Train your AI Employee
Outcomes
Audit what it actually knows
You are here: Partner Center, on the same employee you deployed last step. Your receptionist already answers most questions well. Here you find the ten it gets wrong, fix them, and set a guardrail so it never invents a price.
Run the ten-question audit again, now that the employee is doing real work: open Chat and ask everything a prospect is likely to ask, prices included. A wrong answer is not a defect, it is a training gap, and every training gap has an address, either a knowledge source that needs fixing or an instruction it never received.
Open Chat in a fresh incognito window and run the audit before you change anything. Keep the answers that need fixing; the rest of this step fixes them.
Fix the source, not the AI
A wrong price usually means the source it searched never had the real number, not that the AI is broken. Copy the package's exact pricing into a clean document and upload it as the authoritative source: auto-imported marketplace content can carry broken links or no usable pricing at all.
Labupload a clean price list
Go to your employee's Configure page → Knowledge sources.
For every knowledge source type and what belongs in each, the Knowledge Base guide covers the rest.
Add a guardrail
Knowledge is what the AI looks up when a question calls for it. A rule that must hold on every response belongs in a capability instead, so a guardrail like "quote only from the published list, never estimate" holds even on a question the price list does not directly answer.
Labwrite the pricing guardrail
Go to your employee's Configure page → Capabilities.
Keep the instruction short. Everything an employee has been told shares one working memory, and a long list of rules competes with itself; specific and short beats long every time. Instructions also tend to collect leftover lines from earlier edits, so revisit and trim them now and then rather than only ever adding to them. The same discipline suits a business in a regulated field: rather than answering specifics on medical or legal questions, have the guardrail redirect to a person, and use the word its own customers would use, never a clinical term a business would not use itself.
Re-test and see why it answered that way
Labre-run the audit
Go back to Chat.
Coach it the way you coach a new hire
Treat the employee as a new hire rather than a setting: coach it, correct it, and expect the answers to keep improving, the same way a new team member does. The rule behind every fix here is the one behind every fix to come: a wrong answer to a specific question is usually a knowledge gap, and a wrong behavior across every conversation is usually a capability gap. Knowledge is the degree it holds; capabilities are the skill it practices.
Testing on purpose, on your own account, is how gaps surface before a client ever sees them, and a change that turns out wrong is never permanent: undo it and try again. As an employee earns your trust through this loop, what changes is how closely you review it, not a level it's set to: the next step builds one from a blank start and shows you exactly what it can do from day one.
What you now have
- A receptionist whose pricing answers come from one clean, current source
- A guardrail capability that stops it from ever guessing a price
- The habit of auditing after every knowledge change, with Explanation to show you why an answer happened
- The knowledge-versus-capability question, answered, for every fix you make from here on
Knowledge Check
Three quick questions on fixing a wrong answer, sorting knowledge from capabilities, and reading Explanation.