Can AI Write Toolbox Talks? What a Competent Person Must Check

Yes, AI can write construction toolbox talks: a weekly topic plan, a five-minute script and a sign-in sheet. It cannot decide whether the talk is right for your site. A competent person must read each one against the work being done, correct anything wrong or vague, and sign it off before the foreman delivers it.

The reason this works at all is that a toolbox talk is short and repetitive in form, which is what AI drafts well. The reason it needs a check is that a short talk has no room for error: one wrong instruction about a harness, a trench edge or a lifting sling, delivered confidently to twelve people, is worse than no talk. AI output is a draft, not compliance, and the person who signs is the one answerable for it.

Follow me on Instagram@sagnikteaches

What a good toolbox talk looks like

A toolbox talk is a short briefing given to the crew, usually standing together at the start of a shift or after a break. It covers one subject, takes about five minutes, ends with a question or two, and is recorded. The aim is a conversation, not a lecture. The best ones are tied to what the crew is doing that day: a talk on mobile elevated platforms on the morning the platform arrives is worth ten generic talks on a noticeboard.

Connect on LinkedInSagnik Bhattacharya

That shape suits an assistant. You can ask for a script with four to six points, a plain "what to watch for" list, two questions for the group and a one-line summary to read at the end. It does not suit an assistant to guess which hazard matters most this week, because it cannot see your site.

Subscribe on YouTube@sagnikteaches

Three jobs for AI, and who owns each

Notice that the person owns the facts and the record each time. The assistant owns the wording and the layout. Keeping that split clear is what makes the process safe.

Step 1: Build a 12-week plan from your actual work

Start by listing what is happening, not what is generically dangerous. Give the assistant the trades on site, the work phase, the weather season and any plant arriving. An illustrative prompt for a small contractor building a two-storey extension and refurbishing the existing house:

I run site safety briefings for a small building contractor.
Next 12 weeks on site: groundworks and foundations (weeks 1-2), brickwork and scaffold
(weeks 3-6), roof structure and tiling (weeks 5-8), first fix electrics and plumbing
(weeks 8-10), plastering and decorating (weeks 10-12). Season: late autumn into winter.
Crew: 8 of our own staff plus 2-4 subcontractors on any day.
Give me a weekly topic plan: one toolbox talk topic per week, a one-line reason why it
fits that week's work, and a "trigger" that would make me swap it for something else
(for example a near miss or a new delivery). Do not include topics that need a specialist
to deliver.

A reasonable answer, shown here as an illustration, lists topics such as trench edges and excavation checks in week 1, scaffold inspection tags in week 3, working near roof edges in week 5, cold weather and slippery surfaces in week 7, dust and cutting in week 9. The point of the "trigger" column is that the plan is a starting position. If a delivery of heavy glazing units is due in week 6, you swap in a manual handling talk. The plan stays useful because a person edits it.

Step 2: Write the five-minute script

For each week, give the assistant the topic, the actual task and the specific equipment. Ask for plain speech, short sentences and a "stop and ask" rule. Then read the output as if you were the foreman standing in front of the crew.

Write a five-minute toolbox talk for a building site crew, for a foreman to read aloud.
Topic: working near an open excavation edge on week 1.
Context: 1.2 m deep foundation trench, strip foundations for an extension, 4 operatives,
a mini excavator working at the far end, clay soil, rain forecast.
Format: a one-sentence opening that says why we are talking about this today; five
points in plain words; two questions to ask the group; a one-line stop rule.
Do not quote regulation numbers or specific measurements except the ones I gave you.
Mark with [CHECK] anything you are unsure about.

The result is a fluent, readable script. Here is an illustrative extract and what a reviewer would do with it:

The fourth row is the example that matters. The assistant produced a precise-sounding figure you did not give it. That is the typical hallucination in safety material: a number or standard that looks authoritative and was never checked. Telling the model not to add figures reduces the habit but does not remove it, so the reviewer reads for numbers every time.

Step 3: Create a sign-in sheet that records more than names

Ask the assistant for a table you can print or use on a tablet. A useful sheet has the topic, date, time, presenter, the three or four points covered, any question the crew raised, and a column for each person's name and signature. Add a line at the bottom for "actions arising" such as "install a second barrier by the gate".

Create a toolbox talk sign-in sheet as a table:
Header fields: site name [to fill in], date, time, topic, presenter.
Body: a numbered list of the points covered with a tick box for each.
Attendee table: name, company, signature, time of arrival.
Footer: questions raised by the crew; actions arising (owner and due date);
presenter's declaration; supervisor's review signature and date.

Keep the signed originals. Scan them the same day and file them with the daily site record. The daily report tutorial shows how to log "toolbox talk given" alongside weather, labour and deliveries so the entries sit together and are dated.

Who reviews and signs each week's talk

For a toolbox talk, the competent person is whoever knows the task, the site and the hazards well enough to spot a wrong or missing statement: a site manager, a supervisor, a safety adviser, sometimes a specialist subcontractor for their own trade. What your contract or regulator requires differs, so confirm it there. An AI assistant cannot fill the role, and a vendor's claim that its safety tool produces "compliant talks" does not move responsibility off the person who signs.

A sensible split on a small site is that the foreman delivers the talk, the site manager reviews each week's script on Friday for the following week, and the owner or a safety adviser reviews the full 12-week plan once a month. The same person who reviews should initial and date each script. If you want a wider picture of the paperwork around this, the RAMS sign-off tutorial shows the same author, reviewer and operative roles applied to risk assessments.

A review checklist you can tick in two minutes

Print one per talk. An illustrative filled version for the trench talk above:

Where AI-written talks go wrong on site

Four patterns recur, and each is worth describing so you recognise it.

The generic talk. A talk on "slips, trips and falls" read word for word on a site with fresh mud, no hardstanding and a scaffold lift is accurate and useless. The cure is to feed the assistant the site's current conditions every week, so the draft names them.

The confident wrong detail. An illustrative case: a draft on manual handling states a maximum weight for a single lift. It sounds sensible and comes from nowhere in your own rules. A supervisor who spots a number they cannot trace asks where it came from, and deletes it. Treat every figure as guilty until traced.

The talk nobody listens to. If every talk is the same shape, with the same opening line, the crew stops hearing it. Ask the assistant for variety: a story prompt, a demonstration, a "spot the hazard" photo. The best variation comes from the crew's own near misses, described without names.

The signatures with no talk. The sheet gets signed in the cabin while the talk is skipped. No tool fixes that. A foreman who gives the talk in person, asks a question and notes an answer is the control. The sign-in sheet records it afterwards.

Using a near miss as the topic

The most effective talks follow something that actually happened. Write a three-line factual note without names: what happened, where, what could have happened. Give it to the assistant and ask for a talk built around it.

Near miss (no names): a delivery driver reversed a lorry towards the site gate while two
operatives were unloading plasterboard 3 m behind it. Nobody was hurt. The banksman was
on a phone call.
Write a five-minute toolbox talk that uses this as the example, with three questions for
the group about vehicle movements, and one action we could take at the gate. Keep the tone
factual and do not blame anyone.

The reviewer's job here is different: read the draft for tone and blame, and check the proposed gate action is something the site can do. A talk that sounds like an accusation will be resented. A talk that ends with "we will do X by Friday, owner: site manager" will be remembered.

Seasonal and trade-specific talks

Give the assistant the date and what the weather is doing, and it will suggest talks that match: heat and hydration in a hot spell, cold and ice in winter, wind and loose materials on a roof. Give it the trade and it produces a trade-specific list: for electricians, safe isolation briefings; for plasterers, mixing station housekeeping and dust; for roofers, edge protection and fragile surfaces. Always add a line telling the assistant not to draft talks that need a specialist trainer, such as asbestos awareness or confined space entry. Those belong to a qualified person, not a five-minute chat assistant.

If your firm also quotes and plans work with AI, the scheduling tutorial for small builders pairs well: tie each week's talk to the programme so the briefing appears the day the hazard arrives.

Filing talk records an insurer can follow

A talk record that is not dated is nearly worthless. Use one file name format for all of them, such as TBT_2026-11-06_trench-edges_signed.pdf. Keep three things together per talk: the reviewed script, the signed sheet and any photos of a demonstration. Store them in a folder per site and per month. If the AI draft is kept, mark it "draft" and keep it with the reviewed version so the changes you made are visible. An inspector or insurer asking "what did you tell your crew about trench edges and when?" can then be answered in a minute.

Keep personal data out of what you paste into the assistant. Describe incidents without names, and add attendees only on the paper sheet. A personal ChatGPT, Claude or Gemini account has a model-training switch in its privacy settings that is worth turning off; business plans start with training off for your content, according to the vendors, though their terms are worth re-reading once a year. Our privacy terms guide lists what to look for.

Rewriting the opening line the crew hears first

The first sentence decides whether the crew listens. AI openings tend to be soft: "Today we are going to talk about the importance of working safely near vehicles." Nobody moves on that. An illustrative before and after for the gate talk:

Before: "Vehicle movements on construction sites present a significant risk and it is important that everyone remains vigilant."

After: "Last Tuesday a lorry reversed at the gate while two of us were unloading board behind it. Nobody was hurt. Five minutes on what we change at the gate, starting today."

The second opening is a fact, a date and a promise about time. You will usually write it yourself after the assistant has given you the body, and that is a good division of labour: the model builds the frame, you supply the detail that only someone on site knows.

Questions that show the crew understood

A closing question does more than a signature. Ask the assistant for questions that need an answer about this site, not yes or no. Illustrative examples for three topics:

• Trench edges: "Where is the nearest safe place to put the spoil today, and who decides?"
• Scaffold: "What does the tag on that lift tell you, and what would make you refuse to climb it?"
• Cutting dust: "Which of today's tasks makes dust, and what are you wearing or using for it?"

Note the answers on the sheet in a few words: "[first name]: tag out of date, would stop". A sheet with real answers on it shows the talk was a conversation. If the crew cannot answer, the talk did not land and you repeat it differently tomorrow, which is useful information the signature alone would hide.

A different site: a small fit-out firm

The same method works away from groundworks. Take an illustrative six-person fit-out firm refurbishing a dental practice at weekends. The week's work is stripping a suspended ceiling, running new data cable and installing a small plant cupboard. The manager gives the assistant those three facts plus "building is a clinic with sensitive equipment, hours are Saturday and Sunday only, a cleaner is in on Saturday morning". The assistant proposes talks on dust control near clinical areas, working at height from a tower in a small room and keeping escape routes clear. The manager swaps the second for a talk on the tower the firm will actually use, adds the cleaner's hours to the first and reviews all three on Thursday evening. The AI saved the layout; the manager supplied the building.

Twelve minutes a talk instead of thirty-five

An illustrative small contractor writes a talk by hand in about 35 minutes (finding the topic, writing it up, formatting a sheet). With a saved prompt, the draft takes two minutes and the review about ten, so roughly twelve minutes per talk instead of thirty-five. Over 48 working weeks, the 23 minutes saved each time add up to about eighteen hours a year per site manager, a little over two working days. The gain is modest in hours, but the real value is consistency: a talk is ready every Friday for the next week rather than written at 6.45 on Monday morning.

The Friday-to-Monday cycle for a site manager

1. Friday morning: open your saved Project or prompt, add this week's site conditions and next week's work, and generate next week's talk (five minutes).
2. Read it aloud once. Cut anything you would not say. Trace or remove every figure (ten minutes).
3. Sign and date the review box. Print the script and the sign-in sheet.
4. Monday start: foreman gives the talk, asks the two questions and notes the answers.
5. Collect signatures, scan the sheet, file it with the daily report.
6. Monthly: look back over the sheets for repeated actions or topics the crew keeps asking about, and feed those into next month's plan.

Run this for four weeks and you will have a small library of reviewed talks you can adapt rather than rewrite. For a similar approach in other settings, see the tutorials on farm safety talks and trucking safety talks, where the hazards differ but the author, reviewer and record roles stay the same.

Toolbox talk questions from site managers

How long should an AI-written toolbox talk be?

Aim for five minutes of speaking, which is roughly 500 to 650 words read aloud, or six to eight short points. Anything longer loses the crew. Ask the assistant for a script plus three questions to ask the group, then cut anything the foreman would not actually say.

Does the sign-in sheet prove the talk was understood?

No. It proves who was present. To show understanding, add a question or two to the talk and note the answers or any follow-up. A quick demonstration or a tick against each point covered is stronger evidence than a column of signatures.

Can AI pick the weekly topic for me?

It can suggest a plan based on the work you list: the trades on site, the season and recent near misses. You choose what goes in the calendar. A topic prompted by something that actually happened on your site this week usually lands better than one picked from a list.

Is it safe to put near-miss details into ChatGPT?

Remove names, the client and the exact location first. Describe it as 'an operative reversed a telehandler near the hoarding' rather than naming the person. Business plans do not train on your content by default, but check current vendor terms and keep personal details out regardless.

Who counts as a competent person for a toolbox talk?

Someone who knows the task and the site well enough to spot an error in the talk, usually a supervisor, site manager or safety adviser. Your contract and regulator define the exact requirement, so confirm it there rather than relying on an AI assistant's answer.

Further reads

• How to Write a Method Statement for a Small Build, With AI Help — Drafting method statements that a supervisor can then verify.
• Best AI App for a Construction Company: Logging Near Misses — Log near misses so next week's talk has a real topic.
• How to Draft Risk Assessments With AI and What a Human Must Check — Draft risk assessments with AI and see what a human must check.
• How to Write a Health and Safety Policy With AI and What to Check — The policy that your toolbox talks sit underneath.
• Roofing RAMS Template: AI Drafts, a Competent Person Signs It Off — Roofing RAMS drafted by AI and signed by a competent person.
• Five AI Apps for Construction Site Work: Photos, Logs, Checks — Apps for site photos, logs and checks that store talk records.
• Construction Meeting Minutes: Record, Summarise and Chase Actions With AI — A site meeting minutes template and the AI steps behind it: recording consent, transcript to action table, same-day issue and weekly action chasers.
• Job Hazard Analysis Template: Draft One per Job Type With AI — A worked job hazard analysis for one job type: the AI's first draft, what the supervisor changes, and a table you can reuse for each job type you do.
• Lone Working Policy for Trades Firms: Check-In Rules Drafted With AI — A two-page lone working policy for a small trades firm: tasks needing two people, check-in rules, an escalation ladder and an AI drafting prompt.
• Scaffolding Inspection Records: A Digital Form and AI Summaries — A fourteen-field inspection form, photo rules, an overdue register and a weekly AI summary that never certifies a scaffold.
• Simpro Form Builder: Draft Job Sheets and Safety Checklists With ChatGPT — Form Builder lays out Simpro's job documents and Digital Forms holds site checklists: a one-fact-per-question method with ChatGPT drafts and a sign-off block.
• Tree Surgeon Risk Assessment Template: AI Drafts, You Sign Off — A tree work risk assessment checklist: what AI can draft from site notes, the hazards it misses, and the steps a competent person takes before signing.
• Translation App for Construction Workers: Brief Mixed-Language Crews — How to brief a mixed-language crew: what Google, Apple, DeepL and Jobber each do, a 20-phrase test with a bilingual lead and a checked daily routine.
• Can AI Design a Scaffold? What Software Can and Cannot Do — Why a chatbot cannot design a scaffold, what scaffold design software does, how TG20-style compliance sheets fit, and where AI helps with paperwork.
• Can AI Draft a Construction Phase Plan? What You Must Still Check — A worked CDM 2015 example showing the duties AI drafts leave out, and what the principal contractor must confirm before work starts.
• How Is AI Used in Construction? Site Jobs and Office Jobs Compared — A small firm's site-versus-office split for AI in construction, with a worked week and a rule for which jobs to try first.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general toolbox talk and site briefing practice; vendor help pages for ChatGPT and Gemini data controls. AI output is a draft and is not compliance advice.