How to Train an AI Employee (Without Writing Prompts)
Training an AI employee is closer to onboarding a new hire than to programming a system. You give it the material, you listen to it work, and you correct it out loud. There are no prompts to engineer. Here is the routine that gets an agent from useless to reliable in about a week.
Day one: give it the material you already have
You do not have to write anything new. Almost every business already has the training material sitting somewhere:
- The price list, including the awkward exceptions.
- Opening hours, holidays, and the service area.
- The intake questions your team asks on every job.
- The FAQ page nobody has updated since 2023 — it is still mostly right.
- The five objections your team hears every week, and the answers that work.
Upload those, pick a voice, and the agent has a first draft of the job. Expect it to be roughly seventy percent right on day one. The last thirty percent is what the week is for.
Day two to four: call it and be difficult
Ring your own agent and behave like your worst caller. Mumble. Interrupt. Change your mind halfway through. Ask the price of the thing you never advertise. Ask for a Sunday. Ask for a discount.
Write down every answer that was wrong, vague, or too eager, and fix them in plain English: "never quote a price for a commercial job, take the details and book a survey". That is the whole editing interface. No prompt syntax, no variables.
The most common early problem is not that the agent is wrong; it is that it is too willing. Agents in their first week tend to say yes to things your business does not actually do. Tightening what it may promise is usually the highest-value edit of the entire launch.
Day five to ten: supervised, on real calls
Now let it answer real callers, but read every transcript at the end of the day. You are looking for three things:
- Wrong answers. Rare after the difficult-caller week, but fix them immediately.
- Avoidable escalations. The agent handed off because it did not know something small. Every one is a missing line in the script, and adding it converts a hand-off into a handled call.
- Questions you did not expect. These are the gold. Real callers ask things no internal document ever covered, and this is how you find out.
Nothing reaches a customer before you switch the role live, and a role can go back to supervised at any time.
After launch: a fifteen-minute weekly habit
An agent does not drift on its own, but your business does. Prices change, a service is dropped, a new location opens, a competitor starts running a promotion your callers ask about.
Fifteen minutes a week is enough: skim the escalations, read three random transcripts, and fix anything that reads badly. Businesses that do this have agents that get better every month. Businesses that set one up and never look at it again have an agent that is quietly six months out of date.
Frequently asked
Do I need to know how to write prompts?
No. You correct the agent in the same plain English you would use with a new hire, and Manta turns that into the script the agent follows.
How long until an agent is genuinely reliable?
Most businesses are comfortable after a week of supervised launch. The first two days are the difficult-caller test; the rest is reading transcripts.
Can I roll back a change?
Yes. Script changes are versioned, so a bad edit is reversible and you can compare how the agent handled the same question before and after.
Can I train more than one agent on the same material?
Yes. Shared material — prices, hours, policies — is reused across roles, and each role adds only what is specific to its job.
What if my prices change?
Update the material and the agent quotes the new number from the next call. Prices live in your script, never in the model's memory.
Can my staff correct it, or only me?
Anyone you give access to. The people who answer the phone today are usually the best trainers, because they already know which questions are hard.