Reading PEB Tender Packs with Azure AI Content Understanding
A custom document analyzer that turns several hundred pages of specification, drawings and query replies into a structured Project Information Sheet - every field returned with a confidence score and a grounded reference to the page and region it came from.
Industry
Time Required
Engagement Model
Solution
Current Phase
Technologies
A leading pre-engineered buildings manufacturer bids for steel building projects where a single tender pack runs to several hundred pages across a technical specification, drawing sets, query reply sheets, building area statements and indicative bills of material. Before a price can be built, an estimator reads all of it and fills a Project Information Sheet for every building in the tender.
This case study explores how Zelite Solutions built a custom Azure AI Content Understanding analyzer that turns tender packs into a structured, traceable Project Information Sheet.
The answer to a single field is rarely on a single page:
- The Project Information Sheet carries well over a hundred fields per building, all filled by hand.
- A single tender frequently contains eight or more non-identical buildings, each needing its own sheet.
- Values are scattered and sometimes contradictory – one tender stated an eave height of 18.0 m in the scope of work and 18.5 m in a query reply.
- Many fields are deferred to drawings or resolve only by combining a specification clause with a later query reply.
- Nothing linked a filled value back to the clause it came from, so checking a sheet meant re-reading the source documents.
- Bid capacity was bounded by estimator reading time.
Zelite Solutions built the extraction layer on Azure AI Content Understanding, the document analysis service in Microsoft Foundry, delivered to estimators through a model-driven Power App. A working prototype was proven on the client’s own tender packs before the build began.
- Structured Sheet, Defined Once: The Project Information Sheet is modelled as a field schema, so adding or changing a field is a configuration change rather than a rebuild.
- Extraction Across Documents: Values stated plainly are picked up directly, while values that resolve across several documents are worked out by the service and returned in the same output.
- Confidence and Source on Every Value: Each field carries a confidence score and a reference to the page and region it came from – shown as High, Medium or Refer.
- Review, Versioning and Approval: Corrections are tracked in the Power App and the sheet is sent for approval, with a complete maker-checker trail.
- Man-hours per tender pack: 30 hrs. reduced to 10 hrs. (67% less).
- Tender receipt to checked PIF set: 5 working days reduced to 2.
- Man-hours per building: 3 hrs. reduced to 1 hr.
- Tenders bid per month: increased from 6 to 10.
- PIF fields with a traceable source: from 0% to 100%.
- Conflicting values: now flagged before pricing, with both sources cited.
The solution delivered significant improvements to the bid process:
- Faster Checking: Reviewing a sheet means opening citations rather than re-reading the tender pack.
- Full Traceability: Generated fields are traceable on the same terms as directly extracted ones.
- Conflicts Surfaced: Contradictions between the specification and query replies are flagged for a decision instead of being missed.
- Clear Gaps: Values deferred to drawings are recorded as such, giving the estimator a clear list of what still needs sourcing.
- Audit Trail: A complete record of who changed what, when, and on what basis.
What was a manual reading exercise, repeated building by building, becomes a consistent extraction step that produces a structured Project Information Sheet with confidence and sources attached to every field. The client gains a traceable, auditable front end to the bid process, and a platform the design-specification and quotation phases can be built on without rework.
