What AI Cannot Do for a Trades Business: An Honest List

What AI cannot do for a trades business is anything where being confidently wrong costs you money: it cannot price a job, it cannot know what is in your van, it cannot decide what the regulations require, and it cannot carry the blame when a wrong number reaches a customer. What it does well is language and transcription. It turns your scribble into sentences, your voice note into text, your bullet list into a scope a customer can read. So the honest split for a 1-20 person firm in 2026 is this: let AI touch words, never let it touch figures, and put one review gate in front of anything a customer or an authority will read. Below is the specific list — five admin tasks AI genuinely removes, five it makes worse, and the review rule.

What AI cannot do for a trades business: the short version

The five admin tasks AI genuinely removes

1. Site notes into a customer-readable scope

You come off a survey with 40 words of shorthand. Two days later the quote is still not sent, because writing it up properly is a different job to doing it. Paste the shorthand in and ask for one thing: a scope list and an exclusions list in plain language. You get a draft in seconds, then you type every figure yourself. Realistic saving is 15-25 minutes per quote, and the bigger gain is the quote that goes out tonight instead of Thursday.

2. Voice notes into text

Speech-to-text is the one part of this that is finished. Talk for 90 seconds at the board, in the loft, in the van, and get a written job record. It handles trade vocabulary less well than plain speech, so expect to correct a few product names. It is still faster than typing with one glove off. If the transcription happens inside the job card rather than a separate app, you skip the copy-paste step — which is where most people quietly give up.

3. Chasing emails and awkward replies

The unpaid-invoice email you have not sent because you cannot find the tone. The reply to a customer who is being unreasonable. Give the AI the facts and the tone you want (firm, not rude) and it writes it. Read it, cut a third, send. This one is worth real cash: invoices get chased on day 31 instead of day 60, because writing the email stopped being a task you avoid.

4. Tidying inspection and handover write-ups

You did the checks and ticked the boxes. What you have not done is write the summary the customer or the main contractor wants to read. AI turns your ticks and your two-line comment into a paragraph.

One hard limit: it may only rephrase what you recorded. If it adds a finding you did not make, that is a fabricated inspection record. In Sweden the self-inspection document (egenkontroll) is the thing you personally stand behind; a UK or NZ reader can think of a certificate carrying their signature. The AI writes the sentence. You make the finding.

5. Job ads, intake forms and website words

Low stakes, high avoidance. AI writes a decent job advert or a clear enquiry form in a minute. The failure mode is bland rather than dangerous, and you can spot bland.

The five jobs AI makes worse

1. Pricing

Ask an AI what to charge for a consumer unit change and it will answer. That answer is a blend of scraped forum posts, the wrong country and two years ago. It does not know your hourly rate, your overhead recovery, your material margin, or that this board sits on a first floor with no loft hatch. Your price comes from your own price list and your own recorded hours. If a figure in a draft did not come from a list you maintain, delete it.

2. Reading numbers off documents

Pulling totals off a photographed supplier invoice feels like the perfect AI job. It is the most dangerous one, because the failure is silent. A human misreading 1 380 as 1 830 usually senses something is off. A model does not flag doubt — it returns 1 830 with exactly the same confidence as a correct read.

Photograph the invoice and store it against the job, by all means. Do not let a figure enter your accounts without an eye on the original. Errors here do not stay small: they compound into a VAT return that will not reconcile.

3. Tax rules and regulations

This is where AI is at its worst and sounds at its best. Ask about ROT — the Swedish tax deduction for household labour — and a model hands you a percentage, a ceiling and a deadline. The percentage is stable: 30% of the labour cost. The annual per-person ceiling is shared with RUT and it changes. A model trained last year states the old figure as current fact, and you pass it to a customer as a promise, in writing, in a quote.

Check current figures at Skatteverket. The same goes for Elsäkerhetsverket, Boverket and Arbetsmiljöverket. Outside Sweden the pattern is identical: a model answers a Part P or Building Regs question in a confident British accent and is a revision behind. Regulations have a source. Read the source.

4. Scheduling and dispatch

An AI that reshuffles your week does not know that the customer is only in after 3pm, that the motorway is shut, or that this job needs the big ladder that is on the other van. A drag-and-drop planner where you make the call in eight seconds beats an optimiser you have to argue with. In a 1-20 person firm the bottleneck was never the arranging. It was knowing where everyone actually is.

5. Anything that sends itself

Wire AI output straight to an outbound action — auto-reply, auto-quote, auto-invoice — and you have deleted the only control in the system. Every safe use above works because a human reads the draft first. Automate the drafting, which is the slow part. Never automate the sending, which takes two seconds anyway.

The review rule that keeps you out of trouble

Here is the whole rule, short enough for the wall.

No figure reaches a customer unless a human has just compared it to the thing it came from.

Not "checked it". Not "it looked about right". Compared it to the source. In practice that is three checks and about 90 seconds:

• Prices against your price list. Open the list. Look at the line. Not from memory — memory is where last year's rate lives.
• Hours against the record. Time clock, timesheet or your own note. If the quote says six hours and the job took nine, that gap is a conversation to have now, not after the invoice lands.
• Tax and statutory fields against the source document. Deduction amounts, VAT, reference numbers, property designations. These come from the customer's paperwork or the authority's own site. Never from a model, never from a draft.

Then one habit that catches most of the rest: read the draft out loud once. Your ear catches the sentence you would never actually say to a customer, and that is exactly the sentence an AI writes.

When a wrong number does reach a customer

It will happen eventually, AI or not. What matters is the next hour.

1. Ring them. Do not email. A phone call about your own mistake reads as honesty. An email reads as a paper trail being built.
2. Say the wrong number, the right number and the difference. No hedging, no passive voice, no "an error occurred".
3. Send the corrected document the same day, with the correction visible rather than quietly swapped.
4. If they have already signed at the wrong price, that is your contract. In Sweden a signed offert binds you, and the standard consumer terms (Hantverkarformuläret) plus Konsumentverket's guidance point the same way; UK and NZ consumer contract law lands in much the same place. An obvious typo may be negotiable. An optimistic estimate is not. You eat the small ones and renegotiate the big ones face to face.
5. Find the gate that failed. Nearly always it is a figure that came from a draft instead of from a list.

Why a chatbot is riskier than a system

Most of the danger above comes from AI sitting in a separate window, cut off from your real data. A chatbot has to guess your rates because it cannot see them. When the assistant sits inside the system that already holds your price list, your hours and your job history, the guessing stops being necessary: the AI drafts the words and the system supplies the figures. That is how AI assist works in OdinTask — drafting and voice transcription on the included Cloudflare AI, no keys and no second tool.

A sensible way to test this in your own firm

Two weeks, one rule per week, no opinions and just counting. Week 1: every quote gets an AI-drafted scope and you type every figure — count how many quotes go out same-day compared to the fortnight before. Week 2: every job gets a voice note instead of a written note — count how many job records are actually complete on Friday.

If neither number moves, AI is not your bottleneck and you have lost two weeks of mild curiosity. If both move, you have found the real edge of what AI can do for a trades business: narrower, and more useful, than anyone selling it will tell you.

If you want the drafting and the discipline in one place, start the free 14-day trial and put your own price list behind it. Costs are on the pricing page.