AI & Intelligence for Trades & Home Services — The Practitioner’s Playbook.

SectionThe honest reframe most AI agencies won't tell you

AI-marketing agencies sell electricians, plumbers and roofers a "ChatGPT content pack" — fifty blog templates, a chatbot widget on the homepage, a "powered by AI" badge for the footer. Three months later the trade owner has a Google index full of generic "10 Tips for Your Boiler" posts written in American English, a chatbot that can't book a slot, and zero measurable impact on quote volume or close rate.

Trades & home services is one of the highest-leverage AI categories in the country, and almost nobody is shipping the work that matters. The high-leverage uses are not blog templates. They are: photo-intelligence on the quote form (estimating roof area, spotting an EICR-fail consumer unit, reading a boiler badge from a phone snap); trade-specific qualification classifiers that triage incoming leads by job urgency, complexity and likely close-value; sales-call summarisation that turns a 25-minute survey call into a one-page job spec; customer-FAQ classifiers built from your own call transcripts; RAG over manufacturer install guides so your engineers stop ringing tech support to ask whether a Vaillant ecoTEC plus needs a specific filling-loop pressure.

This playbook fixes the structure. The photo-intelligence pipeline is the conversion lever. The qualification classifier is the close-rate lever. The RAG layer is the engineer-productivity lever. Read it, run it yourself, or have us ship it on retainer.

SectionThe eight-point audit we run on day one

Score your own AI surface red / amber / green this week.

1. Photo-intelligence on the quote form — A buyer with a leaking flat roof or a tripping consumer unit can take a photo from their phone. A vision-capable model can produce a rough estimate (roof square-metreage, likely cause of fault, age-bracket of equipment) in 4–8 seconds. Most trades sites still ask the buyer to "describe the issue" in a free-text box. Photo-intelligence pre-quote pipelines lift form-to-job conversion 30–50%.
2. Trade-specific qualification classifier — Boiler engineers need boiler-make + age + symptom. Electricians need EICR-fail vs pass-with-remedials vs new-install intent. Roofers need pitched-vs-flat + insurance-claim intent. A small trained classifier on the inbound enquiry routes the lead correctly and pre-fills the engineer's job sheet. Without this you are paying CPL on junk.
3. Sales-call summarisation + lead-scoring — Every survey call recorded, transcribed, summarised into job spec + lead score (hot / warm / cold / no-fit). Five minutes of admin per call collapses to thirty seconds of review. Junior estimators stop missing details; the founder gets a daily digest of every quote in flight.
4. Customer-FAQ classifier from transcripts — Your last 500 calls contain the real questions buyers ask. A topic classifier over those transcripts produces the actual FAQ list — usually 12–20 dominant clusters. That FAQ then powers the website, the quote-form micro-copy, and the chatbot's first-turn responses. Most trades publish a generic FAQ written from imagination.
5. RAG over manufacturer install guides + part-numbers — Vaillant, Worcester, Ideal, Baxi, Honeywell, Drayton, MK, Hager, Schneider, Wickes. Engineers spend 15–40 minutes per job ringing tech support or scrolling PDFs. A retrieval-augmented chat trained on your specific manufacturer corpus answers in 3 seconds with the cited page reference. Hours saved per engineer per week, immediately.
6. Engineer-reviewed AI-drafted content workflow — AI drafts service-area pages, FAQ entries, post-job write-ups; a Gas Safe / NICEIC / FENSA-registered engineer reviews and approves before publish. The "AI content" the agencies sell is unreviewed and harms ranking; reviewed AI content is faster than human-only and ranks well.
7. Emergency-vs-scheduled triage classifier — Inbound enquiries split into emergency (sub-90-second response, on-call routing, premium rate) vs scheduled (calendar booking, standard rate, batched response). A simple classifier on the form fields + free-text routes correctly and stops scheduled buyers triggering emergency callouts.
8. Productionisation with fallback to human triage — Every AI surface has a fallback. If the photo-intelligence model returns low-confidence, route to a human estimator within 15 minutes. If the classifier is uncertain, default to the safer triage. If the RAG layer can't find a citation, surface a "ring tech support" link. AI without fallback is a liability; AI with fallback is a force multiplier.

Three or more reds — fix the foundation before you pay for any "AI marketing" subscription.

SectionSix productised deliverables we ship per cycle

Photo-intelligence quote-form pipeline. Buyer uploads a photo from their phone — flat-roof patch, consumer unit, boiler badge, broken window pane, blocked gutter. A vision-capable model returns a structured pre-quote (estimated area, likely cause, age-bracket, recommended next action) within 8 seconds, surfaced to the buyer as a "ballpark estimate" and to the engineer as a pre-filled job sheet. Confidence-thresholded fallback to human estimator on low-confidence frames. Time to first signal: 14 days.

Trade-specific qualification classifier. A small classifier trained on your last 200–500 closed jobs that takes form-fields + free-text and outputs job-type + complexity + likely close-value + on-call vs scheduled. Routes the lead to the correct engineer's diary, pre-fills the quote template, and surfaces the close-probability so your office manager prioritises follow-up correctly. Time to first signal: 21 days.

Sales-call summarisation + lead-scoring. Every survey call recorded (with consent), transcribed, summarised into a one-page job spec + lead score. Daily digest to the founder, weekly trend digest to the operations manager. Stops the "what did we agree on the call?" loop, removes 80% of estimator admin, surfaces deals slipping through follow-up gaps.

RAG over manufacturer install guides. A retrieval-augmented chat layer trained on your engineers' actual manufacturer corpus — Vaillant, Worcester, Ideal, Baxi, Honeywell, Hager, Schneider, plus your own SOPs and Gas Safe / NICEIC / NAPIT compliance notes. Engineers ask in plain English, get cited answers with page references in under three seconds. Saves 2–6 hours per engineer per week from cycle one.

Engineer-reviewed AI content workflow. AI drafts service-area pages, post-job write-ups, FAQ entries and review-response copy. A registered engineer (Gas Safe / NICEIC / FENSA / OFTEC / NAPIT depending on trade) reviews and approves before publish, signed off in a workflow log. Output is technically accurate, ranks well, and ships at 3–5× the cadence of human-only content without the unreviewed-AI ranking penalty.

Emergency-vs-scheduled triage classifier. Inbound form + call enquiries classified within 5 seconds into emergency / scheduled / quote-only / no-fit, with downstream routing to the correct on-call rotation, calendar slot, or batched-response queue. Stops emergency calls hitting voicemail and stops scheduled buyers triggering after-hours premium response. Time to first signal: 10 days.

SectionWhat to do this week

Three actions, ranked by leverage.

1. Pull your last 100 inbound enquiries and tag them by job-type + outcome. Owner: founder or office manager. Time: 2 hours. You cannot ship a qualification classifier without labelled training data, and you cannot rank deliverables against revenue without knowing which enquiry shapes close. Most trades have this in the CRM and have never extracted it.
2. Pick one manufacturer corpus and stack the install guides. Owner: senior engineer. Time: 90 minutes. Whichever brand dominates your callouts (Vaillant for most boiler engineers; Hager or MK for most electricians; Velux for roofers fitting roof windows). PDF the install guides, fault-code references, and parts catalogues into one folder. That is the seed corpus for the RAG pilot.
3. Decide DIY, DWY or DFY for the next 90 days. Owner: founder. See the three ways.

SectionFive questions trades operators ask us about AI

Photo-intelligence on the quote form — what's the realistic ROI? Form-to-job conversion typically lifts 30–50% in the first 60 days. The mechanism is that the buyer gets a same-session ballpark, which removes the "wait for callback" drop-off, and the engineer gets a pre-quote so the on-site survey collapses from 45 minutes to 15 minutes. On a £300 average job-value, even a small operator sees four-figure-monthly revenue lift in the first cycle, plus engineer-time recovered.

Off-the-shelf chatbots vs custom for trades — what actually works? Off-the-shelf wins for one use: deflecting "what are your hours" / "do you cover BH8" / "do you do landlord certificates" — the top 12–20 FAQ clusters. Custom wins for everything else: photo-intelligence, qualification classification, RAG over manufacturer guides, sales-call summarisation. The agencies selling "AI chatbot for tradespeople" at £99/month are sitting on use-case one and ignoring use-cases two through six.

RAG accuracy on manufacturer install guides — is it actually trustworthy? Yes, with two non-negotiables. First, every answer must surface its citation (manufacturer name, document title, page number). Second, every answer must have a fallback when retrieval is low-confidence ("I couldn't find a specific reference — ring tech support on this number"). With those two, RAG over a curated corpus of Vaillant / Worcester / Ideal / Baxi guides is consistently more accurate than a junior engineer's memory and faster than ringing the manufacturer's helpline.

Do we really need an engineer to review AI-drafted content? Yes. Every trade has compliance exposure (Gas Safe Register, NICEIC, FENSA, OFTEC, NAPIT, Trustmark) and unreviewed AI content can publish factual errors that range from embarrassing to dangerous to actionable. Engineer review takes 5–10 minutes per page with a structured approval template, runs at 5× the cadence of human-only content production, and removes the ranking penalty that unreviewed-AI content now triggers.

Can we run this ourselves with the playbook + £750 audit? Some of it, yes. The customer-FAQ classifier from transcripts and the sales-call summarisation pipeline are achievable in-house with a marketing manager + a developer half-week using off-the-shelf API calls to public LLM providers. Photo-intelligence and the qualification classifier need labelled training data and confidence-thresholded fallbacks — most trades buy that one in. The £750 audit gives you a written red/amber/green of all eight points, named-owner / dated next steps, and a build-vs-buy recommendation per deliverable. Credit toward first cycle if you sign for DWY/DFY within 30 days.

SectionWhere to go from here

If you want this shipped end-to-end on a productised retainer, book a 30-minute discovery call.

If you'd rather have weekly senior coaching reviewing your AI surface — pipeline accuracy, engineer-review workflow, RAG corpus quality — the coaching plans start at £750/month. If you have a hard deadline, the two-week embedded sprint lands a senior practitioner in your account for ten working days at £3,000 fixed — the right call for an AI pilot launch (photo-intelligence pipeline + qualification classifier shipping together) or a post-storm-season scale-up where inbound volume has tripled and your office manager is drowning.

Or run it yourself. Eight-point audit, ship one deliverable a month, attend the twice-quarterly office hours, and the £750 audit credits in full toward a future DWY or DFY engagement.