Training a New Office Hire to Use AI Tools Responsibly

Why this matters

A new office hire will use a chat assistant in their first week whether you train them or not. Most arrive already fluent with one from school or a previous job, and fluent is the problem: they know how to get a polished paragraph out of it and have never been the person who answers for what that paragraph promised a customer. The first time an untrained hire sends a confirmation that quotes the wrong warranty term, the customer holds the shop to it, not the tool. A week of structured training costs a few supervised hours and removes most of that risk before it ships.

This is a first-week sequence, day by day, because the order is the method: rules before tools, supervised use before solo use, and a test of the one skill that matters before anyone signs off.

What the week is actually training

Not prompting. A new hire who writes clumsy prompts produces clumsy drafts that are easy to fix. The skill being trained is knowing which output is cheap to check, which is expensive to check, and which must never go out unchecked. A language model produces fluent text by predicting what usually comes next, which means it will state a wrong date, a wrong price or an invented policy in exactly the same confident tone as a right one. The mechanism is covered in its own card (see related: What a Language Model Is Actually Doing When It Answers); the hire needs one sentence of it, and that sentence is "the tool sounds equally sure when it is wrong."

The verification standard itself belongs to its own card too (see related: The Verify-Before-You-Send Rule for AI Output). This week is how a new person comes to actually apply it.

Day 1: the policy, read together, with the hire's own examples

Sit down with the hire and read the shop's AI use policy aloud, section by section. Do not hand it over to read alone. Reading it together forces you to explain the reasons, and the reasons are what the hire remembers when a situation is not on the page.

The policy itself, and especially the list of what may and may not go into a tool, is owned by its own SOP (see related: The Shop AI Use Policy SOP). On day 1 your job is to make it concrete:

• Have the hire name three things they would have pasted in. Most will say a customer complaint email, a past-due invoice list, or a tech's job notes. Walk each one against the policy. The customer complaint email contains a name, an address and a phone number; the policy decides whether that can go in and in what form.
• Show them which tools are approved and which account to use. A hire using a personal account on a personal phone has stepped outside every setting you configured.
• Tell them who to ask. One name, not "anyone in the office." If the answer to "can I paste this" is "ask Maria," Maria needs to know she is the answer.

What breaks if you skip it: the hire learns the policy from watching others, which means they learn the shortcuts others have taken, not the rules.

Days 2 and 3: three supervised tasks, rising in risk

Pick three real tasks from the hire's actual job, ordered from lowest to highest consequence if wrong. The hire drafts with the tool; you watch them do it and review before anything leaves.

Task 1, low risk: tidy an internal note. A tech's rough job note turned into readable text for the job record. Nothing goes to a customer. The point is to watch how the hire prompts and whether they read the output or skim it. Most hires skim. Note it, say nothing yet.

Task 2, medium risk: a routine customer message. An appointment confirmation or a reschedule notice. Now there are facts that must be right: the date, the arrival window, the tech's name, the address. Ask the hire, before they send it to you, to point at every fact in the draft and say where they checked it. If they cannot point to the job record for the arrival window, the draft is not done.

Task 3, higher risk: a message that touches money or scope. A follow-up on an estimate, or a reply to a customer asking what their repair includes. These are the drafts where the tool will fill a gap with something plausible: a warranty length it has never seen, a "free" trip charge, a phrase like "fully covered." Have the hire draft it, then review it with them line by line, and make them find the gaps before you point them out.

Space the three across two days so the hire is doing the work, not watching you. One task a half-day is about right; faster than that and they are rushing to please you.

What breaks if you skip it: the hire's first unsupervised money message is also their first money message, and you find the error in the customer's reply.

Day 4: the verification drill on a planted error

This is the day that matters. Prepare, in advance, a draft customer message that looks like tool output and contains planted errors. The hire is told only that it is a draft and that they should check it before it goes out. They are not told it is a test, and they are not told how many errors there are.

Plant four, each a different kind:

• A fact that contradicts the job record (wrong date or wrong arrival window).
• An invented policy (a guarantee or warranty term the shop does not offer).
• A number that does not add up (two line items that do not sum to the stated total of hours or visits).
• A soft promise ("we will definitely have it fixed on the first visit").

Give the hire the job record, the shop's warranty sheet and fifteen minutes. Then go through what they found.

DRAFT - reply to customer, job record attached

Hi Mrs. Okafor,

Thanks for your patience. Your tech, Luis, will
arrive Thursday between 8 and 10. As discussed,
the repair includes 2 hours of labor plus the
1.5 hour return visit, 4 hours total. Like all
our repairs, it carries our 2-year parts and
labor guarantee, and Luis will definitely have
it running on the first visit.

Thanks again,
Front Office

In this draft the job record says Friday, 10 to 12; the shop's warranty sheet says one year on labor; 2 plus 1.5 is 3.5, not 4; and "definitely" is a promise nobody can keep.

What breaks if you skip it: you will never know whether the hire checks or just reads. Those look the same from outside until something ships.

Day 4, worked: how the drill actually goes

Here is a realistic first run with the draft above.

The hire, call her Dana, took nine of her fifteen minutes and returned the draft with two corrections: she had fixed the day and arrival window against the job record, and she had changed "definitely" to "expects to." She had not caught the warranty term or the arithmetic. Two of four, which is typical. Most first runs land at one or two.

The pattern in what she missed is the lesson. She checked the facts she had been told to check on day 2, the ones that live in the job record. She did not check the facts that live in a different document (the warranty sheet was on the desk, unopened), and she did not recompute the total because "the tool did the math." Both are the same mistake: treating the tool's output as checked wherever she had not been specifically told to check it.

The debrief took ten minutes and went like this. You do not tell her what she missed. You ask her to list every factual claim in the draft, one per line. She lists seven: Mrs. Okafor's name, Luis, the day, the window, the labor hours, the warranty term, the first-visit claim. Then you ask where each one is checked. She can point to a source for four (name, tech, day, window). Of the three left, she had already softened the first-visit claim. She has no source for the labor arithmetic, which she can check herself in five seconds, and none for the warranty, which she cannot check without the warranty sheet. She finds both errors herself, in under two minutes, once the question is "where is this checked" rather than "does this look right."

Then run a second planted draft the same afternoon, different job, four new errors of the same four kinds. Dana caught four of four in eleven minutes. That second run is what you are actually measuring. A hire who goes from two to four in one afternoon has learned the habit. A hire still at two on the second run needs another day of supervised money messages before solo use, and that is not a failure of the hire, it is the training doing its job.

What would change the call. If the hire's role is purely scheduling and they will never draft anything about money or scope, the drill can drop the warranty and arithmetic plants and test facts against the schedule only. If the hire will draft anything a customer could treat as a promise about price, coverage or safety, all four kinds stay in, and the second run has to come back clean before sign-off.

Day 5: sign-off, and what the hire is cleared to do alone

Sign-off is a short written record, not a feeling. It says which tasks the hire may now do with the tool unsupervised and which still go through a reviewer. A reasonable first-week clearance:

How each tier is reviewed in daily practice is its own SOP (see related: The AI-Drafted Customer Message Review SOP). The sign-off says which tier the hire has reached, signed by the hire and the trainer, dated, kept in their personnel file. Revisit it at 30 days with one more planted draft, unannounced.

What breaks if you skip it: clearance becomes whatever the hire assumes, and the hire assumes more than you would have granted.

What a new hire gets wrong first

These are the patterns that show up in the first two weeks, in roughly the order they appear. Each has a specific correction.

• Pasting the whole thing. The hire pastes an entire customer email thread, signature blocks and all, because it is faster than trimming. Correction: show them the trim on their own example from day 1, and make the trimmed version the template they copy from.
• Treating the tool's math as done. Exactly Dana's miss. Correction: any number in a draft is recomputed by the person, every time, or comes from the job software and not the tool.
• Over-polishing. Drafts start sounding like a bank. Customers notice, and a shop that writes "we sincerely apologize for any inconvenience this may have caused" to a homeowner whose basement flooded reads as not caring. Correction: tell the hire to ask the tool for "short, plain, how a person in the office would say it," and then cut it further by hand.
• Asking the tool what the policy is. The hire types "what is our warranty" and the tool answers with a typical warranty. It does not know yours. Correction: the tool drafts; the shop's documents decide. Anything that states a policy gets checked against the policy document.
• Hiding use. A hire who suspects AI is frowned on will use it quietly and skip review to avoid the conversation. Correction: say plainly, on day 1, that using the approved tool is expected, and that skipping review is the only thing that gets someone in trouble.

References

• See related: The Shop AI Use Policy SOP
• See related: The Verify-Before-You-Send Rule for AI Output
• See related: The AI-Drafted Customer Message Review SOP
• See related: What a Language Model Is Actually Doing When It Answers
• See related: Using AI to Draft a Hard Customer Email Without Sounding Like a Robot