AI & Intelligence for Eco / Energy / Heating / Solar — The Practitioner’s Playbook.

SectionThe honest reframe most AI agencies won't tell you

Generic "AI marketing" agencies are selling eco-energy installers a ChatGPT-blog-post package. Ten posts a month, scraped from competitor sites, lightly rewritten by a model the agency does not understand, published unedited under the founder's name, and then they wonder why traffic flatlines, why MCS auditors flag the technical content, and why TrustMark assessors raise eyebrows at the inaccuracies.

That is not AI strategy. That is AI cosplay.

The high-leverage AI use cases in solar, ASHP, GSHP, EV, battery and retrofit are not blog automation. They are: pulling roof angle, shading and panel-yield estimates from a single homeowner photo before the surveyor leaves the office; classifying scheme eligibility — BUS, ECO4, GBIS, SEG — from an intake form and a postcode in under five seconds; running retrieval-augmented generation over the actual scheme-rule PDFs so your sales team gives accurate answers at quote stage instead of guessing; summarising every sales call into a structured CRM record with a lead score; and, yes, AI-drafted content — but only when an MCS-qualified engineer reviews and signs off the technical claims before publication.

Each of those is measurable. Each saves surveyor hours or lifts close rate. None of them is "more blog posts." Read the playbook. Run it yourself, or have us ship it on retainer.

SectionThe eight-point audit we run on day one

Score your own AI stack red / amber / green this week. Three or more reds means the foundation is broken — fix that before any new tooling spend.

1. Survey-photo intelligence pipeline — A homeowner uploads a single roof photo at quote-form stage. The pipeline returns roof angle estimate, shading map, usable area, and an indicative panel-yield range before the surveyor's van leaves the depot. Removes 30–60 minutes of desk time per survey, and the output goes straight into the quote pack.
2. Scheme-eligibility classifier from the intake form — Postcode plus property type plus heating system plus household income band plus tenure go in. Out comes a probabilistic eligibility score against BUS, ECO4, GBIS, Home Upgrade Grant and SEG, with the relevant scheme rule cited. Front-of-funnel filtering that cuts unqualified leads by 30–50% and routes the qualified ones to the right sub-vertical surveyor.
3. RAG over BUS / ECO4 / GBIS / SEG scheme-rule documents — Retrieval-augmented generation over the actual scheme-rule PDFs, statutory instruments and Ofgem guidance notes. The sales team asks "is a flat in a Conservation Area with an LPG boiler eligible for BUS at the £7,500 grant?" and gets a sourced, paragraph-cited answer in three seconds. Not a hallucinated guess from a generic model.
4. Sales-call summarisation and lead-scoring — Every inbound call transcribed, summarised into a structured CRM record (technology, property, scheme, objection, next step), and lead-scored on intent. Surveyor time goes to A-leads first. Marketing reads the actual objection patterns at the end of every week, not at the end of every quarter.
5. MCS-engineer review on AI-drafted content — No unedited AI on the public site for technical claims. Every AI-drafted blog, FAQ entry, or scheme-explainer goes to an MCS-qualified engineer for review and sign-off before it publishes. The engineer's name on the byline. TrustMark and RECC are explicit about installer accountability for published technical content; one wrong heat-loss calculation in a published article is a regulatory exposure.
6. Customer-FAQ classifier from sales-call transcripts — The same call transcripts that feed lead-scoring also feed a clustering model that surfaces the top 20 recurring buyer questions per sub-vertical, ranked by frequency. That list becomes the FAQ schema, the next quarter's blog cluster, and the sales-team objection-handling script. Intelligence harvested from the work you are already doing.
7. Data-hygiene and consent management on AI training inputs — Customer call transcripts, intake-form data and photo uploads are personal data under UK GDPR. Lawful basis logged, retention windows defined, opt-out path documented, sub-processor list maintained, and no customer data shipped to a model provider that retains training rights. Boring, mandatory, and where most installer AI projects quietly fall over.
8. Productionisation with explicit fallback paths — Every AI tool has a documented behaviour for when the model is wrong, slow, or unreachable. The eligibility classifier defaults to "human review needed" on low-confidence outputs. The photo pipeline tells the surveyor "model returned low confidence, do a manual rooftop check." The RAG system surfaces source paragraphs even when the answer is uncertain. AI without a fallback is a liability waiting to happen.

Three or more reds — fix the foundation before commissioning new tooling spend.

SectionSix productised deliverables we ship per cycle

On a Foundation, Compound or Architect retainer, the same six outputs land in your portal each cycle. Industry-tuned, fixed scope, dated, owned by you.

Survey-photo intelligence pipeline. A homeowner uploads a single roof photo at quote-form stage. The pipeline runs an off-the-shelf vision model behind your form, returns roof angle estimate, shading footprint, usable panel area and an indicative kWp / annual-yield range, and writes the output directly into the surveyor's pre-visit pack. Built on public LLM and vision providers (OpenAI, Anthropic, Google) with explicit fallback to manual rooftop survey on low confidence. Time to first signal: 21–30 days. Owned by you, exported as written SOP plus the orchestration scripts.

Scheme-eligibility classifier from intake form. Intake form fields plus postcode plus property attributes go in; out comes a probabilistic BUS / ECO4 / GBIS / Home Upgrade Grant / SEG eligibility score with the relevant scheme rule cited. Threshold rules route eligible leads to the right sub-vertical surveyor; ineligible leads get a polite rejection with reasoning, which protects your CSAT and your review velocity. Time to first signal: 14–21 days.

RAG over BUS / ECO4 / GBIS / SEG scheme-rule documents. A retrieval-augmented generation system trained on the actual scheme-rule PDFs, Ofgem guidance, MCS standards and TrustMark code. Sales team types a question; the system returns a paragraph-cited answer with a confidence score and the source document. Public RAG patterns over public LLM providers — no opaque internal pipeline. Cuts misquoted scheme advice at quote stage to near-zero, which is where most reputational damage happens.

Sales-call summarisation and lead-scoring. Every inbound call recorded, transcribed, summarised into a structured CRM record (technology, property, scheme, top objection, next step, confidence), and lead-scored 1–5 on intent. The CRM updates the lead automatically; the surveyor sees the call summary before they ring back. Drives a 15–25% lift in close rate by the simple act of routing surveyor time to high-intent leads first.

AI-drafted plus engineer-reviewed content workflow. A documented draft-to-publish workflow: AI drafts a long-form post against a sub-vertical pillar, an MCS-qualified engineer reviews and edits, the engineer's byline goes on the published post, and the post ships. Author schema points at the engineer. Throughput rises 3–5x versus all-human drafting; technical accuracy stays at the level a TrustMark assessor would sign off. Never unedited AI on installer technical content.

Productionisation with fallback paths. Every AI surface in your stack — photo pipeline, eligibility classifier, RAG, summarisation — gets an explicit fallback runbook: low-confidence output, model unavailable, sub-processor outage. Documented in plain English, tested quarterly. The difference between "we use AI" and "we run AI in production" is whether the fallback exists and has been tested. Most installer AI projects we audit fail this test.

SectionWhat to do this week

Three actions, ranked by leverage. Same first three steps we ship in week one of a Foundation retainer for an eco-energy installer.

1. Time the lag between an inbound call and the CRM record being readable. Owner: founder. Time: one afternoon. Pull yesterday's inbound calls. How long after each call does a structured, searchable CRM record exist that another surveyor could pick up? If it is more than an hour — or, in most installers we audit, more than a working day — sales-call summarisation is your highest-leverage AI deployment.
2. Audit one of your last ten published technical blog posts for accuracy. Owner: technical director or MCS engineer. Time: 30 minutes. Pick a post on heat-pump sizing, ECO4 eligibility or solar-PV yield. Read it line by line. How many specific claims would survive a TrustMark assessor reviewing them? If the count is low, your AI-drafted content is shipping unedited and you have a regulatory exposure. The fix is not "less AI" — it is engineer review on every draft.
3. Decide DIY, DWY or DFY for the next 90 days. Owner: founder. Time: 30-min discovery call. We will confirm the right way in writing within two business days. See the three ways.

SectionFive questions eco-energy operators ask us about AI

What is a realistic ROI on AI for an installer in 2026? The honest answer: it varies by use case, and the agencies quoting "10x productivity" are selling you a vibe. Survey-photo intelligence saves 30–60 minutes of surveyor desk time per quote — at £45/hr blended, that is £25–£45 per quote, on quote volumes of 80–200 a month. Sales-call summarisation lifts close rate 15–25% by routing time to A-leads first; on a £6,000 average install, that is meaningful five-figure revenue per month from a six-week deployment. Eligibility classification cuts junk leads 30–50%, which protects surveyor capacity in scheme-window surges. Payback on a productised AI stack typically lands inside one quarter.

ChatGPT subscription versus a custom-built AI workflow — what is the difference? ChatGPT, Claude and Gemini consumer subscriptions are excellent for ad-hoc surveyor research, drafting initial copy, and exploring ideas. They are not a production workflow. The moment you want eligibility classification running automatically against every intake-form submission, RAG over scheme-rule PDFs cited at quote stage, or photo intelligence written back into your CRM — you are building a workflow, and that workflow needs version control, fallback paths, sub-processor management and consent handling. The custom build is not "fancier ChatGPT"; it is the difference between a tool one surveyor uses ad-hoc and an asset every surveyor relies on every day.

How accurate is RAG on scheme rules — can we actually trust it at quote stage? RAG over the actual scheme-rule documents is materially more accurate than asking a base LLM. Public RAG patterns with paragraph-level retrieval and citation typically land in the 85–95% range on factual scheme questions, with the remainder flagged as low-confidence by the system itself rather than answered wrongly. The discipline is to surface the source paragraph alongside every answer, set a confidence threshold below which the system says "escalate to a human," and have an MCS engineer audit a sample monthly. Production RAG with that discipline is more accurate at quote stage than the average new starter on the sales team.

Is engineer review on AI-drafted content really necessary, or is that a CYA exercise? It is necessary for two distinct reasons. The first is regulatory: TrustMark, RECC and MCS are explicit about installer accountability for published technical claims. A wrong heat-loss calculation in a published article is enforceable. The second is brand: eco-energy buyers are running a multi-thousand-pound, multi-month decision and they read the technical pages with care. One paragraph that contradicts MCS Standard MIS 3005 destroys the trust the rest of the site is building. Engineer review on AI drafts is not a CYA box-tick — it is the production gate that lets you ship 3–5x more content without ever shipping inaccurate content.

Can we run this ourselves with the playbook plus £750 audit? Yes. Most of the audit-and-fix list above is achievable in-house if you have a marketing manager plus a developer half-week per cycle plus an MCS-qualified engineer who can spend an hour a week reviewing AI drafts. The £750 audit gets you a written red / amber / green of all eight points, a prioritised next-step list with named owners and dates, and a copy of the workflow templates and prompt patterns we use ourselves. If you sign for DWY or DFY within 30 days, the audit fee credits against the first cycle.

SectionWhere to go from here

If you want this shipped end-to-end on a productised retainer, book a 30-minute discovery call. Tailored proposal in writing within two business days.

If you would rather have a senior practitioner reviewing your team's AI deployments each week, the coaching plans start at £750/month with rolling cycles and walk-away rights. If you have a hard deadline — an AI-pilot kick-off, a peak-season prep build, or a scheme-window surge you need eligibility classification ready for — the two-week embedded sprint lands a senior practitioner inside your tools for ten working days at £3,000 fixed.

Or run it yourself. Read this playbook end to end, run the eight-point audit, ship one deliverable a month for six months. Twice-quarterly office hours are open to anyone using the playbooks — bring your work, get reviewed, no charge.