- The jobs worth automating in a trade are the ones that happen after the van door shuts, not the ones that happen on site.
- When researchers graded eight systems that write notes from a recorded conversation, the dominant error was omission, between 8.5% and 24% of the content that should have been there.
- Omission is the dangerous failure for a trade, because a job sheet that is missing a detail looks complete and reads fine.
- Gas work in the UK is a licensed activity under 1998 regulations, so no amount of software changes who may legally do it or sign for it.
- Sole traders over the income threshold have been inside Making Tax Digital since 6 April 2026, which makes the bookkeeping job mandatory rather than optional.
A one van business has two jobs. One of them is the trade, and it pays. The other starts when the van door shuts: the quote that has to go out tonight, the certificate that has to be issued, the invoice from three weeks ago that nobody has paid, the call that came in while you had both hands inside a boiler.
Almost everything useful that AI does for a trade sits in that second job. Almost everything risky about it comes from people pointing it at the first. Working out where the line falls is worth an evening, because the answer is unusually clear in this line of work: some of it is a time saving, and some of it is a licensing offence.
Which jobs can software actually take?
The ones with a written output and a forgiving failure mode. Anything where being wrong is recoverable by a quick read, and anything where the alternative is that it does not get done at all.
| Job | What AI does well | Where it goes wrong | Who still has to be in it |
|---|---|---|---|
| Answering the phone | Takes a name, a number, an address and a description while you are up a ladder | Mishears a postcode or a part number, and sounds confident doing it | You, reading the message before you call back |
| Writing the quote | Turns a voice note and some measurements into a readable document in minutes | Leaves out an exclusion you said out loud, which the customer later treats as included | You, checking the scope and the price before it is sent |
| Job notes and sheets | Converts a spoken walk round into a structured record | Omits detail rather than inventing it, which is harder to notice | You, adding what is missing while the job is fresh |
| Scheduling and routing | Sequences a day of calls sensibly and reshuffles when one overruns | Does not know that one customer needs an hour of talking | You, for anything with a human constraint |
| Invoicing and chasing | Drafts and sends on a schedule, then follows up without embarrassment | Chases the wrong person, or a bill already paid by bank transfer | You, before anything is sent to a customer you want to keep |
| Certification and sign off | Nothing it may lawfully do | Everything, if you let it | A registered person, and only a registered person |
That bottom row is not a caution, it is a legal boundary, and we come back to it below. Everything above it is a judgement call about how much checking you are willing to do.
Why do the notes matter more than the model?
Because the characteristic failure of a system that writes up a conversation is leaving something out, and a job sheet with a gap in it looks exactly like a complete one. Researchers measured this carefully in a field where the stakes are high enough to fund the study.
A 2026 evaluation took eight systems that generate structured notes from a recorded conversation, broke the output into individual facts, and compared them against notes written by the professional who was in the room. Omissions were the main source of degradation, running from 8.5% to 24.0% depending on the system. Outright contradictions were rarer and more stable, between 6.2% and 7.9%.
Two further findings transfer directly to a van. About 92% of the content that first looked like invention turned out to be supported by the transcript, which means most of what reads as the machine making things up is actually the machine reporting something that was said and the reader had forgotten. And repetition rescued roughly 38.6% of errors: when something was said more than once in the conversation, the system was much more likely to get it right.
What does that mean on a site visit?
Say the important things twice, out loud, in different words. If you are recording a walk round so the software can write it up, repeating the boiler model, the access restriction and the thing you are explicitly not quoting for is not pedantry, it is the single cheapest accuracy control available, and the research says it works.
The second habit is to read for what is absent rather than for what is wrong. Scanning a generated job sheet, the eye checks whether the sentences are true. The error you are looking for is the sentence that is not there. This is the same problem transcription has at the audio layer, which we went through in what a transcription accuracy number actually hides.
What is legally reserved to a certified person?
In the UK, gas work, and the boundary is drawn by statute rather than by custom. The official listing is blunt about it: registering with Gas Safe Register is a legal requirement for anyone carrying out gas work in the United Kingdom.
The detail is worth knowing precisely, because it shapes what can and cannot be delegated to anything. The regulator is the Health and Safety Executive, the legislation is the Gas Safety (Installation and Use) Regulations 1998, and the regulation type is a licence. Registration is not general: an applicant holds certificates of competence for the specific areas of gas work they want to be registered for, from an accredited certification body. A person is registered for particular work, not for gas in the abstract.
So a certificate is an attestation by a named, licensed human that they inspected a thing and found it in a given state. Software can lay out the document, pull the appliance details across and remind you which fields are blank. It cannot form the opinion the document records, and nothing about the output of a model changes who is answerable if that opinion was wrong. The same logic holds in most countries that license this trade, and it holds for electrical certification in the UK too.
The practical version of this rule is simple to apply. If a document ends with your registration number on it, a machine may help you write it and may not help you decide it. If it ends with a price, the machine may draft it and you must still read it.
Is the paperwork becoming compulsory anyway?
For a lot of sole traders in the UK, yes, and that changes the economics of the bookkeeping job specifically. Digital records stopped being a preference and became a filing requirement.
The eligibility rules set out three steps. A sole trader or landlord registered for Self Assessment with qualifying income above 50,000 pounds for the 2024 to 2025 tax year should have started using Making Tax Digital for Income Tax from 6 April 2026. The threshold falls to 30,000 pounds for the 2025 to 2026 tax year, with use required from 6 April 2027, and to 20,000 pounds for 2026 to 2027, required from 6 April 2028.
Two details catch people out. The test is on qualifying income, which for a trade means turnover rather than what is left after materials and the van, so a business that does not feel like it earns 50,000 pounds can easily be over the line. And the threshold is read from an earlier tax year than the one you start reporting in, so the decision is made on figures you have already filed.
None of that requires AI. It does mean the receipts and the mileage have to live somewhere structured, every quarter, and once they do, the marginal cost of letting software categorise them is close to zero. If you have not set that side up, the sequence and the traps are in running the books of a one person business with AI.
How do you check a generated job sheet quickly?
Against the job rather than against the text, and in under two minutes. The temptation is to reread the document looking for mistakes, which is the one method the research suggests will fail you, because the sentences that are there are mostly fine.
Work from a short mental list of the things that must appear on every sheet in your trade, and tick them off. For most jobs that is the address and access notes, the appliance or circuit identified properly, what you found, what you did, what you did not do and why, parts used, and anything the customer must act on. If an item on your list is not in the document, add it before the memory fades. The list is fixed, so the check does not get slower as the documents get longer.
The study behind this is worth reading as a warning about a specific instinct. When the researchers looked at the content that appeared to have been invented, roughly 92% of it turned out to be present in the transcript after all. Applied to a van, that means the sentence that surprises you on a job sheet is usually something you said and forgot, not something the software dreamt up. Deleting it because it looks unfamiliar is as much an error as leaving a gap.
Does the customer need to know a machine wrote it?
For a quote or a job sheet that you have read and are standing behind, you are the author and the tool is a typewriter. The disclosure question gets real when the software is talking to the customer directly rather than drafting for you, and the rules on that are genuinely different, which we covered in when you have to tell people they are dealing with a machine.
Where do the photographs of a customer's home end up?
Somewhere you should be able to name before you take the next one. A trade accumulates an unusual archive: the inside of people's houses, their meter cupboards, occasionally their post on a worktop, and increasingly all of it is being uploaded to something that writes a report from it.
That is personal data belonging to the householder, held by you, and the fact that it was gathered in order to fix a boiler does not change what it is. The questions to answer are short. Which app holds the photographs, for how long, and can you delete them on request. Whether that vendor uses customer content to improve its own systems, and whether that setting is on by default. And whether your phone is syncing the job folder to a personal cloud account that your business has no control over, which is the most common quiet failure of the lot.
None of this is specific to a trade, but the exposure is, because the images are of homes rather than of products. The general rules on what may be pasted into a tool and what may not are in what you can safely put into an AI tool, and the trade specific version is to assume a customer will one day ask where the picture of their kitchen is kept.
What should you automate first?
The phone, because a missed call is the only one of these failures that costs you a job you would otherwise have won. Everything else on the list costs you an evening.
The argument is about asymmetry rather than about the technology. A badly drafted quote can be rewritten before it goes out. An invoice chased a day late is still chased. A call that rings out while you are under a sink is a customer who phones the next name on the list, and you never learn it happened. Anything that converts a ring into a written message with a number attached has already paid for itself, and the practical questions about doing it without sounding like a call centre are in what happens when AI answers your business phone.
Second is the quote, with a caveat that is specific to trades. A quote is not marketing copy. In most places a written quotation that the customer accepts becomes the basis of a contract, including the parts you did not think about, and a generated document that silently drops an exclusion has created an obligation rather than a draft. We set out where that line falls in what happens when the estimate your AI wrote becomes a promise.
Third is the chasing, which is the job most sole traders hate most and therefore do worst. Software has no feelings about sending a third reminder, and consistency is most of what collection is.
And what should you leave alone?
Anything that produces a legal attestation, as above. Anything where the value is that you personally turned up, which includes most of the conversation at the end of a job. And, for now, anything that would have your voice on it without your knowledge, because a customer who works out that the voice reassuring them about a leak was synthetic has learned something about you that no time saving repairs.
Does AI for tradespeople actually save a measurable hour?
It saves a fairly predictable one on documentation and a much less predictable one everywhere else. The honest framing is that it compresses the evening admin rather than changing the day.
Work out your own number before buying anything, because the arithmetic is specific to how you already work. Count the quotes you send in a month and the minutes each takes end to end. Count the invoices and the chasing. Count the calls you missed last week, if your phone will tell you. That list is your addressable time. A tool that halves one line of it is worth paying for; a tool that promises to transform the business is selling you the line that does not exist.
The hour it does not save is the one on site, and the vendors are increasingly honest about that. Diagnosis from a photograph is improving and still needs the person who can smell gas and hear a pump. The physical trade is close to the last thing in commerce that a model cannot touch, which is a reasonable thing to be in.
What this does not fix
It does not fix pricing. A system that writes quotes faster will write underpriced quotes faster, and most one van businesses that feel busy and broke have a pricing problem rather than an admin one.
It does not fix the record keeping that already was not happening. Software will categorise the receipts you photograph. It will not find the ones in the footwell, and the quarterly obligation above is indifferent to how the data failed to arrive.
And it does not move the liability. The certificate carries a name, the quote carries a business, and the customer standing in a kitchen with a complaint is not interested in which tool drafted the paragraph. Every control in this piece reduces to the same instruction: let the machine do the writing, and keep the deciding.