How to teach an AI agent your business — without prompt engineering
· 2 min read
An AI agent is only as good as what it knows about your business. The common mistake is to paste everything into one long prompt and hope. A better way is to treat knowledge like a new employee’s onboarding: give them the material, check what they understood, and correct them as they go.
Start with what you already have
Your website, price lists, FAQ, delivery and return policies already describe your business. Let the agent read them: in Sales Core AI you point it at your website or upload documents, and it sorts the facts into topics — about the company, products and services, prices and payment, delivery, hours and contacts, terms, FAQ, objections.
Approve before the agent uses it
What the import finds becomes drafts. The agent doesn’t see a draft until you approve it — so an outdated page or a misread price doesn’t reach customers. Reviewing drafts takes a few minutes, and you can edit any of them before approving.
Keep facts and rules apart. Facts (“delivery to the islands takes 3 days”) belong in knowledge. Rules (“never promise same-day delivery”) belong in the agent’s instructions. Mixing them makes both harder to maintain.
Let customers show you the gaps
No import covers everything. When a customer asks something the knowledge doesn’t answer, the agent says it doesn’t know instead of guessing — and the question lands on a list with how many times it was asked. Answer it once, and every agent knows it from then on. After a couple of weeks the list gets short.
Correct replies in plain words
Sometimes the facts are right but the reply isn’t: too long, too pushy, offering a bundle the customer doesn’t need. Under any reply you can say what was wrong, like you would to a colleague: “don’t offer the bundle when the customer already has a grinder”. The fix goes where it belongs — a missing fact into knowledge, a rule into the agent’s behaviour.
Test after every change
Save the conversations that matter as tests: the common question, the tricky one, the one that went wrong last week. Each test says what a good result is and which words must or must not appear. Run them after a change — in a sandbox that never creates real orders or bookings — and see at once if something broke.
A weekly routine
- Approve or fix new drafts.
- Answer the unanswered questions.
- Read a few real conversations and correct one or two replies.
- Run the saved tests.
Twenty minutes a week is usually enough to keep an agent sharp. More on why facts must come from your data in how to stop an AI from inventing prices.