Key takeaways
- Text tools are the most common kind of AI in use[1], and admin is one of the top areas firms apply them to[2].
- Dictate facts on site, let the tool structure them, and keep measured readings, gas safety results and part numbers under the engineer's own control.
- Use a fixed prompt and a fixed report layout so every customer gets the same sections in the same order.
- Check every AI draft before it is sent; most firms that use AI already review its output[3].
- Personal data must be accurate and limited to what the job needs[4], so do not paste whole customer histories into a general chatbot.
Why notes are worth fixing first
Job notes are where a heating business quietly loses money. A vague note means a second visit to find out what was done, a warranty claim that stalls, or a customer who disputes the invoice because nothing explains the extra hour. Writing them properly at the end of a ten-hour day is the bit everyone skips.
That makes notes a sensible first use of AI. It is low risk, because a person reads the result before anyone else sees it, and the time saving is easy to measure. Construction is still a light user: 6.7% of UK construction businesses used large language model text generation in autumn 2026, against 19% of all businesses[5]. Among businesses that do use AI, 85% use natural language processing and text generation[1], and administration is one of the most common areas, alongside marketing[2].
Source: Department for Science, Innovation and Technology (GOV.UK), 2025[2]
Show the data
| Item | Value |
|---|---|
| Marketing | 72% |
| Administration | 72% |
| IT | 64% |
What a usable job note contains
Before any tool is involved, agree the fields. If the engineer's voice memo covers these, the AI only has to put them in order.
- Customer, address and appliance: make, model and location of the boiler, cylinder or heat pump.
- Reported fault in the customer's words, and what you found.
- Work done, step by step, with parts fitted and part numbers.
- Readings and test results, copied exactly from your instruments or forms.
- Advice given, including anything the customer declined.
- Follow-up needed: parts on order, return visit, quote to send.
- Photos taken and where they are stored.
From voice memo to finished note
Dictate before you leave the property
Spend two minutes on a voice memo while the details are fresh. Speak in the order of the field list. Read numbers out digit by digit.
Transcribe
Use the transcription built into your phone or job app. Fix obvious mishearings of model names before going further.
Structure with a fixed prompt
Paste the transcript into your approved AI tool with the same instruction every time (see the prompt below). Ask for the internal note and the customer summary as two separate outputs.
Check against the source
Compare every number, part and appliance detail with your forms and photos. Delete anything the tool added that you did not say.
File and send
Save the internal note to the job record. Send the customer summary only after the check.
A prompt that works well: Turn this transcript into two outputs. Output 1: an internal job note with headings Appliance, Fault reported, Findings, Work done, Parts, Readings, Advice, Follow-up. Copy numbers exactly as spoken and write MISSING where a heading has no information. Do not add anything not in the transcript. Output 2: a short, polite summary for the customer in plain English, under 150 words, with no readings and no prices.
The instruction to write MISSING is the most useful line in the prompt. It stops the tool filling gaps with plausible guesses.
What AI gets wrong in heating notes
| Error | How it shows up | Check |
|---|---|---|
| Invented detail | A flue check or pressure test appears that was never mentioned | Delete anything not in the transcript |
| Number drift | A reading or part number changes by a digit | Compare every number with the form or photo |
| Wrong appliance | Model names confused, or a combi described as a system boiler | Check against the data plate photo |
| Overconfident wording | Fault fixed when the engineer said probably fixed | Keep the engineer's own level of certainty |
| Advice softened | A warning about an at-risk appliance becomes a suggestion | Safety advice is copied word for word, never summarised |
| Tone mismatch | Customer summary sounds like a sales letter | Plain, short and factual |
Checking is standard practice. In 2025, 84% of UK businesses using AI said they checked outputs at least to some extent, and 67% reported significant checking[3]. For job notes, the person who did the work should do the check.
Records the engineer must own
Some records should never be drafted by a language model at all. Results on a landlord gas safety record, flue gas analysis, commissioning data and unsafe-situation paperwork come from the engineer's instruments and judgement. AI can remind you that a landlord must give tenants the record within 28 days of the check[6], but it should not fill in the record itself.
Commissioning records are also changing. The paper Benchmark checklist is no longer maintained and the record has moved online and to the Benchmark apps[7]. If your notes feed into that record, keep the numbers in one place and copy them across, rather than retyping from an AI summary.
Under consumer law, a service must be carried out with reasonable care and skill[8]. A clear note showing what was found, done and advised is often your best evidence that it was.
Data protection without the jargon
Job notes contain personal data: names, addresses, sometimes health details such as a customer who needs heat because of a medical condition. UK GDPR requires that data to be limited to what is necessary and kept accurate, with inaccurate data corrected without delay[4].
- Use a business account for any AI tool, with settings that stop your inputs being used for training where that option exists.
- Strip names and full addresses from transcripts before pasting them into a general tool; add them back in your job system.
- Do not record health or vulnerability details unless they change how the job is handled, and then keep them in the job system only.
- Tell customers in your privacy notice that you use software to prepare reports.
- Keep the original voice memo for a short, fixed period so you can settle disputes, then delete it.
Writing the customer report
Customers read reports to answer three questions: what was wrong, is it safe now, and do I need to do anything. Order the summary that way. Keep readings out unless the customer asks, but include any advice in full, and say plainly when a part is on order and what happens next.
Example (fictional) customer summary: We attended because your boiler was losing pressure. We found a weeping pressure relief valve and replaced it. The system now holds pressure. We also noticed the magnetic filter has not been cleaned for some time and recommend this at your next service. No further visit is needed unless the pressure drops again, in which case please call us.
Example: one engineer's week
Example (fictional): an engineer in Bristol does 26 jobs in a week. Before the change, notes took about 12 minutes each in the evening and four were too thin to support an invoice query. With two-minute voice memos and a fixed prompt, each note takes about five minutes including the check. Two drafts that week had a part number wrong by one digit and one invented a pressure test; all three were caught at the check because the engineer compared numbers with photos. The time saved is useful, but the bigger change is that every job now has a note.
Rolling it out across the team
Introduce the change with one engineer first, ideally the one who already writes the best notes. Their transcripts show what good dictation sounds like, and their checks will catch the tool's habits early. After two weeks, write down the mistakes they found and add a line to the prompt for each one.
- Give every engineer the same field list on a laminated card or in the job app.
- Share three model voice memos: a breakdown, a service and a landlord check.
- Agree that any note the office cannot understand goes back to the engineer the same day, not the next week.
- Review ten random notes a month for missing fields, invented detail and tone.
- Keep the prompt in one shared place so nobody works from an old version.
Expect resistance from engineers who see notes as office work. The argument that usually lands is practical: a good note is what stops a customer disputing a bill or a manufacturer refusing a warranty claim, and the voice memo takes less time than typing on a phone in the van.
Measuring whether it helps
Pick three measures before you start and record them for a month either side of the change: average minutes per note, the share of notes with every field filled in, and the number of follow-up calls from the office asking engineers what they did. If the first measure improves but the other two do not, the tool is producing faster notes without better ones, and the prompt or the check needs work.
Also count corrections at the check stage. A falling number means the prompt is improving; a sudden rise usually means someone has started using a different tool or an old prompt.
Where this fits with other tools
Once notes are reliable, the same habits carry over to AI drafts of review replies and to scheduling tools that read job descriptions. Before you widen use, read the main AI risks for a trade business and agree who in the firm signs off each kind of output.
Frequently asked questions
Can I let AI fill in a gas safety record?
No. Results on a gas safety record, commissioning data and unsafe-situation forms come from the engineer's instruments and judgement. AI can help with the customer summary around them, not the record itself.
Which AI tool should we use for job notes?
Any tool you can use through a business account with clear data settings. The prompt, the field list and the check matter more than the brand.
How long should a job note be?
Long enough that another engineer could pick up the job without phoning the first. In practice that is the field list filled in, with MISSING where something genuinely was not recorded.
Is it safe to put customer addresses into an AI tool?
Only with a business account and settings you have checked. A simpler habit is to remove names and addresses before pasting and add them back in your job system.
References
Department for Science, Innovation and Technology (GOV.UK) (2025). AI adoption survey report 2025 (opens in a new tab) ↑
Department for Science, Innovation and Technology (GOV.UK) (2025). AI adoption survey report 2025 (opens in a new tab) ↑
Department for Science, Innovation and Technology (GOV.UK) (2025). AI adoption survey report 2025 (opens in a new tab) ↑
legislation.gov.uk (2016). Regulation (EU) 2016/679 (UK GDPR), Article 5: principles relating to processing of personal data (opens in a new tab) ↑
Office for National Statistics (2026). Business insights and impact on the UK economy: wave 165 dataset (AI tables) (opens in a new tab) ↑
legislation.gov.uk (The National Archives) (1998). The Gas Safety (Installation and Use) Regulations 1998, regulation 36 (opens in a new tab) ↑
Heating and Hotwater Industry Council (HHIC) (2026). Benchmark goes digital - what homeowners need to know (opens in a new tab) ↑
legislation.gov.uk (The National Archives) (2015). Consumer Rights Act 2015, section 49 (opens in a new tab) ↑