Auto-monitoring of public tenders for a small marketing agency

§01Challenge

A 5-person marketing agency specialising in the public sector (local government units, cultural institutions, education). Annual revenue ~€800k, 70% from won tenders. Market monitoring — 4 portals (eZamówienia, BIPs of individual units, sectoral portals) — was critical but too time-consuming.

Process in 2024: one employee spent 1.5h per day on manual review of new announcements. Annual: ~330 hours = ~€11k cost. Plus a constant dose of stress over "did we miss something".

Critical incident in 2024: the agency missed a large tender (€85k value) because the announcement appeared on Friday evening on a BIP they did not usually check at weekends. Submission deadline was the following Tuesday — without weekend work they would not have made it.

§02Approach

Setup in 3 weeks. 4 portals parsed every 30 minutes. Every new announcement passes through AI pre-screening (Claude Haiku) which scores "is this for us" based on:

- industry + scope of work matching their competencies

- budget range (filters too small and too large)

- submission deadline (filters too urgent)

- contracting entity (filters excluded clients)

Output: 3 Slack buckets — #tenders-priority (match all criteria, manual review within 4h), #tenders-watch (partial match, daily review), #tenders-skip (auto-rejected, available for audit).

Stack: single VPS, Playwright for scrape, Claude Haiku API (~€8/month in tokens), PostgreSQL, Slack integration. Total tech cost: ~€50/month. Rest is retainer.

§03Outcome

Employee monitoring time: 1.5h/day → 0.3h/day (mostly priority bucket review). 1.2h/day × 5 days × 4 weeks = 24h/month saved. At €33/hour = ~€800/month time value, system costs €450/month.

3 tenders won from signals previously missed (weekend announcements, niche portal they "did not get to" check). Combined revenue: ~€140k in the first 5 months. That is the main project value — the time calculation is secondary.

Stress reduction: the owner reports the team stopped worrying about "what are we missing". Manual review now applies only to AI-flagged opportunities, every one is a real decision.

System operating 5 months. AI prompts iterated 3 times in the first month (client narrowed definition of "for us" as they used the system), then stable. Zero false negatives reported (client randomly samples the #tenders-skip archive weekly, in 10 reviews: 0 missed opportunities).

Stack

Metrics

• 4 — Portals monitored
• ~800 — Announcements screened / month
• 30h / month — Time saved
• +3 — Tenders won
• ~€140k — Revenue from signals
• €450 / month — Operating cost