SkynetLabs

AI-Drafted, Human-Approved Customer Updates for Trades

Published 17 November 20267 min read

A customer rings a plumber about a leaking tap. The job is booked for Thursday. On Thursday morning the technician is running late because the previous job turned up a corroded valve. Somebody now has to tell the customer, politely, with the right new time, before the customer starts wondering.

This is the kind of message an AI can draft in seconds and a person can approve in a few more. It is also the kind of message that goes wrong in a specific way when nobody checks it. This article shows what an AI-drafted, human-approved customer update looks like in practice for a plumbing, HVAC or electrical business: what the draft contains, where the approval step sits, what the AI should never decide and how to start without a big build.

What "drafted by AI, approved by a human" means

The assistant writes the message. A named person on your team reads it and sends it, edits it or throws it away. Nothing reaches the customer until that happens.

That is different from an assistant that replies on its own. For a first system, or for any message that carries a promise (a time, a price, a warranty), the approval step is the point. It is the same idea the OWASP Gen AI Security Project gives for language-model applications in general: OWASP's guidance on excessive agency says to require a human to approve high-impact actions before they are taken. A message that tells a customer you will be at their house at 2pm is a small action, but it is one your business has to keep.

A worked example (illustrative, not a real customer)

Everything below is made up to show the shape. The names, times and job are examples only.

What the team already knows. The job card says: leaking kitchen tap, booked Thursday 9:00 to 11:00, technician Sam. At 8:40 Sam marks the previous job as "running over" and types a note: "corroded valve, need another hour".

What the assistant sees. Only the fields you chose to share: the customer's first name, the booking window, the technician's first name, the delay note and your approved wording rules (more on those below).

What the assistant drafts.

Hi Priya, it's the team at Example Plumbing. Sam's previous job is taking longer than planned, so he will reach you later than the 9:00 to 11:00 window. We will confirm the new arrival time shortly. Sorry for the wait, and reply here if you need to change anything.

What the approver sees. The draft, the job card beside it, and three choices: send, edit, discard. The approver notices that the draft deliberately gives no new time, because nobody has confirmed one. They ring Sam, learn he can be there by 12:30, edit one sentence to say "around 12:30" and send.

That edit is the whole value of the approval step. The assistant did the part that costs time (tone, structure, consistent wording). The human supplied the one fact only they could confirm.

What the draft should contain

A good trade-business update is short and has a fixed shape:

  1. Who it is from, in the business's own name.
  2. What changed, in one sentence.
  3. What happens next and when you will confirm it.
  4. How to reply or reach a person.

Keep a short set of approved phrasings for the situations that repeat: running late, job finished, quote ready, parts on order, reschedule needed. The assistant then writes within those phrasings instead of improvising. This also keeps the voice consistent across technicians, which customers notice even if they never say so.

What the assistant should not decide

Draw the line in writing, before anything is built. A sensible starting list:

  • Prices and quotes. It can say a quote is being prepared. It does not state or change an amount.
  • Arrival times it cannot see. If nobody has entered a time, it says the team will confirm, as in the example above.
  • Warranty, fault or liability wording. These go to a person every time.
  • Urgent wording. A message that mentions flooding, a gas smell or no heating for a vulnerable person should not wait in an approval queue. It should route to your emergency number or a person at once.
  • Anything about another customer. The assistant gets one job's details, not the whole diary.

The last point matters more than it sounds. Customer messages are untrusted text. A customer can paste anything into a reply, including instructions aimed at the assistant. OWASP's prompt injection entry recommends human-in-the-loop controls for privileged operations and keeping untrusted external content separate from your instructions. In a plain customer-update flow that translates to two rules: the assistant cannot take actions on its own (it can only draft), and the draft it produces is always read by a person before it leaves.

Where the approval step lives

You do not need new software for this. Three common places:

  • A shared inbox or WhatsApp Business inbox. The draft appears as a suggested reply that a team member sends.
  • A message in your team chat. The draft is posted with the job reference, and a person copies, edits and sends it.
  • Inside your job-management tool. Where the tool supports it, the draft attaches to the job card. Whether that is possible depends on the tool, which is covered in how an AI receptionist works with ServiceM8 and Jobber.

Pick the place your team already looks at during the day. An approval queue nobody opens is worse than no assistant, because messages sit unsent.

The WhatsApp timing rule to know about

If your updates go out on WhatsApp, there is a platform rule that shapes how drafts are used. Twilio's documentation says a customer service window opens when the customer sends you a message and lasts 24 hours after the most recent message received. Inside the window you can send free-form messages. Outside it, you must use a pre-approved message template.

In practice: a reply to a customer who has just written to you can be a free-form, approved draft. A proactive "we are running late" message to someone who has not messaged for a few days usually needs a template that was approved in advance. So the approved phrasings above are useful twice: they keep the assistant on script, and they can double as the template wording you submit. Check the current rules with your messaging provider before you build, since platform terms change.

Measuring whether it works

Keep the measurement plain and your own. For two weeks, note for each draft whether the approver sent it as written, edited it or discarded it. If most drafts are sent unchanged, your approved phrasings are good and you can consider letting the lowest-risk message types go out after a quick glance. If many are edited for the same reason, fix the rule, not the person. Do not set a target in advance; the counts from your own jobs are the evidence.

How to start in a week

  1. List the five update messages your team writes most often.
  2. For each, write the approved phrasing and the facts the assistant may use.
  3. Write the "never decide" list above and add your emergency wording rule.
  4. Choose the approval place and name one approver per shift.
  5. Run it on drafts only for two weeks and count sent, edited and discarded.

If you want help scoping this, it is the kind of workflow we map in the AI audit: where messages are written today, who approves them and what a safe first version looks like. For the call side of the same problem, see AI receptionist vs answering service for trades.

FAQ

Is a human-approved draft slower than just sending the message?

Usually not for the person approving, because they read a finished message instead of writing one. The saving is in writing and consistency, not in removing the check. If a message type is routine and low risk, you can decide later to relax approval for that type only.

Can the AI send updates on its own once I trust it?

It can be set up that way, and some businesses will want that for simple confirmations. We would keep approval on anything that states a price, a time the team has not confirmed, or a warranty position, and on anything urgent. Autonomy should grow one message type at a time, based on your own sent, edited and discarded counts.

What happens if a customer replies with something unexpected?

The reply goes to a person. The assistant can draft a response, but unusual replies, complaints and anything that sounds urgent are flagged for your team rather than answered automatically.

Sources

  1. OWASP Gen AI Security Project, LLM06:2025 Excessive Agency, 2025. Accessed 2026-11-17. Lists excessive functionality, permissions and autonomy as root causes, and recommends requiring a human to approve high-impact actions.
  2. OWASP Gen AI Security Project, LLM01:2025 Prompt Injection, 2025. Accessed 2026-11-17. Recommends human-in-the-loop controls for privileged operations and separating untrusted external content.
  3. Twilio documentation, WhatsApp API with Twilio, undated documentation page. Accessed 2026-11-17. Describes the 24-hour customer service window and the need for a pre-approved template outside it.

See the AI audit