Most AI advice for small business is written by people who have never had to answer a phone at 7am with a van already loaded.
TL;DR Use it where a human checks the output before it leaves. Drafting, triage, transcription and repurposing. Do not let it speak to customers unsupervised, and do not let it near your pricing without review.
The rule that decides everything
Does a human see the output before a customer does?
If yes, AI is usually worth it. If no, the failure mode is a customer receiving something wrong in your name, and that costs more than the time saved.
| Task | Human checks first? | Verdict |
|---|---|---|
| Drafting a quote | Yes | Good use |
| Summarising a call | Yes | Good use |
| Sorting enquiries by urgency | Yes, before action | Good use |
| Turning one article into five posts | Yes | Good use |
| Answering the phone out of hours | Partially | Careful |
| Sending pricing directly to a customer | No | Avoid |
| Handling a complaint | No | Avoid |
Proposal and estimate drafting
Where it saves real time. Not calculating the price, writing the words around it.
The workflow
- You dictate the notes on site. “Combi swap, 1930s semi, existing pipework needs replacing to the airing cupboard, two days, awkward loft access.”
- AI drafts the proposal using your template and your standard clauses.
- You add the price yourself.
- You read it, correct it, send it.
What this saves. Twenty to forty minutes per proposal, mostly on the descriptive sections you rewrite from scratch every time.
What it must not do. Generate prices, invent specifications, or promise timescales you did not state. Give it your template and your notes, not a free hand.
Call and meeting summaries
The highest-return low-risk use.
Record the site visit or the call, with consent, and get a structured summary out: what was discussed, what was agreed, what needs quoting, what the customer was worried about.
Why it works. The notes get written, which they otherwise do not. The information that usually lives in one person’s head ends up in the record.
Consent matters. Recording rules vary by jurisdiction and some require all parties to agree. Ask, plainly, and note the answer.
Enquiry triage
Sorting, not answering.
Incoming enquiries get classified: emergency, quotable job, existing customer, out of area, supplier, spam. Tagged and routed accordingly.
What this fixes. The emergency sitting unread in a shared inbox at 8am behind fourteen marketing emails.
Keep the action human. The AI sorts. A person decides.
Build in an escalation default. Anything it cannot classify goes to the top of the pile, not the bottom. Uncertain should mean urgent, not ignored.
Content repurposing
One piece of work becomes several.
A completed job with photos becomes a social post, a newsletter item, a website case study and a short video script.
Rules that keep it usable
- Feed it your own material. Your notes, your photos, your voice. Not a topic prompt.
- Give it your voice rules. No exclamation marks, no corporate padding, whatever your standards are.
- Never let it generate a statistic. This is the single biggest risk with AI-written marketing. It will produce plausible, specific, entirely fabricated numbers and they will look exactly like real ones.
- Rewrite the first and last line yourself. Openings and closings are where generated text reads most obviously as generated.
The statistics problem, in detail
This deserves its own warning.
AI models generate text that resembles true statements. A fabricated statistic looks identical to a real one: specific figure, plausible source, confident phrasing.
Practical policy
- No number goes into published material without a named primary source you have personally opened.
- If it cannot be verified, describe the direction instead. “Response speed materially affects conversion” rather than an invented percentage.
- Be suspicious of round, memorable figures, particularly ones attributed to a well-known institution with no link.
This is not a theoretical risk. A single fabricated statistic in a published article damages your credibility more than the article gained you.
Where not to use it
Complaints. A generated apology is detectable and makes an upset customer angrier.
Pricing decisions. It has no idea what your costs are.
Anything legal or compliance-related. Certificates, regulations, contract terms.
Fully autonomous customer contact. Covered next, with caveats.
Getting started without a project
Pick one task you do more than five times a week and do it with AI for a fortnight.
Most likely candidates: proposal drafting, or turning job notes into a follow-up email.
Measure the time honestly, including the correcting. If editing takes as long as writing did, the task is a bad fit or the prompt needs work. Usually it is the prompt.
Do not buy a platform first. Start with a general assistant and your existing templates. Buy tooling only when you know exactly which repeated task justifies it.
Pick your most-repeated writing task this week and run it through an assistant with your own template attached. Measure the time saved after correction. That single test tells you more than any amount of reading.
Need a pro to set it up? [BOOK A CALL]