An AI knowledge base for your customer work

An AI knowledge base gives people and agents a maintained source for company facts and procedures. In Customermates, those pages live in the customer workspace, so a person, Mate or an external MCP client can consult the guidance alongside the customer record before preparing work.

This fits a small team that handles customer conversations and next steps across several tools. It is not a public help centre, a learning platform or a replacement for an enterprise search system.

What belongs in an AI knowledge base for customer work?

Start with facts that change what your team does: what you offer, who can approve a commitment, how to hand over a customer, and where to check an exception. Give each page an owner, a source and a review date. Keep a customer's current stage, owner and agreed next step in the customer record rather than copying them into a permanent policy page.

The Knowledge Base documentation describes the page types and editing. The broader agentic CRM guide explains how knowledge, customer context and actions fit together.

How do people, Mate and MCP clients use the same source?

A person opens the workspace's Knowledge Base. Mate can consult company knowledge as part of supported work in Customermates Cloud. An external client connects through MCP and the API, reads the workspace context, searches the Knowledge Base and fetches the relevant page or section.

The source should help finish a concrete task. For a handover, read the customer record and the handover procedure, prepare one internal task, assign its owner, link it to the customer and read it back. Creating a task does not send a customer message or establish a new commercial commitment.

1. Read the source

   Start with the operating guide and the procedure that matches the situation. Follow relevant reference links.

2. Check the current customer

   Read the customer's actual record and agreed scope. A generic procedure does not supply missing customer facts.

3. Prepare the bounded next step

   Create an internal checklist or draft supported by those sources. Show where each factual statement came from.

4. Read the saved result

   Check the task, assigned owner and customer relationship. A successful request alone does not prove the intended state was saved.

The integration overview separates external clients, REST, webhooks and other connection options. No local Markdown synchronisation or shared-drive connector is implied by storing company knowledge here.

What did our four-page example actually demonstrate?

Checked on 29 October 2026 in an isolated local workspace with fictional data: we created four bilingual pages, fetched their contents through the actual MCP endpoint, searched in English and German, and prepared a manually written answer with source references. We then created an internal task, replayed the same idempotent request and read back the task's owner and customer link.

A separate reader with API access but no Knowledge Base read permission could not fetch the guide. This checks the workspace permission boundary; it does not establish per-page permissions or a retrieval-accuracy benchmark. No external message, paid AI call or third-party account connection was part of the example.

You can download the four-page Knowledge Base starter pack. It contains original Markdown examples and instructions for reviewing and copying them into pages. It does not add a dedicated Markdown import/export feature. Replace the fictional content and confirm your own rules before use.

What happens when the source does not answer the question?

Do not turn a search result into a promise. In our example, a question about backup retention returned a fuzzy match after spelling correction: the handover procedure said to read a saved task "back". That result contained no retention policy. After checking all four source pages, the appropriate draft was: "The sources do not state a retention period. Ask the responsible person before promising one."

This is why an AI knowledge base does not eliminate incorrect answers. Read the returned section, check its scope and date, and distinguish a supported fact from a related word. Record an unanswered question instead of inventing a response time, legal term or product capability.

Which permissions and hosting limits matter?

Knowledge Base read and management rights belong to workspace roles. There is no separate access list for each page, and no field-level permission system. Store only material appropriate for the people and connected users who can read that Knowledge Base.

External MCP clients act with the connected user's permissions. The server's confirmation instructions guide the client; they do not create a server-enforced approval gate for an otherwise authorised tool call. Mate has its own approval policy in Customermates Cloud.

Hosted Mate depends on the eligible cloud plan, available usage and global controls. External clients use their own model provider, but semantic search can still count toward Mate usage. Self-hosting provides the Starter core and MCP for external clients, with keyword/full-text retrieval; it does not include hosted Mate or self-hosted semantic AI search. See self-hosting and the Mate documentation.

Text-bearing attached files can provide additional searchable material where supported. Scanned documents and images are not read through OCR. This is not a promise to index every shared drive, file type or external system.

When is another knowledge tool a better fit?

Local Markdown can be enough for one person's coding instructions or a small repository. A public help centre is a better fit for a customer-facing documentation portal. A dedicated knowledge or enterprise-search platform may be needed for broad drive synchronisation, large-scale document ingestion or more granular document access.

Customermates fits when the maintained guidance should sit beside customer records, conversations and the next piece of work. Start with a small, useful source set and keep its owner and review routine clear.