Phoenix Journal · Ductwork
AI is starting to move extract maintenance from fixed calendar dates towards live, evidence-led timing - here is what that means for scheduling, cost and compliance.
The shift under way
For years, cleaning and servicing your extraction system has run on the calendar - a date in the diary, a visit, a certificate, repeat. AI is quietly moving that model from fixed intervals towards evidence, using live data to suggest when work is genuinely due rather than when the diary says so.
The traditional approach is time-based, or planned preventative maintenance. You clean the ductwork every three, six or twelve months depending on how hard the kitchen runs, and that rhythm has served the industry well because it is simple and defensible. The weakness is that it treats every quarter the same. A kitchen that switched to heavy frying halfway through the period is treated exactly like one that went quiet over a slow season, so you either over-clean and waste money or under-clean and carry risk you cannot see.
What AI adds is a feedback loop. Sensors on the fan, in the plenum and along the duct feed readings into software that learns what normal looks like for your specific system - airflow, static pressure, motor current, run hours - and flags when the pattern drifts. Grease-density sensors in some modern canopies now monitor build-up directly and raise an alert as deposits climb towards a set threshold. Instead of guessing, the system points at the exact asset and the likely date it will need attention, and industry data suggests these tools can surface a developing fault a couple of weeks or more before it would otherwise stop the kitchen.
Crucially, this is decision support, not autopilot. The better platforms surface a recommendation - clean this run, inspect that fan - and leave a competent human to approve it. That distinction matters enormously once compliance enters the picture, which it always does in a commercial kitchen.
Up to a point, yes - and the direction of travel is clear. Grease build-up in extract ductwork is the single biggest fire risk in a commercial kitchen, with around 70% of kitchen fires linked to fat and grease accumulating in the extract system. That build-up is not random. It tracks cooking hours, cooking method and volume, which are exactly the things a monitored system can measure. Feed enough of that history into a model and it can estimate how quickly deposits are forming and when they will reach a level that needs clearing.
Where AI is strong is spotting the early signals of a problem: a fan drawing more current than usual, static pressure creeping up as a duct narrows, airflow dropping at the canopy. Each of those is a proxy for restriction, and catching them early turns a costly emergency call-out into planned work - the kind of unplanned failure that typically carries a far higher price than the same job scheduled in advance, before you count lost covers during a forced closure.
What AI cannot yet do is replace a physical, verified clean. A model can tell you a duct is probably dirty; it cannot reach inside, remove the grease to bare metal and prove it is gone. Predictive tools are best understood as a smarter trigger for the visit, not a substitute for it. They tell you when to act with more confidence than a calendar ever could - then a trained engineer still has to do the work and evidence the result.
This is the part that stops AI running the show on its own. In the UK, kitchen extract cleaning is judged against TR19 Grease, the BESA specification for fire risk management in kitchen extraction. It sets out how a system should be cleaned - to bare metal across canopies, plenums, ductwork, fans and filters - and how the result is verified, using deposit-thickness measurements taken before and after so you can prove the grease has actually been removed. Many insurers expect that evidence: some will only indemnify a kitchen where a recognised TR19-compliant contractor has carried out the cleans, and a fire traced to inadequate ductwork cleaning can put your right to claim at risk.
Sitting alongside that is LEV testing. If your extraction is classed as local exhaust ventilation controlling exposure, COSHH Regulation 9 requires a thorough examination and test at least every 14 months, carried out by a competent person to the standard set out in the HSE's guidance, with records kept for at least five years. That is a legal interval, not a suggestion - you are in breach from the day after the window closes.
Here is the tension. A predictive model might reasonably conclude a lightly used duct does not need cleaning as often as the calendar says. But your insurer, your fire risk assessment and the LEV regime are all built on documented, human-verified intervals. So AI does not get to overrule them. In practice the sensible pattern is to let AI tighten scheduling within the compliance framework - bringing a clean forward when the data says the kitchen is dirtying faster than expected, never pushing a statutory test beyond its legal limit. The machine advises; the competent person, the certificate and the specification still govern.
Start modestly and let the value prove itself. You do not need a wholesale technology programme to benefit - you need better information feeding decisions you already make. The most useful first step is simply capturing good data: accurate run hours, honest cooking profiles, and the deposit-thickness readings your TR19 cleans already produce. That history is the raw material any predictive tool depends on, and it is worth gathering even before you buy a single sensor.
From there, monitoring on your most critical assets - typically the extract fan and the longest duct runs - gives you early warning where a failure hurts most. Treat the output as a prompt for a conversation with your cleaning contractor, not an instruction. The goal is a schedule that flexes with real usage while staying anchored to your legal and insurance obligations, so you spend where it counts and never drift out of compliance.
Done well, this also steadies your costs. Fewer surprise call-outs and cleans timed to genuine need make spending far easier to forecast, which feeds naturally into how you build a maintenance budget that prevents surprises. AI is not replacing the engineer, the certificate or the specification any time soon - but as a way to decide when to act, it is starting to earn its place.
Questions
Not on its own. TR19 Grease frequencies and your fire risk assessment are built on documented, human-verified intervals that insurers rely on. AI can bring a clean forward when live data shows the system is dirtying faster than expected, but it should never be used to push cleans or statutory LEV tests beyond the limits your insurer, COSHH and your fire safety obligations require. The safest approach is to let AI tighten scheduling within that framework, not override it.
No. Sensors and AI can tell you a duct is probably dirty and roughly when it will need attention, which is genuinely useful for timing. What they cannot do is physically remove grease to bare metal or prove it has gone. A trained engineer still has to carry out the clean and record deposit-thickness measurements before and after, and where LEV testing applies a competent person must complete the thorough examination. Treat the technology as a smarter trigger for the visit, not a substitute for it.
Phoenix Duct Clean · by the numbers
Phoenix surveys and cleans kitchen and building ductwork to the TR19 standard - measured, cleaned and certificated, UK-wide.