The challenge
The firm's core offering was the building energy performance audit (U-values, thermal bridges, regulatory PENB / energy class certificates). Each audit was manual: an auditor would read the uploaded architectural drawings, key dimensions into a spreadsheet, run the regulatory calculation engine, cross-check against the building product catalogue and generate the audit report. Typical turnaround was 5 working days and the firm was losing deals — especially in the B2C channel, where homeowners expected an immediate online quote and audit when they listed a property for sale.
The solution
We designed a web application with an agentic backend that takes an uploaded drawing (PDF or DWG) and returns a priced, energy-compliant audit report. A vision-capable reading agent extracts the geometry and element schedule from the drawings. A calculation agent runs the regulatory energy model and cross-checks it against the firm's reference product catalogue via internal API. A pricing agent assembles the audit fee under the right channel rules (B2C fixed pricing, B2B frame contract pricing). A delivery and CRM agent pushes the audit report to the customer, files it with the regulator-required archive, and creates the opportunity in Salesforce — assigning a senior auditor for B2B engagements over a threshold for human review and sign-off.
Why this was hard
It is not the LLM part that was hard. The hard part was building a reading agent that could reliably extract geometry from real architectural drawings — which come in every possible format, every possible draftsman convention, and every possible level of neatness. We assembled a golden set of 400 drawings, co-labelled them with the firm’s senior auditors, and built an evaluation harness that the agent had to pass on every release.
The architecture
Four specialised agents behind a single web application:
- Drawing reader — vision-language pipeline that extracts geometry, element schedule, and annotations from the drawing. Outputs a structured building model.
- Energy calculator — runs the regulatory calculation against the structured model using the firm’s existing, audited calculation engine. The agent does not invent the calculation — it orchestrates the tool call.
- Pricing & report agent — applies channel-specific pricing rules (B2C fixed pricing, B2B frame contracts), and assembles the final audit report against the regulator’s required template.
- Delivery & CRM agent — for B2C, triggers the customer delivery flow and files the report to the regulator archive; for B2B, creates a structured opportunity in Salesforce and assigns a senior auditor for sign-off.
Every step is logged. Every drawing extraction is stored with its input, output, and confidence score. Senior auditors can inspect any calculation after the fact and flag it for re-review.
Governance for a regulated calculation
Building energy performance audits are regulated work. We did not try to replace the existing regulatory calculation engine — we wrapped it. The agent produces the inputs, the engine produces the answer, and every run is archived for audit. The external supervisor signed off after a two-week walkthrough.
What the head of the firm said
“Our B2C channel was losing customers at ‘wait five days for an audit’. Now the audit report is ready before they finish their coffee. Our B2B auditors are now doing the interesting part of their job again — review and judgment, not data entry.”
Services deployed
- Custom AI Agents (four specialised agents)
- Knowledge Engineering (drawing-to-structure extraction, regulatory ontology)
- AI Governance (audit log, evaluation harness, regulator sign-off)