Driver Walkaround Check: Triage Defect Reports With AI

Put every defect into one of four severity bands, let AI sort the free-text reports into them, then rank the rest by which vehicle is due out, parts in stock and repeat reports. Which defects take a vehicle off the road comes from the DVSA or FMCSA pages and the manufacturer's manual, never from the AI's memory.

That split is the whole design. AI is good at reading forty scruffy driver notes before the workshop opens and putting them in order. It is not an authority on what is legal to drive. Where it is unsure, the rule is simple: when in doubt, the vehicle stays parked until a qualified person looks at it.

Follow me on Instagram@sagnikteaches

The form drivers fill in decides how good the triage is

A walkaround check is the driver's look over the vehicle before it moves. Many operators also have a duty to record defects. The FMCSA rule at 49 CFR 396.11 requires a driver to prepare a written report at the end of each day's work for each vehicle driven, listing defects they find or are told about that would affect safe operation or cause a breakdown. The report must cover at least the service brakes, parking brake, steering, lighting and reflectors, tyres, horn, wipers, mirrors, coupling devices, wheels and rims, and emergency equipment, and the driver signs it. If no defect is found, no report is required under that rule, though most firms keep a "nil defect" tick for evidence. The section also exempts a carrier that runs only one commercial vehicle, and it was amended in 2026, so read the current text rather than an older summary. DVSA's "Guide to maintaining roadworthiness" (listed on gov.uk as "Keeping commercial vehicles safe to drive") covers daily checks and has annexes with example driver's vehicle defect reports for goods vehicles, buses and coaches, and light goods vehicles. Read whichever applies to you before you design anything.

Connect on LinkedInSagnik Bhattacharya

A form that triages well asks for five things, each in its own field:

Subscribe on YouTube@sagnikteaches

1. Vehicle ID: fleet number or registration, chosen from a list (a typed field invites errors).
2. Area: choose one from a list such as brakes, steering, tyres and wheels, lights, mirrors and glass, coupling, body and load, cab and controls, fluids and leaks, other.
3. What is wrong: one sentence in the driver's own words.
4. Can the driver drive it safely? Yes, No, Not sure. This one question is surprisingly informative.
5. Photo: optional but encouraged.

The "area" list matters more than it looks. A structured area lets the AI check for mismatches, such as a note about "air hissing" filed under "body", and it lets the workshop filter by trade (the tyre fitter's list differs from the electrician's).

Severity bands with plain-language actions

You need bands the workshop can act on. Four is enough. The wording below is an illustration of the structure; the content of the top band comes from outside your business.

Band A is not something you invent. For roadside enforcement, DVSA publishes How vehicle defects are categorised in roadside checks and vehicle tests, a separate publication on gov.uk that sets out how defects are classed, and the roadworthiness guide itself explains the operator's duties. On the FMCSA side, 49 CFR 396.9 lets authorised inspectors declare a vehicle out of service when its mechanical condition or loading would likely cause an accident or a breakdown, and the vehicle must then not be operated until the listed repairs are completed. Print the pages that apply to your vehicles and keep them next to the form. Your Band A list is a restatement of those categories in your own workshop's words, checked by a qualified person.

The manufacturer's manual adds items the regulations do not spell out for your model, such as minimum brake pad thickness, warning-light meanings or specific tyre limits. Add those to Band A or B as your workshop lead decides.

What to give the AI, and what never to ask it

Give it three things: your four bands with their definitions, your written list of Band A triggers (typed in from the official pages and the manual, by a person), and the day's reports as a table. Ask it to assign a band to each report, explain the match in one line and quote the trigger it matched.

Never ask it: "Is this defect a prohibition?" or "Is this vehicle legal to drive?" Those questions invite a confident, plausible answer built from memory, and categories and thresholds change. The safe question is narrower: "Does anything in this report match an item in my list below?"

You are sorting driver walkaround defect reports for a small fleet workshop.
Below are (1) our band definitions, (2) our Band A "do not move" list, copied by a
qualified person from the regulator's guidance and the manufacturer's manual,
and (3) today's reports.
For each report return: report number, band (A, B, C, D or REVIEW), the exact item
in our list it matched (or "none"), and a one-line reason.
Rules: if a report might match a Band A item, return A. If you cannot tell, return REVIEW.
Never move a report to a lower band than the driver's own "can't drive safely" answer implies.
Do not state what is legal; only match against our lists.

BAND A LIST:
[paste]

REPORTS:
[paste table: no., vehicle, area, driver note, drive safely? yes/no/not sure]

Notice the three safety rails. Doubt goes up (an uncertain match becomes A, a muddle becomes REVIEW), the driver's own "not safe" answer can only raise a band, and the AI is told not to make legal statements. These rails are the difference between a handy sorter and a liability.

A morning's batch, sorted

Here is an illustrative batch from a fleet of nine vehicles, with the AI's bands and what a workshop lead would do with them. The Band A list in this illustration is a made-up placeholder, so treat the matches as demonstrations of how the sort behaves, not as guidance about real defects.

Report 5 is the important one. On its own it is Band D noise, but it shares a vehicle with report 1, and a "funny noise" on a vehicle with a suspected brake issue deserves the same inspection. Ask the AI to group by vehicle and show the highest band per vehicle first. The workshop lead then works down a list of vehicles, not a list of notes.

Ranking the repair list beyond severity

Severity answers "may it move?" Ranking answers "what do we do first?". Once Band A items are parked and dealt with, the order of B, C and D repairs depends on a few practical factors. Give each a number if you like, but keep the arithmetic visible.

• Job needed tomorrow? Is the vehicle scheduled for a long or critical run?
• Parts on the shelf? A bulb can be done at once; a window needs an order.
• Repeat? A defect reported three times in a fortnight moves up a band, because repeated faults tend to be the real ones.
• Time to fix. Short jobs between longer ones keep the workshop moving.

The assistant can build the list as a table with these columns once you give it the facts: which vehicles are due out, which parts are in stock, how long each job usually takes. Of course it can only use facts you provide, so give it a small parts table and a schedule. A reply (illustrative) might read: "1. V-04 parking-brake check, 90 minutes, vehicle due out 10:00, parked pending inspection. 2. V-07 tyre, 45 minutes, tyre in stock. 3. V-02 indicator, 15 minutes." The workshop lead overrides it when they know better; the point is to have a starting order in two minutes rather than twenty.

Reading photos: helpful, not conclusive

Current chat assistants can look at a photo and describe what they see: a cracked reflector, a worn tyre shoulder, oil on the underside. They cannot measure tread depth to a millimetre from a phone image, judge whether a bulge is in the structure or only the paint, or tell a harmless weep from a serious leak. Treat a photo as extra context for the person reviewing, and use it to catch mismatches: the note says "small chip" but the photo shows a long crack.

Look at this photo from a driver's walkaround report. The driver wrote: "[note]".
Describe only what is visible. Say whether the photo appears to match the note,
list anything visible that the note did not mention, and say what is not visible
that a person would need to check in person. Do not say whether the vehicle is safe or legal.

The prompt's last sentence matters as before. A typical sample answer (illustrative): "The photo shows a crack running about two-thirds across the lower screen, longer than a 'small chip'. The wipers are parked over it, so the extent in the driver's eyeline cannot be judged." That answer sends a person to look, which is exactly the job.

When the report is bad: vague, duplicated or too cheerful

Three recurring problems spoil triage. The first is the vague note: "something wrong with the brakes". Have the AI draft a follow-up question for the driver or the office to send straight away, such as "Is the pedal low, the noise grinding or the vehicle pulling to one side? Is the warning light on?". The second is duplicates: five drivers on a shared vehicle each report the same cracked mirror. The AI can cluster them, but make sure the cluster is shown as one defect with five mentions, because a defect reported by five people is more worrying than one reported once.

The third is the too-cheerful report: "no defects" ticked at 06:02 on a vehicle that came back at 22:00 with a note about a noisy bearing. Compare today's check against the previous day's open defects. If a defect was reported yesterday and the vehicle is "clear" today, the system should ask who cleared it. The record the FMCSA rule asks for makes this explicit: before the vehicle is driven again, the carrier or its agent must certify on the report that the listed defect was repaired or did not need repair, and carriers keep the report, the certification and the driver review for three months. Whatever your rules, closing a defect should need a name and a date.

A realistic wrong turn: the confident wrong band

Suppose an assistant, given a loose list, sorts "steering feels a bit loose over bumps" into Band C because the phrase "a bit" sounds mild. The workshop plans it for Friday. The vehicle is driven for two more days. Nothing happens, but the next check finds worn track-rod ends. The lesson is not that the AI is useless; it is that "steering" belongs on the Band A list as a category until a qualified person has looked, and the prompt should say "any steering report is at least REVIEW". The fix is a change to the list and the rules, not a blame session. Make the same sort of change whenever a near miss shows a gap, and write the date and the reason next to the change.

A second illustration, about process: picture a small coach operator whose drivers mark "Not sure" on the drive-safely question for one report in five. Ignoring the answer would be the worst response. A better rule makes "Not sure" automatically Band A pending a quick call from the workshop lead. If those calls stay short and never turn into a telling-off, drivers learn that "Not sure" costs them five minutes, not a reprimand, and the answer becomes honest rather than rare. Triage rules shape behaviour, so design them as you would a safety procedure, not an algorithm.

Pipeline in practice: three ways to run it

Pick the lightest option that fits your fleet.

1. Copy and paste. Drivers use a paper or digital form; the office keys the day's reports into a sheet and pastes it into a chat assistant on a business plan with the standing prompt above. Time: ten minutes each morning. Cost: a business-plan seat (ChatGPT Business, for example, lists $20 a user a month billed annually or $25 monthly, with a two-user minimum; check the vendor page).
2. Form to sheet to assistant. A web form writes straight to a spreadsheet; a no-code automation sends new rows to an AI step and writes the band back. Time: a few hours to build. It is faster, but test it with a month of old reports before trusting it.
3. Fleet software. If your maintenance or fleet software already has driver inspection and defect workflows, use those and add AI only for the free-text sorting. Our survey of fleet admin for small businesses covers the services and fuel side, and predictive maintenance for a small workshop is a sober look at what AI can and cannot foresee.

Whichever route you take, do not send personal data you do not need. A defect report needs vehicle, area and note; it does not need the driver's full name or home details. Use driver IDs.

Checking that triage is working

Run a back-test before you rely on it. Take a month of past reports, including some where you know the outcome, and run them through the prompt. Compare bands with what the workshop actually did. You want zero cases where the AI put a defect lower than the workshop eventually treated it, and a handful of cases where it was more cautious than needed. The cautious misses are the cheap errors; the lenient ones are the expensive ones.

Then keep three numbers on a weekly sheet: reports received, reports moved up a band by a human, and reports moved down. Moves up are the signal that the prompt or list is too lax. After a month, adjust the Band A list or the rules, and note the change date. Pair this with the monthly paperwork check described in the tutorial on tracking driver qualification file expiry dates, since the same office habits cover both.

If the stakes feel high, that is because they are. A defect system is part of your safety management, and your insurer, your regulator and your drivers will each judge it by what happens on the bad day. Keep a person in the loop for every Band A and REVIEW decision, keep the official pages current and keep the records. If you want a second pair of eyes on the set-up, a 1:1 AI implementation call can walk through the form, bands and prompt with you.

One more worked morning: a three-vehicle van firm

To see how light this can be, take a firm with three vans and a part-time mechanic who visits twice a week. Drivers photograph a paper checklist in the yard and send it to a shared inbox. Each evening the owner pastes the notes into the standing prompt. The next morning the owner has a sorted list with Band A items on top, phones the mechanic about any B items and parks any vehicle with an A or REVIEW result until the mechanic or a garage has looked.

The saving is not in the mechanic's time, which stays the same. It is the owner's: the old routine was reading eleven handwritten notes before the first van left, sometimes after it had gone. The new routine is a four-minute paste and a three-line read. And the paper trail is better, because each note now carries a band, a reason and, once fixed, a name and a date.

Questions about walkaround checks and AI triage

Can AI decide that a vehicle is unsafe to drive?

Not as the final word. It can match a driver's wording to your own list of stop-now defects and flag anything uncertain, but the list itself must come from the regulator's published guidance and your manufacturer's manual. When the AI is unsure, the rule is to escalate to a person and keep the vehicle parked.

How long should I keep walkaround reports?

It depends on the rules that cover your operation. For the FMCSA rule on driver vehicle inspection reports (49 CFR 396.11), carriers keep the report, the repair certification and the driver review certification for three months from the report date. DVSA guidance has its own expectations, so read the current guide.

Should drivers report minor things like a scuffed mudflap?

Yes, and the form should make it easy. A culture where small defects get reported is how brake or tyre trouble gets caught early. Triage sorts the trivial from the serious afterwards, so never discourage reporting by making it a chore or by teasing small reports.

What photo should a driver attach to a defect?

One wide photo that shows which vehicle and which side, and one close photo of the defect with something for scale such as a coin or a hand. Add the registration or fleet number in the filename or the form field. AI reads photos poorly, so the driver's one-line description still matters.

Do paper forms still work?

Yes. Photograph the paper form each day, or key the defects into a shared sheet. The triage steps are identical; only the first step changes. A fleet can start on paper and move to an app once the severity bands are settled.

Further reads

• AI Software for Trucking Companies: What to Check for a Small Fleet — Check what fleet software to buy and what to ask it.
• Rental Car Companies Using AI: Seven Uses a Small Fleet Can Copy — See other ways a small fleet uses AI.
• Driver Scorecard Template: Fair Scoring From Telematics With AI — Combine defect reports with fair driver scoring.
• DOT Audit Checklist for a Small Fleet: Organise Records With AI — See where inspection records fit in an audit.
• AI Maintenance Costs: What You Pay After an Automation Goes Live — Budget for running costs once an automation is live.
• How Property Managers Use AI to Triage Maintenance Requests — Borrow the triage idea from maintenance requests.
• AI for a Car Rental Business: Enquiries, Availability and Damage Notes — Enquiry, availability and damage-note workflows for a small rental fleet: two source sheets, safe prompts, a photo shot list and where a person must decide.
• Dumpster Rental Software: Bookings, Swaps and Routes With AI — What a small yard needs from dumpster rental software, how AI can take bookings and plan swap routes, and a trial script to test any system.
• Driver Debrief Notes: Turn Voice Memos Into an Exceptions Log — A two-minute voice-memo debrief for drivers, turned by AI into a structured log of damage, delays, access problems and follow-ups.
• Driver Handbook Template: AI Drafts Rules for a Small Haulier — The sections a small haulier's driver handbook needs, a prompt to draft each with AI, and what a lawyer or adviser must check before it goes out.
• Driver Onboarding Checklist: From Offer to First Route With AI — A stage-by-stage checklist for onboarding a new driver, with AI-drafted letters, reminders and induction packs and care for personal data.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: 49 CFR 396.11 and 396.9 (current eCFR text, read November 2026); the DVSA 'Guide to maintaining roadworthiness' with its example defect report annexes and 'How vehicle defects are categorised in roadside checks and vehicle tests' on gov.uk, which you must read yourself for current prohibition categories; OpenAI's ChatGPT Business pricing page; vehicle manufacturer manuals.