You do not need to digitise the entire company before AI becomes useful. For most teams, the quickest improvement comes from a compact context pack: five maintained files that explain the business, the customer, the way work is done, the standard of communication and the decisions already made.
This pack can sit inside a ChatGPT or Claude Project, an approved company drive, Obsidian or another controlled knowledge system. The software matters less than the discipline.
Business snapshot
State what the company does, who it serves, current priorities, key constraints, regions, terminology and named owners. Keep it concise enough to review in ten minutes.
Include: current offer, audience, commercial model, team roles, active priorities and words the business uses differently from the market.
Customers and offers
Describe each core customer, the problem they are trying to solve, what they buy, what is included, what is excluded and the evidence supporting the offer.
Include: qualification rules, pricing source, objections, proof, service boundaries and escalation points.
Ways of working
Document one real workflow at a time. Show the trigger, inputs, actions, decisions, owner, approval and completion condition. Do not write an idealised process that nobody follows.
Include: systems, hand-offs, exceptions, permission levels and what must never be automated.
Voice and standards
Give the AI examples of communication the business would approve. Explain audience, tone, structure, banned claims, formatting and accessibility expectations.
Include: good examples, bad examples with reasons, required disclaimers and a short quality checklist.
Decision and evidence log
Record durable decisions with a date, owner, reason and source. This prevents an old chat or proposal from quietly becoming the new truth.
Include: decision, status, evidence, affected work, review date and superseded information.
How to make the pack reliable.
Give every file an owner.
Unowned knowledge decays. A named person should approve changes and check the file on a sensible cadence. Frequently changing commercial information may need weekly review; a stable brand principle may need quarterly review.
Separate source from summary.
A summary helps the model work quickly, but it should point to the approved underlying source. If they disagree, the system needs a rule for which wins and a human needs to resolve the conflict.
Label sensitivity and permission.
Not every useful document belongs in every AI workspace. Mark files public, internal, confidential or restricted. Give each task only the access it needs. A sales-drafting assistant rarely needs payroll, legal disputes or the full customer database.
Prefer examples over adjectives.
“Sound professional” is vague. Two approved emails, one rejected email and a note explaining the difference give the model a usable standard. Examples are especially valuable for tone, classification and judgement calls.
If nobody can say who owns a file, when it was checked and what it replaces, it is reference material, not business truth.
A one-hour starting session.
- Choose one recurring business task with a clear owner.
- Create the five files using only material needed for that task.
- Mark uncertain statements with CHECK rather than guessing.
- Upload the pack to an approved AI project and explain source priority.
- Test three ordinary cases and two awkward cases.
- Record what failed, amend the pack and test again.
The result is not a finished business brain. It is something more useful: a controlled first version that can improve from real work.
Sources and further reading
Use the ready-made structure.
The free starter kit contains the folder map, templates, maintenance rules and first-week checklist.
Download the starter kit