Introducing the Product Brain: product knowledge for AI agents

Pablo Omenaca

Co-Founder at Karumi

Ask an AI support agent built on a general model how to import contacts into your product, and you get an answer that sounds right: look for an Import button, or check the help center. It is the answer for every product, which makes it the answer for none.

Ask an agent that has seen your product, and it says: Contacts › Import, top right. Upload a file, paste a list or pull from another tool, and the columns map to contact attributes for you. Here is the screen, and here is a 16-second clip of the flow.

The difference is not the model. It is what the model knows about your product. Today we are opening the Product Brain, the knowledge layer behind every Karumi agent, to the agents you build.

In this article we explain what the Product Brain is, why AI agents need it, how it works and how to connect your own agent.

A cloud of product screens and dialogs wired to a central Product Brain, with the path from the question "How do I import my contacts?" to the Contacts and Import contacts screens highlighted
A question lights the path through the Product Brain to the screen that answers it.

What is the Product Brain?

The Product Brain is a live map of a software product: its screens, flows, features, settings and roles, plus the business context around them. Karumi builds it by sending its own agents into the product to navigate it the way a person does, keeps it current by never stopping, and serves it to AI agents as structured answers with screenshots and video.

It is not a search index over your help center. A help center describes the product as it was when someone last wrote an article. The Product Brain is built from the product itself.

Why product agents give wrong answers

Most teams building agents hit the same three walls:

  • Stale. Docs and recorded demos drift from the interface within a few releases, so the agent describes buttons that have moved.
  • Shallow. A paragraph of text does not answer "how do I". People need the path, the steps, the permission they need and something to look at.
  • Unresolved. "Change my billing email" can point to three different screens. Without a map of the product, the agent picks one and is confidently wrong.

Better prompts and bigger models fix none of these. They are knowledge problems.

The question "How do I import my contacts?" answered twice: a generic answer without Karumi, and an answer with Karumi that names Contacts › Import and shows the screen with the Import contacts button highlighted
The same question, answered from general knowledge and from the Product Brain.

How the Product Brain works

Your live product, every release, the help center and demo recordings flow into the Karumi Product Brain, which agents use to search, explain a feature or watch for changes
Four sources feed one Product Brain. Agents search it, ask it to explain a feature, or watch it for changes.

1. Karumi's agents navigate your live product, continuously

You give Karumi a login to a demo environment, and Karumi sends agents into your product. They sign in, open every page, click through each flow and record what they see, the way a new hire explores a product in their first week, except they never stop. They keep scanning around the clock, so the map follows the product as it changes, and a new product maps itself in about a day.

A browser page with a coral scan line sweeping down it, the blocks it has crossed lit up, above the chips "Oct 14 release" and "Re-scanned 9:15"
Karumi's agents keep scanning the product, so every release is picked up the same day.

2. It builds a graph, not a pile of documents

Every screen becomes a node, linked to the flows that pass through it, the features on it, the roles that can see it and the plan it belongs to. Product knowledge sits next to business knowledge, such as plans, pricing and the questions customers actually ask, so an answer can account for both.

3. It stays current when you ship

Because the agents keep navigating the product, the next pass after a release notices what changed. It retakes the screenshots, re-records the clips and updates every feature that screen touches. You ship on Tuesday; your agents know on Tuesday, and nobody rewrites a guide.

4. Your agents query it

Agents ask in plain language, the way a user would. The Product Brain returns the feature that answers the question, with what an agent needs to act on it:

  • the path to it, such as Contacts › Import
  • the steps, in order
  • who can use it, by role and plan
  • current screenshots and a short clip of the flow
  • when Karumi last saw it in the product

Retrieval is indexed for speed, which is why Karumi's own agents answer in about two seconds.

A plain-language search for adding contacts to an existing list returns three features, Import contacts, Sign-up forms and Lists, with the first one's API response open
A plain-language query returns the features that answer it, each with its path, steps and media.

What you can build on it

AgentA question it getsWhat the Product Brain hands back
Support agent"How do I add contacts to a list I already imported?"Import contacts and sign-up forms, with steps and clips
Onboarding agent"What should I set up in my first week?"Sender domain, first list and first campaign, in order
Sales agent"Can I send SMS and WhatsApp, not just email?"Where each channel lives, with a clip of the switch
Documentation agent"Which help articles are out of date?"The articles whose screens changed in the last release, with fresh screenshots

These are the jobs Karumi's own agents already do, and the Product Brain is what they run on. StackAI's onboarding agent, trained on its product, docs and setup flows, onboarded more than 40 enterprise teams in a quarter and cut manual onboarding time by 85% (read the story). Re-Leased's demo agent answers 83% of prospect questions live (read the story).

How to connect your agent

There are two ways in:

  1. MCP. Add Karumi to Claude, Cursor or any other MCP client with one block of config: {"mcpServers": {"karumi": {"url": "https://api.karumi.ai/mcp"}}}. The MCP documentation covers the setup.
  2. API. For agents that run in your own stack, request API access.
An mcp.json config block pointing an MCP client at https://api.karumi.ai/mcp
One block of config connects any MCP client to the Product Brain.

Either way, the first step is the same: point Karumi at your product, and it maps itself in a day.

Frequently asked questions

What is a product brain?

A product brain is a continuously updated map of a software product, its screens, flows, features and roles, that AI agents can query for grounded answers. Karumi's Product Brain adds current screenshots and short video clips to each answer.

How is it different from RAG over our docs?

Retrieval-augmented generation over docs returns text that someone wrote about the product, as current as its last edit. The Product Brain is built by agents that navigate the product itself, so it has screens, steps, roles and video, and it updates when the interface changes.

Does it stay accurate after a release?

Yes. Karumi's agents keep navigating your product around the clock. When a release changes a screen, its screenshots, clips and every feature that uses it update the same day.

Which agents can use it?

Any agent that can call an MCP server or an HTTP API: support bots, onboarding guides, sales assistants, documentation writers and internal copilots.

Is our data used to train AI models?

No. Customer data stays inside your boundary and is not used for training.

How long does setup take?

You provide a login to a demo environment, and Karumi maps the product in about a day.

See it on your own product: explore the Product Brain, or talk to our team.