A company that remembers everything
Every process and role documented once, answering questions forever: for people and their AI alike.
The company only existed in people's heads
A multi-store retail company with no documentation at all. How an order is processed, who signs off what, which supplier's terms apply, what to do when a delivery arrives short: all of it lived in the memory of whoever had been there longest. Work stopped when they were on holiday. New starters learned the business by interrupting people for weeks. Every answer cost somebody else's afternoon, and nobody could point at the company and simply read it.
One place that holds the company in writing
The brain is a single workspace holding the operation itself: every process documented, every person's role and what it actually covers, and hundreds of supplier documents, guides and terms lifted out of drawers and inboxes. Core operations and inventory work were raised into versioned procedures with a named owner and a path to follow when something goes wrong. When someone asks how something works here, the answer is a page, not a person's memory.
The team's AI reads it too
This is the part that makes it a brain rather than a wiki. The whole team's AI accounts are connected to the workspace, so when anyone asks their assistant a question, the answer comes back grounded in the company's real processes, real roles and real data, not in generic advice about how retail businesses usually work. New starters onboard from it, the AI reasons from it, and every agent we build draws its context from it.
It speaks the company's shorthand
Every business runs on a private language: the grades, the trade terms, the nicknames for things that appear on no invoice and in no manual. The most valuable job the brain does is decoding it, so a request gets parsed the way a colleague would parse it rather than taken literally by something that has never stood in the stockroom.
It keeps itself current, and it knows when to stay quiet
A weekly agent scans the key pages for material change, and only material change. If something real moved, it updates the affected sections, writes a line in the changelog, and nudges the team to refresh. If nothing moved, it says nothing at all. And if it cannot be certain it is posting in the right place, it drafts the message for a human instead of broadcasting it. Restraint, designed in.
The writing is cheap, the trust is the build
A knowledge system nobody uses is a graveyard with better formatting. This one shipped as part of the company's first joined-up tech stack, on a one-month migration plan, with the CEO's sign-off and training run live in the room with every head of department rather than sent round as instructions. Adoption was designed, not hoped for.
- One month
- From no documentation at all → a live system the whole company works out of.
- Every department
- Heads trained in the room rather than sent instructions, with the CEO signed on before rollout.
- Hundreds of documents
- Supplier terms, guides and process knowledge out of drawers and inboxes → one place, searchable.
- Asked once
- "How does this work here?" is answered by a page instead of by the longest-serving person in the room.
For the person who wants the detail.
Everything above is what it does. This is how it is actually built.
- 01A single Notion workspace as the source of truth: processes, role definitions, and hundreds of supplier documents, guides and terms digitised out of paper and email.
- 02Core operations and inventory workflows raised to versioned procedures with named owners and explicit escalation paths.
- 03The team's Claude accounts connected directly to the workspace, so retrieval happens against the company's own pages and every answer is grounded in its actual processes and data.
- 04A two-tier memory system for the AI: a compact working memory loaded into every session (who everyone is, the company glossary, live project state, and a log of hard-won platform quirks), sitting over a library of deep-context documents pulled in on demand.
- 05A shorthand decoder for internal vocabulary: grades, trade terms and nicknames mapped so an agent parses a request the way a colleague would.
- 06A weekly context-pack agent diffs the key pages for material change, updates only the affected sections of a versioned pack, writes a changelog line, and posts a refresh nudge to Slack, staying silent when nothing moved.
- 07Posting is gated on certainty: if the agent cannot confirm the correct channel, it drafts for a human instead of broadcasting.
- 08The same workspace is the context source for the reporting fleet, so people and agents are reading one set of definitions rather than two.
What would you stop doing by hand?
Three questions, three minutes. We come to the first call already understanding the problem.