How Bimmer.AI Assesses Used BMW Risk

The pipeline

1. Input normalisation. The user supplies a listing (text or screenshot), a VIN, or a UK registration plate. We extract year, model, badge, mileage, price, and seller claims.
2. Engine identification. A deterministic chassis × engine lookup (180+ row internal reference, published publicly) matches the model + year to the exact engine code. This step is not LLM-driven — we saw too many hallucinations on model → engine mapping in early tests.
3. Failure database lookup. If the resolved engine has a record in our engine failure database (N47, B47, N57, B57 today — petrol and M engines in development), we inject that record into the buyer-report generation context. Each record contains structured failure modes, severity, frequency, mileage windows, and UK repair-cost ranges.
4. Recall prompts. Known recalls relevant to the resolved engine are flagged for user verification at gov.uk/check-vehicle-recall.
5. Context-aware synthesis. A Claude-family language model synthesises the structured inputs into a decision-led buyer report — buyer/seller perspective, buy/negotiate/walk verdict, negotiation points. The model's role is language, not fact invention.
6. Editorial review. Our public SEO content (this page, the engine guides, the pre-purchase inspection checklist) is hand-written and edited, not LLM-generated.

Data sources

• BMW chassis × engine reference — compiled internally from BMW press archives, manufacturer spec sheets, and UK trade press. 180+ rows, updated as new generations launch.
• Engine failure database — per-engine records covering failure modes, severity/frequency, onset mileage ranges, UK repair-cost ranges, preventative schedules. Sources per record: DVSA and NHTSA recall databases (public), Wikipedia engine pages (CC-BY-SA), BMW press archives, forum consensus (linked, not copied), DVSA MOT advisory aggregates, and independent specialist maintenance schedules.
• DVLA Vehicle Enquiry Service — registration → factory spec (MOT due, tax status, emission class). Integrated at the Worker layer; frontend integration in progress.
• DVSA MOT History — mileage and advisory timeline by registration. Integrated at the Worker layer; frontend integration in progress.
• Listing-text parsing — we extract price, mileage, seller claims, and presentation-quality signals from any pasted listing text or screenshot.

See Sources and editorial standards for the full citation and review policy.

What we deliberately don't do

• We don't invent engine codes. If the model/year doesn't match a chassis in our reference table, the report says so rather than guessing.
• We don't fabricate failure patterns. Engines not in the failure database (currently: most petrol engines, all V8s) get a clean general-knowledge report without our structured failure block — graceful degradation, not hallucinated detail.
• We don't copy competitor content. We explicitly rejected scraping bimmerboom.com and similar AI content farms as sources — they compound errors. All our failure records are first-party-curated.
• We don't replace a physical inspection. Bimmer.AI is the first filter and the negotiation-prep tool, not a substitute for a specialist workshop.
• We don't replace an HPI Check. See HPI Check vs Bimmer.AI — provenance checks are a different layer.
• We don't fake reviews or testimonials. If and when we publish user testimonials, they'll be real and verifiable.

Limitations

• Engine database coverage. Phase 1+2 covers UK diesels from 2007 (N47, B47, N57, B57). Petrol engines and V8s are on the roadmap. Until they're in, reports on those engines fall back to general knowledge without the structured failure block.
• Cost ranges are indicative. UK repair costs vary significantly by region, workshop, parts supply, and labour rates. Our ranges are typical indie-specialist figures — get written quotes before committing.
• No model can predict the future. A clean-history car can still have a failure we didn't flag. We reduce surprise; we don't eliminate it.
• AI components are used for language. The language model helps us frame decision-led advice in buyer-friendly English. It does not invent facts about engines; those come from the structured database.

Editorial and correction policy

Our public SEO content is reviewed and updated on a rolling basis. If you find an error — a failure-cost range that's off, an affected-model list missing a variant, a recall reference that's stale — email [email protected] with the specifics. We publish corrections on the affected page with a dated update note.

We're a small team and our expertise is in used-BMW buyer intelligence specifically. We don't claim to be a general automotive encyclopaedia. When we're uncertain, the content says so.

Related