The verdict in three sentences
An AI classifier for inbound mail pays off for an insurer from about 20,000 items a month, provided you first target 70 to 80% automatic routing on the simple categories and leave the rest to people. The real work is not the model but the quality of the 30 categories, per-category confidence and the manual fallback queue. With health data kept in EU, certified hosting, a DPIA and full audit logging, the project runs in 3 to 5 months.
Starting point: 45,000 items and 9 people sorting
The operations director of a 120,000-member health insurer in Amsterdam handles a mixed flow: scanned paper letters, emails, web forms and attachments (dental quotes, hospital invoices, certificates). Nine people read each item, pick one of 30 categories and forward it. Average time before pickup exceeds 3 working days at peak (January renewals, autumn switching season).
| Indicator | Today | Target with AI |
|---|---|---|
| Monthly volume | 45,000 items | 45,000 items |
| Items sorted by hand | 100% | 20 to 30% |
| Time before pickup | 2 to 3 working days | under 4 hours for 80% |
| FTEs on sorting | 9 | 2 to 3 (review and fallback) |
| Routing error rate | 6 to 8% (estimate) | 3 to 4% on automated flows |
| Categories handled | 30 | 30, 18 automated at launch |
The 6 freed FTEs move to complaints and complex claims, where the insurer actually loses members during switching season.
Architecture and compliance: GDPR and the EU AI Act
A serious build chains four blocks: capture (OCR for letters, mailbox connector), data extraction (member number, claim type, amount), classification with a confidence score, then routing into the document management system or claims CRM. Anything below the confidence threshold lands in a human review queue.
The items contain health data, a special category under GDPR article 9. In the Netherlands that means a DPIA, EU-based hosting with strong certification (ISO 27001, NEN 7510 for healthcare), and a contractual ban on the model vendor reusing data for training. The EU AI Act treats internal triage with human oversight as limited risk, but you must document data, performance and supervision.
| Technical option | Indicative 2026 cost | Strengths | Limits |
|---|---|---|---|
| Off-the-shelf capture software (licence) | 30,000 to 80,000 EUR a year, order of magnitude | fast to deploy | rigid categories, weak on 30 business classes |
| General-purpose cloud AI platform | 0.01 to 0.05 EUR per item + integration | powerful models | check EU hosting and no-training clauses |
| Model on dedicated NEN 7510 hosting | 1,500 to 4,000 EUR a month, order of magnitude | full data control | needs an ops team |
| Kolonell custom build | 2,500,000 to 15,000,000+ FCFA (3,810 to 22,870+ EUR) depending on complexity | fitted to your 30 categories and tools | scope must be tight |
| Kolonell maintenance | 35,000 / 75,000 / 120,000 FCFA a month (53 / 114 / 183 EUR) | fixes, monitoring | advanced retraining quoted separately |
For this case a medium to complex project is realistic: DMS connectors, a per-category confidence dashboard, fallback queue and audit log. Hosting and third-party licences are quoted separately. See our AI services and cybersecurity.
Rollout in four steps
- Flow audit (3 weeks): pull 5,000 real, pseudonymised items, review the 30 categories, merge ambiguous ones.
- Prototype (4 to 6 weeks): classify a labelled sample, measure precision per category, set thresholds.
- Shadow mode (4 weeks): the AI suggests, people decide. You compare with zero risk.
- Progressive switch-over: categories above 95% precision go automatic, the others stay assisted.
That gradual approach is what separates a useful system from one abandoned after six months.
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Mini case study
Sanne, operations director at the insurer, runs the numbers. Loaded cost of a sorting FTE in the Netherlands: 50,000 EUR a year. Freeing 6 of 9 FTEs equals 300,000 EUR a year of redeployed capacity. Against that: a complex build at 22,870 EUR, dedicated hosting at roughly 2,500 EUR a month (30,000 EUR a year, order of magnitude) and maintenance at 183 EUR a month (about 2,200 EUR a year). First-year cost: about 55,000 EUR. Even if only half the gain materialises in year one (150,000 EUR), payback comes in about 4.5 months. The less visible benefit, pickup time cut from 3 days to 4 hours, also reduces follow-up calls.
FAQ
What automatic routing rate should we target first?
Aim for 60 to 70% in the first three months, then 80% once categories settle. Beyond 85%, errors often cost more than the time saved.
Do we need certified hosting for a triage AI?
Yes, as soon as items contain health data, which is most insurer mail. The model vendor must also commit by contract not to reuse your data, and hosting should stay in the EU.
How many labelled items do we need to start?
Count 150 to 300 examples per category, so 5,000 to 9,000 items for 30 categories. Part of that can come from your DMS history.
Does the EU AI Act block this project?
No, internal triage with human oversight is limited risk. You still need to document performance, keep logs for at least 12 months and train staff.
How soon will we see measurable results?
The prototype gives precision figures within 6 to 9 weeks. The first categories usually go live in month 4.
Let's scope your project. We map your categories, connectors and hosting constraints, then deliver a measurable prototype within weeks. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.
Mohamed Bah
Fondateur, Kolonell
Passionate about digital and entrepreneurship in Africa, Mohamed has been helping Sénégalese businesses with their digital transformation since 2020. Founder of Kolonell, he believes every SME deserves a professional and accessible online présence.
