Case study
Designing AI to sort building safety case documents and draft reports
Tetra Consulting's building safety managers start every building safety case by sifting a client's bulk upload of documents. Alongside the QA agent, we designed an AI classifier to sort, rename and tag those documents, and a report-writing assistant that works from the in-scope set with every claim linked to its source.
2025
- 18 hours
- average to write a building safety case report2025
- 80+
- building safety case reports a month2025
- 1–2 hours
- per project finding the relevant documents2025
The challenge
When a client commissions a building safety case, Tetra receives the building's documents in bulk: fire strategies, certificates, test reports, floor plans and policies, often badly named and sometimes out of date. Building safety managers spent one to two hours per project clicking through files to find what mattered, then cross-referenced sources that disagreed to compile building and structural descriptions. A report took 18 hours on average, and a team of nine was producing more than 80 a month.
What we did
- September 2025 — mapped the building safety case process from Tetra's requirements and example document sets: organising documents, writing the report, quality assurance
- Designed a classifier with a trainable taxonomy and an approve-or-reject loop for building safety managers
- Designed a report assistant pre-loaded with building safety legislation and guidance, with a one-click prompt for each report section
- Produced interface concepts for both with our in-house AI toolchain
What we built
The classifier reads each uploaded document against the taxonomy, marks it in scope, out of scope or an outdated version with a confidence score, extracts dates and building references, and renames it to a standard convention. Building safety managers approve or reject each suggestion, and their decisions train the model. The in-scope set then feeds the report assistant: pick a building, run the prompt for each section, and check every statement against a preview of its source document.
- In-scope, out-of-scope and outdated classification
- Confidence scores for human review
- Date and building-reference extraction
- Standard file naming
- One-click prompts for each report section
- Source links with document previews
The result
The design moves building safety managers from searching to checking: a bulk upload becomes an organised, named document set, and the report is drafted from sources they can verify. It was scoped as the foundation for automatic gap analysis, first-draft reports and publishing straight into PRISM. Tetra took the QA agent proposed alongside it into build.