UC1: Bid comparison ยท stage: research and design
Five bids, one comparison you can defend.
Dundir is designing a bid comparison that computes instead of generates: the same bids give the same result, every run, with the reasons written next to it. It is in research and design; the pilot tests it on your closed tenders.
The problem
A buyer receives five bids for the same work. Each has its own structure, its own grouping of line items, its own assumptions about materials and labour. To compare them fairly, someone has to line them up by hand: unit prices, omissions, delivery dates.
In a test at Skanska, comparing six bids by hand took 30 minutes. A generative AI tool did it in 2, but its numbers changed between sessions, so the time saved went into checking its errors.
€8,354
swing on one subcontractor's total when a generative AI tool was given the same bids in a new session
Up to a 28% difference on a single line, for identical input. A tool that does not repeat itself cannot give the reasons a public award needs. That is why Dundir's design computes the comparison with an optimisation engine and keeps the language model to reading documents.
Hamppi 2025, Aalto University / Skanska Talonrakennus Oy: Table 6, p. 62; test timings p. 64.
The design, in three steps
Bids are put on one line
The language layer is meant to read bids as they arrive (PDF, email, spreadsheet) and put unit prices, delivery times, certificates and assumptions side by side. A buyer checks what it read.
The engine computes
Your hard requirements (certificates, delivery window, budget ceiling) bound the choice; your weights for price, reliability and sustainability rank what is left. Same input, same answer.
You decide, the record is written
You see the ranking with its source data and the rule that produced it. You decide, and the decision is recorded with its reasons, in the form a public buyer has to give them (Aanbestedingswet 2012, art. 2.130).
What the pilot has to show
Time
Whether the comparison of your closed tenders takes less time than your current process, checking included.
Reproducibility
Whether the same bids give the same result on every run, across buyers and over time.
The record
Whether the reasoning written at award is good enough for your internal approval and for an audit.
What we do not claim yet
That the engine picks better than your best buyer. That is what the pilot measures.
Procurement scan
Calculate your procurement cost
Five minutes. No registration. Straight away, your own estimate of the time and money your organisation loses to manual procurement.