What we heard in our conversations
In at least five conversations with planning firms, the same topic came up: bid work eats time nobody can bill for. The process has grown over the years and looks almost the same everywhere. Several people share the steps: screening and sorting notices on the procurement portals, writing bid documents, picking staff by competency, contacting partner firms.
Depending on firm size, the owners named anything from several hours to a good 10 hours a day. At the upper end, that's more than one full-time role spent on bidding instead of planning.
Another thing almost every firm mentioned: they had already tried keyword filters and tender tools. Those don't know which projects fit a firm, so they return lots of notices that don't fit, and a person has to sort them out again.
Architecture
The agent doesn't use a generic keyword filter; it works with the firm's own profile. For that, we index:
- Indexing references: the firm's references are tagged by practice area, contract size, project phase and regional distribution.
- Capturing competencies: staff competencies come from the personnel database: specialties, software knowledge, certifications, typical project sizes.
- Partner firms: regular partners are tagged by practice area, along with the track record of past collaboration.
- Target contract sizes: for each practice area, the firm sets the contract size below which a bid isn't worth pursuing.
The agent (Claude Sonnet) sorts incoming notices from the procurement portals against this firm profile. For hits with a high rating, it drafts the first bid straight from the tender documents. Hits with a medium rating land on a review list for a single bid manager. If a bid moves on to the next round, the same agent creates the presentation using the same criteria from the firm profile.
Compliance: the agent only does the groundwork. Self-declarations such as the ESPD (European Single Procurement Document) and running the procedure in a legally compliant way stay in the firm's existing workflow. VgV, GWB and HOAI risk remains entirely in the firm's hands.
Model calculation
The projection starts from the upper end of the figures planning firms gave us (10 staff hours/day on bid work) and is validated against the notices the prototype processed:
- 10 staff hours/day → 30 minutes of review in bid work (factor of 20)
- Built into the architecture: once the sorting bottleneck is gone, the firm can handle more bids without adding staff.
- Built into the architecture: competency matching becomes more consistent because the agent uses an indexed staff database instead of manual assignment.
The time freed up goes back into actual planning work, with the same people.
What we take from it
A tender AI that knows nothing about the firm can only return general hits. What matters is not the model but indexing the firm profile. Without that step, even a strong model like Claude Sonnet only returns average hits.
That is exactly what we built: a working prototype that indexes the firm profile and runs against live notices. For your firm, it is the starting point for a 10-day pilot, not a concept on paper.
Methodology in detail: assumptions and math
To keep the factor of 20 traceable, here is the math for a generic planning firm. It builds on the figures from our conversations with planning firms and the target processing time we validated in the fully built prototype.
Typical industry assumptions
- Volume on the procurement portals (EU and Germany): we assume 800–1,200 relevant notices per week across all practice areas.
- Hit rate after indexing the firm profile: typically 0.5–2 % of incoming notices are real matches for a specialized firm. The rest fits the practice area, but not the specific profile.
- Manual screening effort without pre-sorting: roughly 3 min per notice × 200 notices/day = 600 min/day = 10 staff hours/day (upper end of the figures from our conversations).
- HOAI fee zones: specialized firms are mostly interested in contracts in the upper fee zones IV and V, where fee rates are higher.
- Procurement law: self-declarations such as the ESPD stay in the firm's existing workflow. The AI makes no procurement-law assessment. VgV, GWB and HOAI risk stays in the firm's hands.
The math
- Factor of 20: 600 min/day (upper end from our conversations) / 30 min/day (target processing time validated in the prototype for reviewing the filtered top hits) = 20×. We measured the target time in the fully built prototype against real notices; it is not modelled.
- In headcount terms: at 8 hours per working day, 10 staff hours equal roughly 1.25 full-time roles in bid work. Bid work earns no fees, planning does. Every hour that moves from bidding into planning becomes billable.
What moves the numbers
- Smaller firms save fewer hours in absolute terms; the factor stays similar.
- In highly specialized firms (e.g. structural engineering only, HOAI zone V), filtering can be even more selective. The factor can rise to 30–40×, because the hit rate is below 1 % and manual screening is correspondingly less efficient.
