AI knowledge capture agent: forward an email, it feeds the wiki

In a small company, useful knowledge travels by email. The delivery time negotiated with a supplier, the procedure the newest hire must follow, the contact at a customer who really decides: all of it exists, in black and white, in a thread that three people received and nobody can find. The day someone leaves or forgets, the information goes with them. Writing a wiki is the right answer, but it is a task nobody schedules, because it is never urgent. It is patient, methodical work that an AI agent can take on, provided you give it a simple way in: an email address.

The inspiration: forward instead of write

The starting point is a use reported by Hermes Agent users, summarized in a card titled "knowledge capture: from email to the company wiki". The idea fits in one sentence: instead of asking everyone to document what they learn, ask them only to forward the email to the agent, which does the rest. The gesture costs ten seconds, where writing a page costs ten minutes and never gets done.

What we take from it is not the technical feat but the change of effort. The person who knows keeps their role, which is to recognize that an exchange deserves to be remembered. The agent takes the thankless part: reading the history, sorting out the noise, finding the right page, avoiding the duplicate, citing the source. Atako applies the same split, and adds explicit limits on what the agent may do with an email.

What the agent does

The scenario deploys a single agent, the Knowledge keeper. It gets its Atako email address at deployment and gives it in the chat in its first message, with the steps to follow.

When a thread is forwarded to it, it extracts what lasts: decisions (who decided what, and when), figures, contacts with their role and company, procedures and deadlines. It ignores pleasantries and signatures. It then files each fact on the wiki page of the matching section, among those you chose, and creates a new page only if no existing section fits.

Two habits make the difference with an automatic summary. First, the agent updates the page in place: if a supplier's delivery time changes, the line changes, there are not two versions side by side. Second, every fact carries its source, the email subject and date, so anyone can trace an item back to its origin and judge how fresh it is.

Finally, the agent always compares what it learns with what the wiki already says. If the new email contradicts a page, it keeps both versions, flags the contradiction on the page and mentions it to you in the chat. Every Friday, it publishes a "What's new this week" page listing the pages created or changed, the new contacts and the contradictions still open.

What the agent does not do

The limits are part of the design. The agent never replies to the external sender of a thread, never chases anyone and writes to no one outside the company: it is a reader, not a correspondent. It has no access to your mailbox, only to what is forwarded to it. It files only what the thread says: when a fact is vague or incomplete it writes "to be confirmed" instead of guessing, and it invents no figure, name or date.

It does not file credentials or passwords that might sit in an exchange either: it flags them in the chat so you can handle them separately. And it does not decide between two contradictory versions, that is a human decision, which is the principle of a human-in-the-loop design.

How to set it up in Atako

From the page of this use case, deployment takes four steps.

1. Choose the sections. List one to eight sections for your wiki, for example Customers, Suppliers, Procedures. The agent creates one page per section, with a short description of what belongs there.
2. Say who forwards. You can name the people or roles who will send threads to the agent, which helps it understand the context of the emails it receives.
3. Deploy. A single action creates the agent, its email address and the "Knowledge base" project with its how-it-works page. No integration to connect.
4. Forward a first email. The agent gives its address in the chat. Forward it a thread and watch the wiki page update.

If you do not have an account yet, creating one starts your trial and the scenario is configured right after. If your administrator has not allowed you to create agents, the deployment tells you and stops there.

Where to see the results

The result is read in the Atako project "Knowledge base": one wiki page per section, each fact with its source, and the "What's new this week" page on Fridays. On the Atako floor, the agent appears with its current activity, so you can tell whether it is processing a thread without asking. Each forward also gets a confirmation in the chat: what was filed, where, and what contradicts the existing content.

You can talk to the agent in the chat to ask for one more page, a correction or a rewording. Requests are handled in the same conversation, with the same memory of what has already been filed.

Guardrails

The agent has no integration: it touches no CRM, no billing tool and no mailbox. Its only write access is the wiki of the project the scenario created. It contacts nobody outside the company, and the only mail it receives is what someone decides to forward to it.

Human validation sits where it matters. You choose what comes in, by forwarding a thread or not. You settle the contradictions the agent flags. And you read the pages like any company document: an item marked "to be confirmed" waits for a person to confirm it.

What to expect

Expect a wiki that becomes useful by accumulation, not from the first email. In the first weeks the pages are short and the sections evolve: that is normal, and the agent adapts when you ask it to split or merge a page. The value shows when someone looks for an item and finds it in ten seconds, with its date and origin, instead of digging through a mailbox. For a team of a few people, it is the difference between knowledge that lives in everyone's head and knowledge that stays in the company when people move on.