKF), released in June 2026, sets out to close. Here's what it is, why it matters for any business building with AI, and how to start without a big project.

The real reason AI agents get your business wrong

Large language models know a lot about the world and almost nothing about your company. To answer correctly, an agent needs things like:

• What your terms actually mean — "active customer", "qualified lead", "gross margin"
• Which database table or report is the trusted source for each number
• Your internal rules — pricing exceptions, approval limits, refund windows
• Which documents are current and which are outdated

In most companies, that knowledge is scattered across wikis, Google Docs, README files, Slack threads, and people's heads. Every time a team builds a new agent or chatbot, someone writes custom glue to pull from those sources. It's fragile, it's duplicated, and when two sources disagree, the agent has no way to know which one to trust — so it picks one, confidently.

Better prompts don't fix this. Better-organised knowledge does.

What OKF actually is

OKF is not a product or a platform. It's a format — a simple, open agreement on how to write knowledge down so that both humans and AI agents can use it.

An OKF knowledge bundle is just:

• A folder of Markdown files, where each file explains one concept — a metric, a product, a process, a policy, a dataset
• A short YAML header at the top of each file. The only required field is the page's type; fields like title, description, tags, and the resource it describes are recommended
• Normal Markdown links between related pages, so an agent can follow connections the way you'd click through a wiki
• Index pages that act as a table of contents for each folder

Here's what a single page might look like:

---
type: metric
title: Active Customer
description: A customer with at least one paid order in the last 90 days.
tags: [sales, reporting]
---

# Active Customer

A customer counts as **active** if they placed at least one paid order in the
last 90 days. Refunded orders do not count.

- Source of truth: the `orders` table in the sales warehouse
- Used in: [Monthly Revenue Report](/reports/monthly-revenue.md)
- Not the same as: [Registered User](/metrics/registered-user.md)

That's it. No database, no special editor, no vendor account. Because it's plain text, it lives happily in Git alongside your code.

OKF vs RAG vs MCP: who does what?

These three terms often get mixed up. They solve different problems and work best together:

A typical setup: your agent uses MCP to pull live data from your systems, uses RAG to retrieve the relevant OKF pages that explain what that data means, and combines both to answer correctly. Clean, labelled, one-concept-per-page knowledge also tends to make retrieval more accurate — though, as with anything in AI, it's worth testing on your own questions.

Why this matters for your business

No lock-in

Your knowledge stays as plain files in your own repository. Switch AI models, vendors, or tools, and the knowledge comes with you.

One source for people and agents

The same pages that onboard a new employee can brief an AI agent. One place to update means far less "the docs say one thing, the bot says another."

Governance with tools you already have

Because it's just files, you can review changes through pull requests, track history in Git, and run automated checks for broken links or missing fields — the same way your team already manages code.

Trust and freshness

Once agents start helping write and update knowledge, you need to know where each page came from, whether a human has checked it, and whether it's still current. Newer versions of the spec are moving in exactly this direction, with metadata for sources, verification, and when a page should be considered stale.

A practical way to start (in an afternoon)

You don't need to document the whole company. Start small and prove the value:

1. Pick one painful area. Somewhere agents — or new hires — keep getting things wrong: sales metrics, product catalogue rules, support policies, or developer onboarding.
2. Write 10–20 short pages. One concept per page, a type in the header, a clear definition, and links to related pages.
3. Add an index page so people and agents can see what's in the folder at a glance.
4. Test real questions. Ask your agent the questions it used to get wrong — before and after — and compare the answers.
5. Put it under review. Store it in Git and require a quick review for changes, so the knowledge stays accurate.

If the answers improve, expand domain by domain.

Honest limits

• It's young. OKF launched in 2026 and is still evolving, so check details against the current specification before building tooling around it.
• It won't fix wrong information. If your definitions are inaccurate or contradictory, writing them in OKF just makes the mistakes easier to read. Someone still has to own the content.
• It isn't a database, search engine, or security layer. Sensitive knowledge still needs proper access controls.
• It may not replace what you have. If you already run a well-governed data catalog, OKF is more likely to be an export format alongside it than a replacement.

The takeaway

The next big improvement in your AI agents may not come from a bigger model. It may come from finally writing down how your business works — clearly, one concept at a time, in a format both your team and your agents can read.

How Brahmweb can help

We build AI agents and chatbots that work with your real business knowledge — not just a demo. We can help you identify the knowledge your agents need, structure it in an open, agent-readable format, connect it to your systems through RAG and MCP, and test that the answers are actually right.

Explore our AI agent development services, or talk to us about making your business knowledge AI-ready →