AI Tender Drafting for a Construction Firm
A Swiss construction firm specialising in residential renovations was losing tenders on response time. We built an offer-drafting assistant that parses incoming tender requests, extracts the scope of work, cross-references the firm's past projects and catalogue prices, and produces a structured first-draft offer for the project manager to review. Draft time dropped from days to hours.
Background
The firm has around forty employees and a strong reputation for renovating older residential buildings in the German-speaking part of Switzerland. They receive somewhere between eight and twelve tender requests per week through a mix of channels: architect briefs as PDF attachments, direct customer enquiries by email with photographs, and occasional formal call-for-tenders documents. Each request had to be read in full by one of the two project managers, then translated into an internal scope-of-work document, costed against the firm's catalogue and against comparable past renovations, and finally written up in a presentable offer document with the firm's standard clauses. End-to-end this took two to four days per offer, and the firm was visibly losing time-sensitive opportunities to faster competitors. The owner's instinct was to hire another project manager; he wanted to test whether a tool could buy them that capacity instead.
Solution
We built a drafting pipeline that ingests tender requests from a dedicated mailbox, extracts structured scope information from the attached PDFs and email body, and produces a first-draft offer in the firm's own template. Document understanding runs through Claude — both the text content and the layout cues in floor plans and PDF tables. The pipeline cross-references the request against two internal data sources: a vector index of the firm's past projects (built in PostgreSQL with pgvector) and the current catalogue of materials and labour rates. The orchestration layer in Langflow walks through scope extraction, comparable-project retrieval, line-item drafting, total estimation with explicit confidence flags, and finally rendering into the standard Word template the firm has used for years. Everything runs on EU infrastructure in Frankfurt under Article 28 terms; for tenders involving cantonal or federal authorities, an on-premise variant was scoped but not yet activated. The output is never sent automatically — it lands in the project manager's inbox as a draft, with the source PDF and the comparable projects linked next to each line item.
Outcome
First-draft turnaround dropped from two to four days down to a few hours, with the project manager spending most of that time on review and pricing judgement rather than on writing. The firm responded to several tenders within twenty-four hours of receipt during the first two months of operation, something they had previously considered impossible. The owner's working hypothesis — that response time alone would shift their hit-rate on competitive tenders — has so far held in the project manager's qualitative reports, though the firm wants another quarter of data before claiming a hard number. Importantly, no offer has gone out without project-manager review, which is the design intent.
Lessons Learned
Three honest lessons. First: tender PDFs are messier than any cleanly synthesised demo. Our first prototype handled the firm's nicely formatted reference tenders beautifully and choked on a real-world scanned architect brief on day one of pilot. We rebuilt the ingestion layer around the assumption that every PDF is partially garbage and that the pipeline must surface its own uncertainty rather than hide it. Second: comparable-project retrieval was the single highest-impact component. The first version drafted offers from the catalogue alone and produced numerically plausible but commercially naive line items; once we wired in the vector search over past renovations, the drafts started reading like the firm's own work. Third: the project managers initially treated the tool as a threat to their judgement. We addressed that head-on by designing the review UI to make the AI's reasoning visible — every line item links to the source paragraph in the tender and the comparable past project — and within two weeks the conversation shifted from "is it correct" to "can it also do the cover letter".
A similar problem?
Drop us a short note and we will reply within one business day.
Request a project