The Challenge
CIRFOOD's bid team had a math problem. Leadership wanted 50% more tenders submitted this year, the same team, no drop in win rate.
The work that stood in the way wasn't writing or strategy. It was the hours every bid manager lost upfront, sorting through 12 to 22 documents per dossier in mixed PDF, Excel, and Word formats, then re-keying 72 CRM fields and a few hundred calculation inputs into a spreadsheet by hand.
That pre-work was eating the time that actually wins tenders.
The questions leadership couldn't answer:
How do we add 15 more tenders a year without adding headcount?
Why is the most expensive part of our bid process the part with the lowest strategic value?
If we hand this to a generic AI tool, who owns the data, the prompts, and the workflow six months from now?
And how do we move fast without breaking a calculation model the business has trusted for years?
Without an answer, the 50% growth target was a hiring conversation. Either the team grew, or the win rate fell. CIRFOOD didn't want to choose.
The Approach
We started by refusing to scope a tool. CIRFOOD asked for capacity, not software, so we treated the engagement as discovery first, build second.
Step 1. Map the actual work. We sat down with bid managers and walked through three live tenders (Spaarnelanden, Amsterdam VU, Maastricht UM) field by field. The output was a complete inventory: 72 CRM fields, the documents each one lives in, the time each one takes to extract manually, and the failure modes when a tender package is incomplete or formatted unusually.
Step 2. Find the calculation bottleneck. The Aanneemsom calculation isn't just a number entry job. It's a transformation: gross wages from the previous caterer's Bijlage 9 (Personeelsovernamelijst) become all-in hourly rates inside CIRFOOD's own model, with SROI handling, scale parsing, and roster defaults applied per role. Every caterer hands over Bijlage 9 in a different column layout. The bottleneck wasn't the math. It was reconciling shape-shifting input formats with a fixed calculation model.
Step 3. Pressure-test the risk. CIRFOOD's calculation workbook contains roughly 230 formula cells the business has trusted for years. Any tool that touches it has to guarantee, mechanically, that those formulas stay intact. We designed the safety layer (formula protection with hard-fail on mismatch) before we wrote the parser.
Step 4. Define the phasing. Discovery made the phased roadmap obvious. Fase 1 is the Analyst (extraction and calculation). Fase 2 is the Schrijf Assistent (draft writing from prior tenders). Fase 3 is the Strateeg (benchmarking, win-rate analysis). Each phase compounds on the data the previous one produces.
The discovery didn't just scope Fase 1. It made it possible to commit to a 12-month roadmap without guessing.
The Solution
What we built is an AI Tender Agent that grows from assistant to sparring partner. Fase 1 is in CIRFOOD's hands now, with two production tools.
CRM Tender Extraction Tool.
A web interface where a bid manager uploads a ZIP of the full tender package, hits run, and gets back an Excel with 72 structured CRM fields, metadata, and a validation tab. Every field carries a confidence score from 0 to 1, the source document, the page number, and the exact citation. Powered by Claude Sonnet 4.5, deterministic, running on CIRFOOD's own GCP project.
Aanneemsom Calculator.
A tool that combines the extraction output with the incumbent caterer's Bijlage 9 and fills the existing calculation workbook. It populates roughly 80 input cells and never touches the 230 formula cells. If the formula protection check fails after a run, the program halts hard. Approximated all-in rates are flagged for manual review on a separate validation tab.
The cost story is where this stops being a productivity tool and starts being infrastructure.
Approach | Cost per tender | Cost at 45 tenders/year |
|---|---|---|
Manual bid manager hours (3-4 hrs at €75/hr loaded) | €225-300 | €10,000-13,500 |
Owned stack (Claude API + GCP) | €0.55 | €26 |
That's not 30% cheaper. It's roughly 400 times cheaper per tender, and it scales the right direction: a second tender costs the same minutes as the first.
But the cost story is the surface. The deeper point is ownership. The tool runs in a GCP project CIRFOOD owns. The prompts, the extracted data, and the configuration belong to CIRFOOD. The architecture is stateless and reproducible. Every run is auditable. If What's Next disappeared tomorrow, CIRFOOD would still have a working asset, not a dead subscription.
The human-in-the-loop design is intentional, not a limitation. Processing only starts on manual command. Low-confidence fields get a warning. Approximated rates get flagged. The bid manager remains the decision-maker. That's the difference between an agent that replaces judgment and one that returns hours.
The Results
On three measured tenders, the extraction tool delivered 85% of CRM fields in an average of 128 seconds, at an average LLM cost of $0.58 per tender.
The pre-work that consumed several hours per bid manager now completes in roughly two minutes with an explainable trail per field.
Operationally, the bid team's day has changed. The first hours of every tender used to be sorting and data entry. Now those hours are strategic work: pricing approach, win themes, response narrative. The calculation model the business has trusted for years stays exactly as it was, with the input layer automated and the formulas mechanically protected.
Strategically, the 50% growth target moved from a hiring conversation to an execution one. Leadership can now plan tender volume against bid manager capacity instead of against bid manager hours. And Fase 2 and Fase 3 turn the same dataset into writing acceleration and win-pattern intelligence, compounding the value of the work already done.
What differentiates them is not only their technical expertise, but also their deep understanding of AI governance and its pitfalls, ensuring that innovation is implemented in a controlled, compliant, and sustainable way. This is not a one-off project, but a long-term partnership built on trust, professionalism, and measurable outcomes.
Stefano — CIRFOOD Netherlands
Total Fase 1 timeline: discovery to MVP in roughly five months. The discovery half wasn't optional. It's the reason the build half didn't break anything.
Key Results
cheaper per tender vs. manual processing
per tender, down from 3-4 hours
more tenders submitted, same team
