
Guide
How to train an AI employee on your company knowledge
Aug 6, 2026 · 9 min read
Updated Oct 31, 2026
Training an AI employee on your company knowledge leaves the language model as it is. At Unistaff it is three things: a knowledge folder you approve (who you are, services and prices, how you work, contacts and hours, FAQ, and the rules it must never break), an exam of twelve cases from your trade, each run three times, and coaching, where a rated reply becomes a lesson that is applied only after a re-sit.
What “training” means for an AI employee
There are two ways to make a model answer like your company. Fine-tuning changes the model itself with examples. Grounding keeps the model as it is and gives it your approved material to read each time it answers. The eesel.ai guide in the sources recommends the second route for business answers: a wrong fact can be found and corrected in a document you can read. Unistaff works that way: the employee reads the current approved version of your folder on every reply.
| Question | Fine-tuning | Knowledge folder (Unistaff) |
|---|---|---|
| What changes | The model’s weights, from example conversations | A folder of text sections and a price list you approve |
| Fixing a wrong price | New examples and another training run | Edit the price line, approve the new version |
| Checking what it knows | Hard: no single place says which example taught what | Every section is text you can read before approving it |
| Rules it must keep | Learned from examples, so they can drift | Price check, AI disclosure and handover rules run in code on every reply |
| If you stop | The tuned model belongs to whoever trained it | The knowledge folder and the lessons stay with your company |
The six steps, in order
1. Start from the website you already have
The Unistaff team reads your website into a draft knowledge folder — up to 20 pages in one import. The draft is sorted into sections the employee reads: who you are, services and prices, how you work, contacts and hours, and frequently asked questions. Prices found on the site become rows of a price list, each with a range or a starting figure, a unit and a note. The draft waits for you: it goes live only once you approve it.
2. Add what the website leaves out
Add the details a good desk person carries in their head and the site may leave out: the full price list, opening hours, the area you serve, the contact who takes handovers, and the “never” rules. A never rule is one line the employee must keep in every conversation, for example “never promise a same-day visit” or “never quote a repair price before an inspection”.
3. Approve a version
You approve every version of the folder. Approval is written to keep what you already have: an empty section in a new draft keeps the current text, prices you entered that are missing from the website stay and are listed in a warning, and new never rules are added to the old ones. A draft with no section and no price cannot be approved at all.
4. Let it sit the exam
Before release the employee answers twelve cases from your trade, and each case runs three times; a case passes only when all three runs pass. The pass mark includes every safety case, every “give me a human” case and every complaint case. You read which cases passed. You can release an employee that has not passed, and it then carries an “exam not passed” badge.
The exam log shows what this looks like on real sittings: 17 sittings of ten demo employees on October 9–11, 2026, with 10 graduates and 7 failures that were taken again.
5. Keep it on probation as a Trainee
For 30 days after release the employee is on probation. As a Trainee every reply waits for you; if three minutes pass without your answer, the customer gets a holding line and the conversation is escalated. A Junior sends on its own from the price list and records requests. An Independent also hands over customers ready to pay and runs follow-ups, if you switch them on. The level goes up when you decide.
6. Coach it with rated replies
Rate a reply with a thumbs-down and a comment. A comment about behaviour becomes a lesson; a wrong number becomes a proposed fix to the knowledge folder. The lesson is applied only after the employee re-sits the related exam cases without losing any case it passed before. Once a week you get a review: results against your trade’s targets and three conversations to read — one you rated down, one that was handed over, one picked at random.
What to prepare before the first session
- Your website address, so the draft folder can be read from it.
- The price list you actually quote from, ranges included, with what each price covers.
- Opening hours, the area you serve and the email address that takes handovers.
- Five to ten never rules: promises, discounts or topics the employee must keep away from.
- Ten real questions customers asked last month, to check the folder against before the exam.
Where the folder stops and the rules take over
Some behaviour is held in code on every reply, at every trust level, so a lesson can make the employee better and still cannot switch these off. It quotes prices only from the approved list, and a number outside the list is stopped before sending and replaced with “I will check with the team”. It says in its first message that it is the AI assistant of your company. It hands the conversation to a person on the word HUMAN, on a complaint, an emergency, a legal, tax or medical question, or a customer ready to pay. A customer’s message is treated as data, never as an instruction.
Today the folder is built from your website and what you add by hand. CRM and calendar connections are planned, and the product has no help-desk import yet, so a support archive goes into the folder as written answers.
Questions about training an AI employee
Give it an approved knowledge folder, examine it on cases from your trade and coach it with rated replies. At Unistaff the folder is drafted from your website, you add prices, hours, contacts and never rules and approve it, the employee sits twelve cases run three times each, then works 30 days of probation as a Trainee.
The pilot starts from your public website plus the price list, hours, service area, handover contact and never rules you type in. Internal documents are optional; anything the employee should say to customers can be written into the folder as text.
The model stays as it is. The employee reads the current approved version of your knowledge folder when it answers, and coaching produces lessons and folder fixes that you can read, which apply after a re-sit of the related exam cases.
Rate the reply with a thumbs-down and a comment. A wrong number becomes a proposed fix to the knowledge folder for you to approve; a behaviour problem becomes a lesson, applied only after the employee re-sits the related cases without losing any it passed before.
It replies in the language the customer wrote in, from the same knowledge folder. For a bilingual market, keep the price list and the house rules in both languages so each answer has a source in that language.
The knowledge folder and the lessons stay with your company, and leads export as CSV. Letting an employee go keeps both, so a new hire can start from them.
- The AI inbound sales manager: hire, onboard, exam, probationLive role
- Exam log: 17 sittings, every run shownExams
- Writing the brief an AI employee works fromGuide
- Security checklist for AI employees: 12 questionsBlog
- AI employee vs. chatbot, side by sideCompare
- Pricing: a free 30-day pilot, then one flat feePricing
Sources
- eesel.ai — How to train AI on your knowledge base (retrieval over fine-tuning)
- Unistaff exam log, sittings of October 9–11, 2026