An AI opportunity engine for the London commercial demolition market

The constraints

What made this hard?

The hard part wasn't any individual data source. The London Gazette publishes insolvency notices daily. Planning portals publish applications. Derating registers are public. Consultant websites name their pipeline projects. The problem is that each source, alone, is noise: London generates around 522 insolvency, liquidation, and CVL notices a month, and the vast majority have nothing to do with a £200k-plus commercial demolition or strip-out job. Getting from raw signal to a weekly feed of 10 to 15 actionable leads meant building a qualification layer that understood the commercial profile of a specific contractor — their geography, their value floor, their target work types, their no-go signals — and composing all the source streams around that profile so each additional source refined the shortlist instead of inflating it.

The other constraint was interpretation speed. Dense commercial text — a Gazette insolvency notice, a planning application, a derating register entry — carries the relevant signal buried inside several paragraphs of procedural language. Reading it manually would take 20 minutes per record. At source volumes this high, that's not a research workflow; it's a full-time job.

The approach

How did Tincture frame the problem?

I built an agentic pipeline with five layers working in sequence. Each layer was designed so additional sources refined the shortlist rather than inflating it — the composition problem is the design problem, and tuning it to a specific contractor's commercial profile is what turns public noise into actionable intelligence.

The build

What was shipped?

Signal ingestion. The system monitors five source categories in priority order: Gazette insolvency and CVL notices cross-referenced against charge holders and asset addresses; derating and vacancy triggers for buildings past the three-month rates-relief cliff; charge-holder and asset-recovery activity from bridging lenders and banks managing recovered commercial property; commercial vacancy listings beyond the inflection point where derating becomes the rational landlord move; and consultant-named project activity from architect and project-manager websites. Planning data is treated as the noisiest source and included as a sixth stream, heavily filtered.

AI interpretation. Each raw signal is processed by an AI layer that extracts the address, the registered company, the asset type, the named contacts, and the scale indicators from whatever form the source presents them in — a Gazette notice, a planning PDF, a derating register entry, a consultant's pipeline page. The AI layer cross-references those extracted fields against the contractor's qualification criteria: geography (London), value floor (£200k-plus), work-type relevance (strip-out, enabling works, refurbishment, change-of-use). A record that would take 20 minutes to manually parse and assess gets scored in seconds.

Contact resolution. Once an opportunity clears the qualification threshold, a contact-finding workflow identifies the right route into each lead. The route depends on the signal: an administrator named in a Gazette notice, a bridging lender holding a recovered asset, a lead architect or project manager named on a planning application, a charge holder listed at Companies House. Each qualified opportunity arrives with a named contact and a recommended approach channel.

Outreach drafting. Qualified opportunities are drafted as outreach messages before they reach the contractor for review. The draft is contextualised to the specific opportunity, the specific contact, and the specific signal that surfaced it — so the message reads as informed awareness, not a cold blast. Nothing goes out automatically; the contractor reviews and decides. The system sits upstream of the decision, not parallel to it.

Dashboard and delivery. Surfaced opportunities are presented with source context, confidence level, and recommended next move. The contractor's input is a focused 30-minute weekly review rather than ongoing monitoring.

The outcome

What were the results?

AI summarisation of dense commercial text, agentic signal monitoring across all five source categories, contact resolution, and outreach drafting are all working components. Based on the source volumes running through the build — 522 Gazette notices per month in London alone, with planning, derating, vacancy, and consultant signals layered on top — a live four-week pilot was projected to surface 50 to 70 qualified opportunities, with 20 to 25 carrying enough signal for direct outreach and 5 to 8 likely to convert to a warm conversation with a commercially relevant decision-maker inside the pilot window.

The pipeline has not yet been deployed to a live client.

The takeaway

What's the transferable principle?

The signal problem in a niche B2B sector is almost never data scarcity. The sources exist, they're largely public, and they're updated constantly. The problem is composition and interpretation at volume: each source individually is noise, and a human reading them manually can't cover enough ground fast enough to make the intelligence useful. An AI layer tuned to a specific commercial profile — the geography, the value floor, the work types, the contact routes — is what turns a set of public databases into a lead feed you'd actually act on. Building that tuning layer is the one-time investment. Running it is close to zero marginal cost per week.