The problem
Clipboards record intentions, not outcomes
Walk into the back of most mid-sized restaurants and you'll find the hygiene system on a wall. A clipboard. A temperature log with neat columns. A cleaning rota with initials.
Here's the problem: those sheets are usually filled in after the fact, in a hurry, by people who were busy doing the actual work. The log says the walk-in fridge was 3°C at 10am. Whether anyone actually looked is another question.
That isn't a character flaw. It's a design flaw.
Who experiences it
Kitchen managers carry the risk but can't be everywhere. Line staff are asked to stop and record things in the middle of service. Owners of two or three outlets have almost no view of what happens when they aren't there. And customers experience the consequences only when something has already gone wrong.
Why the current process breaks
- It depends on memory. Checks happen when someone remembers, and busy shifts are exactly when people forget.
- It's retrospective. A paper log tells you what was written down yesterday. It can't tell you a fridge door has been ajar for forty minutes right now.
- It doesn't aggregate. One missed check is noise. The same check missed every Saturday night is a pattern — but nobody reads three months of clipboards to find it.
- It's built for the inspector, not the kitchen. The record exists to be shown, not to change what happens next.
What makes it hard
Kitchens are hot, wet, loud and fast. Any solution that adds steps for staff during service will be quietly abandoned. Hygiene rules vary by jurisdiction, and the consequences of getting it wrong — for customers and for the business — are serious. Whatever replaces the clipboard has to be less work, not more.
The AI thought
Let the system notice. Let people act.
This is where AI changes the equation — not by "running the kitchen", but by doing the one thing humans are bad at in a busy environment: continuous, boring attention.
Three capabilities matter here, and none of them is exotic any more:
- Anomaly detection on sensor data. Cheap wireless temperature sensors can report every few minutes. A model learns what "normal" looks like for each unit — the post-delivery spike, the overnight dip — and flags what isn't.
- Image understanding. A quick photo of a prep station, a sanitiser bucket or a labelled container can be checked for the obvious: is it there, is it labelled, is it dated?
- Language generation. Instead of a cryptic alarm, the person on shift gets a plain instruction: what's wrong, where, and what to do first.
What should NOT be automated
Judgement stays human. Deciding to discard stock, close a section, or call a supplier is a manager's call. AI should never sign off a check on someone's behalf, and it shouldn't be used to monitor individual staff performance by camera. The point is to watch the food and the equipment, not the people.
A checklist proves someone was asked to check. A system proves the food was safe.
The goal isn't more monitoring. It's moving attention to the moment it matters.
The solution
An AI-assisted hygiene loop
The solution isn't one product. It's a small operational loop with four parts.
1. Automatic evidence where possible
Temperature sensors in every fridge, freezer and hot-holding unit replace the hand-written readings entirely. That's the single biggest win: the data that matters most is also the data that's easiest to automate.
2. Quick photo check-ins where it isn't
For things a sensor can't see — station cleanliness, labelling, allergen separation — staff take a single photo at defined moments (opening, mid-shift, closing). An image model checks for the basics and flags anything that looks off for a human to review.
3. Smart, sparing alerts
The system's job is to say less, not more. Alerts are ranked: act now (a fridge out of range during service), act today (a missed closing photo), and review this week (a recurring pattern). Only the first category interrupts anyone.
4. A daily summary that writes itself
At close, the manager gets a one-screen summary: what was checked, what was flagged, what was resolved, what's still open. Over time, it becomes the audit trail — one that reflects what actually happened.
How it works
From problem to outcome
Problem
Hygiene checks are logged on paper, often after the fact, and nobody sees a lapse until an inspection or a complaint.
Input
Fridge and freezer temperature sensors, photo check-ins at key stations, delivery logs, cleaning schedules and staff rotas.
AI / Agent
A model watches the readings and photos for patterns — a fridge drifting warm, a station not photographed, a cleaning task always skipped on the late shift.
Decision
It ranks what needs attention now versus what is a trend to fix, and flags only what a manager should see.
Action
A short, specific prompt goes to the right person: 'Walk-in fridge 2 at 7°C for 40 minutes — check the door seal and move high-risk stock.'
Outcome
Fewer silent failures, an audit trail that reflects reality, and managers who spend time fixing problems instead of chasing paperwork.
Level 1
Manual
Paper logs filled in by staff, reviewed occasionally, mostly for inspections.
Level 2
AI-assisted
Sensors and photo check-ins; AI flags anomalies and drafts the daily hygiene summary.
Level 3
AI-automated
Routine alerts routed and escalated automatically; logs compile themselves.
Level 4
AI-agentic
An agent coordinates follow-ups across sites, schedules maintenance and prepares audit packs — with managers approving actions.
Implement this
0/8 done
Identify the workflow
Map one kitchen's hygiene day: opening checks, deliveries, temperature logs, cleaning, closing. Write down who does what and when.
Identify repetitive tasks
Circle the tasks that are done the same way every day — temperature readings, cleaning sign-offs, delivery checks. These are the candidates.
Define required inputs
Decide what evidence each check needs: a sensor reading, a photo, a tick. Prefer data that is captured automatically over data someone has to remember.
Select the appropriate AI capability
Anomaly detection for temperatures, image checks for station photos, and a language model to turn alerts into clear instructions. You rarely need all three on day one.
Build the workflow
Start with one high-risk area — cold storage is usually it. Connect sensors, set thresholds, route alerts to a single shared channel.
Add human approval where required
Discarding stock, closing a station or changing a supplier stays a manager's decision. The system recommends; people decide.
Measure the result
Track how many issues were caught before service, how long they took to resolve, and how much time managers spend on hygiene paperwork.
Iterate
Tune thresholds that cause noise, add the next area, and retire paper steps only once the digital version has proven itself.
Business impact
Where the value actually shows up
It's tempting to promise a percentage. We won't. The real impact depends on the kitchen, the starting point and how well the system is introduced. What we can say is where the value comes from.
- Time. Managers stop reconstructing logs and chasing initials. Staff stop pausing service to write numbers on a sheet.
- Cost. Catching a failing fridge early can mean moving stock instead of throwing it away. Sensor data also tells you which equipment needs maintenance before it fails.
- Quality. Issues are caught during the shift, not discovered at the next inspection.
- Scalability. For multi-outlet operators, every kitchen reports in the same way, so the owner can see all sites on one screen.
- Trust. An honest record is a better defence than a perfect-looking clipboard.
The takeaway
The opportunity isn't to replace the hygiene process with AI. It's to redesign the process around what AI can now notice.
Share this thinking
The newsletter
Think better about AI.
One practical AI idea. One real-world problem. Delivered periodically.