Farm Safety Plan: Draft Risk Assessments and Toolbox Talks With AI

Use AI to draft three documents: a hazard list for each enterprise on the farm, a risk assessment table that scores each hazard and names a control, and short toolbox talks for each seasonal job. Then walk the farm with those drafts and have a competent person correct and sign them. AI writes the paperwork; people judge the risk.

The order matters because the walk-round is where farm risk lives: the gate that does not latch, the slurry store with no fence, the loader that has no seat belt. A model has never seen your yard. What it does well is turn your rough notes into a tidy table in minutes, remember hazards you forgot (lone working, fatigue at harvest, children visiting a farm shop) and write a ten-minute talk in plain words. What it cannot do is notice what is actually wrong with your farm.

Follow me on Instagram@sagnikteaches

What the plan is made of and who does each part

A workable farm safety plan has five parts. Splitting them by who does the work shows where AI helps and where it must stay out of the way.

Connect on LinkedInSagnik Bhattacharya

Every row has a person doing the part that involves being physically on the farm. If a draft contains something nobody has looked at, it is a guess dressed as a finding.

Subscribe on YouTube@sagnikteaches

Gather your farm notes before you open the chatbot

The quality of the draft depends on the quality of the notes. Spend twenty minutes writing a plain list in a document, with no polishing, and keep names out of it. Cover:

• Enterprises: what you actually run (for example 90 breeding ewes, 140 acres of cereals, a small workshop, a roadside farm shop).
• Machinery and vehicles: tractors with loaders, a telehandler, an ATV, a mower, a baler, a PTO-driven mixer wagon. Note ages, because older machines often lack guards and cab protection.
• Buildings and structures: grain store, slurry or silage pits, fuel tank, workshop, roofs with fragile sheeting, ladders and platforms.
• People: number of full-time staff, seasonal and casual workers, family members, contractors and visitors, and whether anyone works alone or is under 18.
• Past near-misses: described without names ("a bull pushed through a gate in the handling pen").
• Existing controls: guards, training, rules, personal protective equipment, written routines.

The near-miss list is the most valuable item and the one farms most often leave out. A near-miss tells you where the real risk sits better than any general list can.

Step 1: Build a hazard list for each enterprise

Start with one enterprise at a time, because a single prompt that asks for "all farm hazards" returns the same generic list every farm gets. Give the model your notes for that enterprise and ask it to separate hazards from controls, since models like to mix them.

You are helping a farm manager prepare for a safety walk-round.
Farm enterprise: 90 breeding ewes, lambing indoors in February-March,
handling pen with a crush, ATV for feeding, two casual workers each lambing.
Notes (no names): one near-miss with a ewe pushing through a gate in the pen;
one worker works alone at night checks; ATV has no roll protection.
Task: list the hazards for this enterprise only, as a table with columns:
Hazard | Who could be harmed | How | Controls already in my notes.
Do not invent controls I have not mentioned. After the table, list five
hazards that farms like this often overlook, and mark each "to check on site".

Illustrative output, trimmed to three rows:

The "to check on site" list is the useful bit. In this case it might raise zoonotic infection from birthing fluids (handwashing, gloves, and pregnant workers kept away from lambing ewes, newborn lambs and afterbirth), fatigue in a lambing fortnight, and the handling of lambing ropes and sharp tools. Read each one, cross out the ones that do not apply, and keep a note of what you add by hand. When you repeat this for arable, machinery and the workshop, you will end up with a hazard list of 40 to 60 items for a mixed farm, which is normal.

Hazards that deserve a second look on every farm

Whatever the enterprise, ask the model to sort your list against the situations behind most fatal farm accidents. One national regulator's five-year figures put being struck by a moving vehicle at the top, followed by injuries from animals, falls from height, being struck by objects such as bales and gates, and contact with machinery. Add the rarer killers too: drowning or gassing in slurry, silage or water, entry into grain stores and silos, electricity (overhead lines and the workshop), and chemicals. Presence of children and visitors matters too if you have a shop, a campsite or a visitor trail. Ask the model "which of these have I not covered?" and read the answer critically. Do not accept a hazard simply because it appears; accept it because you can picture it on your farm.

Step 2: Turn hazards into a scored risk assessment table

A risk assessment answers four questions for each hazard: who might be hurt, how likely is it, how bad would it be, and what will you do about it. The scoring bit is the part owners like least, so it is the part AI helps most. Use a simple five-by-five matrix: likelihood from 1 (rare) to 5 (almost certain) multiplied by severity from 1 (minor) to 5 (fatal or life-changing).

Ask the model to apply the matrix twice for each hazard: once with the controls you currently have (the existing risk) and once with the controls you plan to add (the residual risk). Showing both makes it clear what the action is for.

Using the hazard table above, produce a risk assessment table with columns:
Hazard | Who | Likelihood (1-5) | Severity (1-5) | Score | Existing controls |
Further action | Owner (role only) | Due | Residual score.
Score = likelihood x severity. Treat 15 and above as "act now",
8 to 14 as "plan within a month", under 8 as "monitor".
Write controls as specific actions I can check ("fit a latch that
closes under its own weight"), not generalities ("improve safety").
Where you are guessing a score, put a question mark after it.

Illustrative output for the first two hazards:

The question marks are what you asked for, and they are where your judgement goes. Someone who has seen a night-time injury knows whether a severity of 5 is realistic for the farm; the model does not. Change the scores where your experience disagrees, and write the reason in the "further action" column so the next reviewer sees it.

What a good control looks like

Models tend to write controls such as "ensure staff are trained" or "maintain equipment". Neither can be checked. Push the draft towards the order that works in practice: remove the hazard if you can, then guard it or separate people from it, then change how the job is done, and only then rely on training and protective equipment. Ask the model to rewrite any control that begins with "ensure", "be aware" or "take care" into something with a person, an object and a date.

A before-and-after: the same hazard, written two ways

Here is a real-sounding difference between what the AI gives you by default and what you should end up with, for a PTO shaft on a mixer wagon.

First draft from the model: "Risk of entanglement with PTO shaft. Controls: ensure guards are in place and operators are trained and aware of the risks."

After the walk-round: "Entanglement with PTO shaft on the mixer wagon (purchased second-hand). The shaft guard has a split cone and a missing chain. Do not run until the guard and chain are replaced. Replacement ordered by the farm manager, due Friday. Operators: two named staff, both given a hands-on demonstration of stopping the engine and removing the key before clearing blockages. Re-check the guard at the start of each feeding round."

The second version names the machine, the fault, an owner and a date. That is what an inspector, an insurer or a new manager can act on. The first could be pasted into any farm's file without anybody walking anywhere.

Step 3: Draft toolbox talks that last ten minutes

A toolbox talk is a short briefing on one job, given at the point of work. The format that works on a farm is five parts: the job, the main danger, the three rules, a question to the group, and a sign-in. Keep it to one A4 page so it can be read aloud.

Write a ten-minute toolbox talk for casual workers starting on the first
morning of lambing. Audience: new casuals, some with no farm experience.
Reading age: plain English, short sentences, no jargon.
Structure: 1) what we will be doing, 2) the main danger, 3) three rules
(each one sentence), 4) a "what would you do?" question with an answer I can
read out, 5) a sign-in table: name, date, signature.
Topics to cover: handwashing after handling lambs and afterbirth, pregnant
workers stay away from lambing ewes, newborn lambs and afterbirth and tell
the manager so their jobs can change, never work in a pen alone with a ram,
report any injury straight away.

The model will return a script of about 300 words. Read it aloud once. Anything that sounds stilted or too clever, cut. Anything you would not say at 6 a.m. in a shed, rewrite. Then put a name on the author line, and change one detail to something only your farm has ("the pen with the broken latch" is better than "any gate").

A short list of talks to prepare before each season

• Spring: lambing and calving (zoonoses, handling, lone working), new casuals induction, ATV use.
• Summer: tractor and trailer movements, working near overhead lines, heat and hydration, children on the farm.
• Autumn: harvest fatigue, grain store entry (never enter alone, lock out augers first, never walk on grain to make it flow), PTO guards, fire in the combine.
• Winter: slurry gases, icy yards, working in the dark, workshop hot work and electrics.

One talk per row of this list takes the model about a minute and you about fifteen minutes to improve. A year of eight to ten talks is a realistic target for a farm with a handful of staff.

Checking the draft against what the model got wrong

A mixed farm in this exercise tends to throw up the same kinds of mistakes. These are worth scanning for in every draft.

• Invented standards. A model may write that a guard must meet a specific numbered standard, or that a legal limit is a particular figure. Strike any number or standard you cannot find on the regulator's own page or the machine's handbook.
• Wrong-sized controls. A dairy parlour control copied into a sheep shed. Read each control and ask "do I have that here?"
• Missing hazards you would not think of. The model sometimes covers these well. Keep them in, and decide on site.
• Rosy scores. Likelihoods tend to run low because the model has nothing to compare against. If you have had a near-miss, raise the score.
• Overlong documents. A 60-page file nobody opens is worse than ten pages that people read. Ask for a summary page of the five highest scores.

A competent person signs it, and your local rules are checked

Workplace safety law differs between places, and the details change: who must have a written assessment, how many employees trigger a written policy, what training records you must keep, and which activities (such as young people on machinery) are restricted. Name your local workplace safety regulator in your own notes and check its farming guidance, which is usually free. Take one instruction from it into your plan for each enterprise, and cite the document name in the file. If your insurer has a farm risk checklist, use that as another cross-check.

The "competent person" is the human sign-off. On a small farm that might be the owner who has had training and many years of experience, an adviser from a farming body, or a consultant brought in once a year. What matters is that they have walked the farm with the table in their hand and can say, truthfully, "I checked this on site." Their name and the date go at the bottom of the assessment. If the person is also the one who prompted the AI, that is fine, as long as they treated the draft as a list of things to verify.

If you would rather map this with someone, that is what my AI implementation consultation covers: finding which of your repeated safety jobs AI can draft reliably, and setting up a routine your manager can run.

A walk-round worked through with numbers

Say a farm of 180 acres with 100 ewes, 40 suckler cows, three full-time staff and four casuals across the year. The owner wants a refreshed safety file before lambing. Illustrative timings:

1. Notes (30 minutes). Owner writes the enterprise list, machinery, near-misses and existing controls.
2. Hazard lists (25 minutes). Four prompts, one per enterprise (sheep, cattle, arable and workshop). Output: 52 hazards. The owner deletes 9 that do not apply and adds 4.
3. Risk table (35 minutes). The model scores 47 hazards. The owner changes 11 scores and marks 6 as "act now" (score 15 or above).
4. Walk-round (2 hours). Owner and farm manager check each "act now" item and a sample of the rest, with the table printed. They find two controls that were on paper but not on the farm (a fire extinguisher past its service date, no first-aid kit in the lambing shed).
5. Toolbox talks (1 hour). Four talks drafted and trimmed: lambing, ATV, new casual induction, slurry and silage.
6. Sign-off (30 minutes). The adviser reviews the file, queries two scores and signs.

Total owner time is about five hours spread over a week, with the model's part under thirty minutes. Typed from scratch, the same file could easily take a couple of evenings, which is one reason safety files stay unfinished. The saving is in the typing and structuring; the five hours that matter are the walk-round and the talks.

Keeping the plan alive after the first draft

A safety file that is signed and filed is only half-done. Put four triggers into a calendar entry and treat each as a reason to reopen the table:

• After any injury or near-miss (add it to the hazard list, rescore, and brief the team).
• Before any new machine, enterprise or building change.
• When a new seasonal worker joins, so they get the induction talk the same day.
• Once a year, ideally just before your busiest season.

A short prompt makes the review lighter: paste the current table and a line of notes ("bought a second-hand telehandler; moved the calf pens; one near-miss at the grain store") and ask the model to list rows that probably need rescoring and any new hazards the change introduces. Check what it returns the same way as before. If you are building a wider set of farm routines, a chatbot that answers from your own farm notes and SOPs can serve the finished plan back to staff in plain language, and rota and timesheet automation helps keep long hours and lone working visible.

Keeping staff and incident details out of the prompt

The prompts above work without any personal information, and that is deliberate. Refer to roles ("shepherd", "casual worker"), not names. Describe incidents by type, not by person. Leave out medical information completely. If you use a free or consumer plan, switch off the model-training setting (in ChatGPT it sits under Data controls; in Claude it is in the privacy settings); business plans of ChatGPT and Claude do not train on your content by default, according to the vendors' own pages in November 2026. For wider guidance on what is safe to paste, the risks of using AI in a small business covers the basics, and drafting risk assessments with AI gives the general method this tutorial applies to a farm.

When the paperwork does not match the farm

A common failure goes like this. The owner produces a polished file in an afternoon, signs it, files it, and a casual worker is hurt doing something the plan does not mention, because the hazard list was built from a generic prompt rather than from the farm. The fix is not more AI. It is a short test for every draft: choose three rows at random, go to the place, and check the hazard, the control and the owner exist. If any of the three fails, the whole table needs another pass on site.

A second failure is subtler: the plan is accurate, but nobody reads it. The cure is the toolbox talk. A plan lives where it is spoken. If your talks use the farm's own words and point at the farm's own gates, people remember them; if they sound like a leaflet, they do not.

A one-page farm safety checklist to print

Use this as the front page of the file. Each line is a yes or no that you can verify.

• Hazard list exists for every enterprise, and was checked on site (date: ____).
• Risk table scored; every score of 15 or above has an owner and a date.
• Machinery guards inspected this season; faults logged.
• Emergency sheet with access points, map coordinates and first-aid locations is on the wall of each building.
• Induction talk given to every new worker on day one, with a sign-in.
• Lone-working rule written and understood (check-in times, who calls whom).
• Competent person has signed and dated the table.
• Local regulator's farm guidance checked for changes since last review.
• Next review date set.

AI can lay this checklist out, create the sign-in sheets and draft the emails reminding people of due dates. The ticks on it are yours.

Questions farmers ask before letting AI near the safety file

Can AI write a farm risk assessment that satisfies an inspector?

AI can produce a sound first draft, but an inspector looks for evidence that someone competent walked the farm, judged the real risks and acted on them. A draft nobody has checked against your yards, machines and people is not that evidence. Treat the AI output as paperwork prepared for your walk-round, never as the walk-round itself.

Is it safe to paste staff names or accident details into ChatGPT or Claude?

Leave names, medical details and injury descriptions out. Describe an incident as 'a casual worker was struck by a reversing trailer' and nothing more. If you want to discuss real cases, use a business plan where your content is not used for training by default, and still keep identifiable health information out.

How often should toolbox talks happen on a small farm?

Match them to the work, not the calendar. A short talk before each new seasonal job (silage, lambing, harvest, a new machine, a new casual starting) does more than a monthly slot nobody attends. Many farms settle on ten minutes at the start of the first morning of each job.

Who counts as a competent person to sign off the plan?

Someone with the necessary skills, knowledge and experience of farm work to spot the hazards and judge the controls, such as an experienced farm manager, a safety adviser or a trainer from an agricultural body. Your local workplace safety regulator or insurer can tell you what they expect for a farm of your size.

Further reads

• AI Hallucinations Explained for Business Owners: Causes and Fixes — Why models invent standards and figures, and how to catch them.
• How to Write a Health and Safety Policy With AI and What to Check — Pair the plan with a written policy statement.
• Can Farmers Use AI for Farm Records and Paperwork? — Where else on a farm AI can handle the paperwork.
• AI Assistant for Farmers: ChatGPT, Claude or a Farm App? — Choosing between ChatGPT, Claude and a farm app.
• How to Build a Staff Training Matrix With AI — Track who has had which training and when it lapses.
• A Simple AI Risk Register for Small Businesses (With Template) — Keep AI's own risks in a register as well.
• Can AI Write Toolbox Talks? What a Competent Person Must Check — AI can plan a year of toolbox talks, script each five minutes and format the sign-in sheet. A competent person still reviews and signs every one.
• Forest School Risk Assessment: Site Checks and Session Plans With AI — How to split forest school risk into a site assessment and per-session checks, with weather triggers, ratios and an AI drafting prompt.
• Fish Farm Record Keeping: Feed, Water and Mortality Logs With AI — A daily feed, water-quality and mortality log for a small fish farm, with voice capture, spreadsheet flags and a clear line on what AI must not decide.
• Selling Beef Boxes Direct: Hanging Weight, Cut Sheets and Pricing With AI — Convert hanging weight to retail cuts, price a quarter, half or box, and build the order form and deposit steps with AI checking your sums.
• Selling Produce to Restaurants: Availability Lists and Chef Emails With AI — A weekly availability list template and four chef email scripts for growers, with AI turning your harvest notes into the list.
• Trucking Company Safety Plan: AI-Drafted Policies and Toolbox Talks — Draft a policy outline, a year of ten-minute toolbox talks and an incident review form with AI, then have a person check the rules.
• AI Chatbot for Farming: What It Can Answer and Where It Fails — What a chatbot can reliably do on a farm, with sample prompts, and the agronomy, livestock and safety questions it should never settle.
• Do Farmers Use AI? Eight Uses of Artificial Intelligence, With Costs — How many farmers use AI, and eight small-farm uses with indicative costs, example prompts and the checks that keep each one honest.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general workplace-safety method (hazard, who is harmed, controls, residual risk, review); regulator guidance to be checked locally by the reader; vendor privacy pages for ChatGPT and Claude plans.