AI & Intelligence for Automotive — The Practitioner’s Playbook.

SectionThe honest reframe most AI-marketing agencies won't tell you

Generic AI-marketing agencies are knocking on dealer doors with a "ChatGPT content package" — fifty AI-drafted blog posts a month, an AI chatbot bolted onto the bottom of every page, and an "AI-powered SEO" upsell that's three Make.com webhooks in a trench coat. Used dealers, franchise dealers, garages, MOT centres, tyre fitters, body shops and EV specialists are paying £1,500–£4,000 a month for output that the FCA wouldn't let them publish on a finance page and that a technician wouldn't sign off on a service description.

The high-leverage AI applications in automotive are sitting unused. VRM-image classification — a model that takes the photos a buyer or appraiser uploads and scores damage, panel-gap drift, tyre wear, dashboard mileage, and condition grade — is a five-minute appraisal that used to take 45. An FCA-aware finance-eligibility classifier — one that actually displays representative APR, the representative example, and the right disclosures depending on whether the customer is being shown a regulated PCP, HP or PCH product — is a compliance asset, not just a conversion lift. Part-exchange valuation from a buyer photo plus a reg lookup, RAG over Auto Trader market-pricing data and manufacturer service-bulletins, sales-call summarisation with lead-scoring against actual buyer intent — these are the levers. AI-drafted blog posts are not.

This playbook is the structural fix. AI where it earns its keep — appraisal, finance compliance, valuation accuracy, knowledge retrieval, sales intelligence — and a hard human-in-the-loop boundary on anything a technician or a finance officer needs to sign. Read it, run it yourself, or have us ship it on retainer.

SectionThe eight-point audit we run on day one

Score your own use of AI red / amber / green this week.

1. VRM-image classification (damage / wear / mileage estimation from photos). A classifier that ingests appraisal or buyer-uploaded photos and returns: panel-by-panel damage scoring, tyre tread estimation, dashboard mileage OCR with cross-check against the V5 / VRM lookup, condition grade (1–5), and an exception flag where the photos contradict the listing. Most dealers do this manually at 30–45 minutes per appraisal. A trained classifier does it in under 60 seconds with a confidence score that flags edge cases for human review.
2. FCA-aware finance-eligibility classifier with proper APR + representative example display. Soft-search eligibility scoring on every VRM page with the right FCA disclosures rendered conditionally — representative APR, representative example with the example deposit / term / total amount payable / total amount of credit, the "subject to status" line, the broker / lender disclosure, and the commission-disclosure language post the recent regulatory tightening. Most AI finance bolt-ons display the APR figure and skip the rest. That's a regulatory exposure, not a feature.
3. Part-exchange valuation from buyer photo + reg lookup. A buyer uploads four photos of their car, types in the reg, and the system returns a part-ex valuation range (low / mid / high) with the model's confidence and a flag where the photos contradict the V5 or the MOT history. Replaces the "tell us about your car in 14 fields" form that converts at 6–9% with a four-photo flow that converts at 25–35%.
4. RAG over Auto Trader market-pricing + manufacturer service-bulletins. A retrieval-augmented system that, when a sales executive or service advisor types a query, pulls back current Auto Trader market-pricing data for the make / model / year / mileage / spec, plus the relevant manufacturer service bulletins, recall notices, and known-fault history. So the SE pricing a part-ex or the SA scoping a job has the actual market data and the known-fault context, not their memory from three months ago.
5. Sales-call summarisation + lead-scoring. Every inbound and outbound sales call transcribed, summarised, and scored: buyer intent (cold / warm / hot), finance vs cash, part-ex flagged, test-drive booked, follow-up window. Summaries land in the CRM against the lead with the next-action recommended. Replaces the SE writing two-line CRM notes that lose half the context.
6. Technician-reviewed AI-drafted service content. Where AI does draft service-page copy, MOT-explainer content, or model-specific maintenance guides, the workflow is AI-draft → IMI-credentialled technician review → publish, not AI-draft → publish. The technician owns the byline. The AI is a typing-speed multiplier, not the author. Most agency packages skip the technician review entirely.
7. FCA-compliant AI-drafted finance copy review. Same principle, harder rules. Any AI-drafted copy that touches motor finance — PCP explainers, HP comparisons, eligibility-criteria pages, finance-FAQ — runs through a compliance review against the FCA CONC sourcebook before publish. The reviewer can be a senior in-house compliance officer or a contracted FCA consultant. AI cannot publish unreviewed finance copy. There is no version of this rule that flexes.
8. Productionisation with technician fallback. Every AI feature in production has a defined fallback: if the VRM classifier returns low confidence, the appraisal routes to a human appraiser. If the finance classifier returns an edge case, the lead routes to an F&I officer. If the RAG system can't find a confident answer, the SE escalates. The system is designed to fail safe to a human, not to bluff an answer. Most "AI in automotive" deployments skip this and ship hallucinations.

Three or more reds — fix the foundation before you let AI touch a customer journey.

SectionSix productised deliverables we ship per cycle

VRM-image classification pipeline. A trained image classifier deployed against your appraisal flow and your buyer-upload flow. Panel-by-panel damage scoring, tyre tread estimation, dashboard mileage OCR with V5 / VRM cross-check, condition grade (1–5), and an exception-flag pipeline that routes low-confidence cases to a named human appraiser. Built on a public LLM provider's vision API or a fine-tuned open-weights model, depending on your data volume and privacy posture. Replaces 30–45 minutes of manual appraisal with under 60 seconds plus a confidence score. Time to first signal: 21 days.

FCA-aware finance-eligibility classifier. Soft-search eligibility scoring deployed on every VRM page, with conditional rendering of representative APR, the representative example, the "subject to status" line, broker/lender disclosure, and the commission-disclosure language. Integrated with whichever F&I provider you run. The classifier handles the eligibility scoring; the FCA-aware presentation layer handles the disclosures. Compliance-reviewed before launch. The single most under-deployed AI surface in dealer SEO.

Part-exchange valuation from photo + VRM lookup. Four-photo buyer flow plus reg lookup that returns a part-ex valuation range (low / mid / high) with confidence scoring and exception-flagging where photos contradict the V5 or MOT history. Replaces a 14-field form converting at 6–9% with a four-photo flow converting at 25–35%. Lead routes to the sales CRM with the photo set, the reg, the valuation range, and the model's confidence attached.

RAG over market-pricing + service bulletins. Retrieval-augmented system over current Auto Trader market-pricing data, manufacturer service-bulletins, recall notices, and known-fault history. Sales executives and service advisors get a desktop / mobile interface where a typed query returns the relevant pricing and fault context with citations. Source-grounded answers only — the system refuses to answer where retrieval confidence is low rather than hallucinate. Used by the SE pricing a part-ex and the SA scoping a job.

Sales-call summarisation + lead-scoring. Every inbound and outbound sales call transcribed and processed through a structured summarisation pipeline: buyer intent (cold / warm / hot), finance vs cash, part-ex flagged, test-drive booked, next-action recommended. Summaries land in the CRM against the lead. Lead-scoring updates the priority queue for the SE team each morning. Time to first signal: 30 days.

Technician-reviewed AI content workflow. Editorial pipeline for any AI-drafted content that touches a service or finance topic: AI-draft → named human reviewer (IMI-credentialled technician for service content, FCA-aware compliance reviewer for finance content) → publish, with reviewer byline visible. Workflow tooling, review SLA, named-owner schedule, and a publish-block on anything that hasn't been signed off. The discipline that separates AI as a force-multiplier from AI as a liability.

SectionWhat to do this week

Three actions, ranked by leverage.

1. Run a VRM-image-classifier proof-of-concept on twenty recent appraisals. Owner: sales manager or used-car buyer. Time: 90 minutes. Take the photos from twenty appraisals you've already completed, run them through a public LLM provider's vision API with a structured prompt asking for damage scoring, tyre wear, dashboard mileage and condition grade, and compare the output against your appraiser's notes. If the model agrees with the appraiser on 16+ of 20, you have a viable production candidate. If it disagrees on 8+, you need a fine-tuning data set before you ship.
2. Audit your current finance disclosures against the FCA CONC sourcebook. Owner: founder or compliance officer. Time: 60 minutes. Open three live VRM pages on your site. Check for: representative APR visible above the fold, representative example with deposit / term / total amount payable / total amount of credit, "subject to status," broker/lender disclosure, commission disclosure. If two or more are missing, you have a regulatory issue independent of AI — but it also means any AI-drafted finance copy you publish will inherit the same gaps.
3. Decide DIY, DWY or DFY for the next 90 days. Owner: founder. See the three ways.

SectionFive questions car-dealer / garage operators ask us about AI

What's the actual ROI on VRM-image classification — is it worth the build? The numbers we see across pilot deployments: 30–45 minutes saved per appraisal, 4–8 appraisals per day per used-car buyer, so 2–6 hours of buyer time per day reclaimed for the activity that actually moves margin (sourcing stock, negotiating part-ex, managing the trade-disposal channel). The build is a 4–6 week deployment if your photo-management pipeline is clean, longer if appraisal photos live across phones, WhatsApp, and the DMS in unstructured form. Payback is typically 2–3 months on a single-site dealer, faster on a group. The bigger second-order effect is consistency — the model doesn't have a Friday-afternoon mood.

Can AI write FCA-compliant finance copy unsupervised? No. And anyone selling you a tool that claims it can is selling you a regulatory exposure. The FCA CONC sourcebook is specific about what must appear, in what order, with what prominence, on what kind of product. AI is a fast first-draft tool — it can produce a structurally-compliant skeleton, populate the representative-example fields from your product data, and flag missing disclosures. But the published copy needs sign-off from a human who understands CONC and can be held accountable for it. The cost-benefit still works heavily in AI's favour — an FCA-aware reviewer signing off AI drafts ships 5–10× the volume of a reviewer drafting from scratch — but the reviewer is non-negotiable.

How accurate is part-ex valuation from photos vs a physical inspection? On well-lit, four-angle photo sets with the dashboard captured cleanly, current models hit 85–92% accuracy against a physical-inspection benchmark on a £/£ basis at the mid-confidence band. Edge cases — flood damage, structural repairs, odometer tampering — the model flags rather than valuing, and routes to a human appraiser. The strategic point isn't replacing the physical inspection on every part-ex. It's qualifying the part-ex at the lead-capture stage so you get a 25–35%-converting four-photo flow instead of a 6–9%-converting 14-field form, then doing the physical inspection on the qualified leads only. Conversion lift compounds on enquiry-to-test-drive and test-drive-to-deal.

Why insist on technician review for AI-drafted service content? Two reasons. The first is liability — IMI-credentialled technicians have a duty of care, and a service-page recommendation that turns out to be wrong (wrong torque spec, wrong fluid grade, wrong service interval for a specific market variant) is a problem you don't want to defend in front of a customer or an insurer. The second is search-quality and trust — Google's helpful-content guidance, MIAFTR's standards on body-shop content, and reader expectations on technical content all reward named-author bylines with verifiable credentials. AI as the silent ghost-writer with a technician as the named, accountable author is a defensible model. AI publishing unreviewed under a generic "Service Team" byline isn't.

Can we run this ourselves with the playbook + £750 audit? Partially. The VRM-image-classifier proof-of-concept on a public LLM provider's vision API is a 1–2 day build for a competent dev. The finance-eligibility classifier needs a compliance reviewer in the loop and an FCA-aware presentation layer — call it 4–6 weeks if your dev is comfortable with regulatory work. The RAG system is achievable in 3–4 weeks if you have access to the source data; the longer pole is sourcing the manufacturer service-bulletins legally and structuring them for retrieval. Sales-call summarisation is the fastest win — 1–2 weeks from a clean call-recording feed. The £750 audit gives you a written red/amber/green of all eight points + named-owner / dated next steps + a worked specification for the highest-leverage build for your specific operation. 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 a senior practitioner reviewing your AI deployments, your FCA-aware presentation layer, and your technician-review pipeline each week, the coaching plans start at £750/month. If you have a hard deadline — a March or September plate-change-window prep, a new-franchise launch, an EV-only sub-brand going live with a finance-eligibility surface from day one — the two-week embedded sprint lands a senior practitioner in your account for ten working days at £3,000 fixed.

Or run it yourself. Eight-point audit + one deliverable a month + twice-quarterly office hours.