The verdict in three sentences
A health insurer with 160,000 members that receives 45,000 letters and emails a month and sorts them by hand ties up 9 staff on a task that AI classification handles at 85 to 92 % accuracy. The useful project combines scanning, classification, document data extraction (quotes, invoices, bank details, certificates) and routing to the right claims queue, with a human validating uncertain cases. At Kolonell, this kind of project ranges from 3,810 to 22,870+ EUR (2,500,000 to 15,000,000+ FCFA) depending on complexity, delivered in 3 to 4 months, with EU hosting and options set in the quote.
What manual sorting costs today
A health insurer's inbound flow mixes very different requests: reimbursement claims with optical or dental invoices, pre-treatment quotes, bank detail changes, enrolments and cancellations, complaints, letters from healthcare providers. Each item must land in the right queue with the right member number, or processing slows and complaints pile up.
| Indicator (2026 estimate) | Manual sorting | AI-assisted sorting |
|---|---|---|
| Monthly volume (letters + emails) | 45,000 | 45,000 |
| Staff dedicated to sorting and indexing | 9 FTE | 2 to 3 FTE (checks and complex cases) |
| Time from receipt to claims queue | 2 to 4 days | under 2 hours |
| Misrouting rate | 6 to 8 % | 3 to 5 % after human review |
| Documents classified automatically | 0 % | 70 to 80 % (high confidence threshold) |
| Loaded cost of a sorting clerk | about 38,000 EUR a year (in the UK roughly 32,000 GBP) | redeployed to claims |
Accuracy of 85 to 92 % depends on scan quality and the number of categories. The system only auto-classifies documents whose confidence score exceeds a threshold (for example 95 %); the rest go to human review. That mechanism is what makes the gain real without hurting quality.
Architecture and compliance
| Component | Role | Watch point |
|---|---|---|
| Ingestion | mailboxes, mailroom scanner, member portal | de-duplicating repeat submissions |
| OCR and reading | turning scans into text, reading invoice tables | quality of phone photos |
| Classification | 20 to 40 business categories | training set of 3,000 to 5,000 labelled documents |
| Extraction | member number, amounts, IBAN, procedures, dates | consistency checks against the member base |
| Routing | push to claims queue or existing tool | connector with the insurer's core system |
| Supervision | dashboard, quality sampling, audit log | traceability required by the AI Act and GDPR |
Health data requires certified health data hosting (HDS in France, equivalent controls elsewhere in the EU), and the European AI Act demands transparency, human oversight and logging for this kind of system. The model choice (EU-hosted model or model deployed on the insurer's infrastructure) is made at scoping. The AI never decides a reimbursement: it sorts and prepares, the claims handler decides.
Budget and timeline
Simple email classification into about ten categories, without extraction, stays in the simple bracket: 3,810 to 6,100 EUR. Adding OCR, document extraction and a supervision dashboard leads to the medium bracket (6,100 to 10,670 EUR). Two-way integration with the core system, 40 categories and continuous retraining belong to the complex bracket (10,670 to 22,870+ EUR).
The 3 to 4 month schedule breaks down as follows: 3 weeks of scoping and building the labelled set, 6 weeks of development and training, 3 weeks running in parallel with manual sorting to measure accuracy, then gradual switchover category by category. Application maintenance then costs 53, 114 or 183 EUR per month depending on service level, excluding hosting and model usage costs, estimated in the quote based on volume.
Mini case study
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Nathalie, head of operations at a health insurer in Bordeaux, starts from 9 sorting clerks at 38,000 EUR loaded cost each, or 342,000 EUR a year. With 75 % of documents auto-classified, she keeps 3 clerks for validation and complex cases and moves 6 people to claims handling, where her team has a backlog. The freed capacity is 6 x 38,000, or 228,000 EUR a year of reallocated time.
Processing time drops by 2 to 3 days, which cuts member chasers. If 15 % of the 45,000 monthly contacts were delay-related chasers and half of them disappear, the insurer avoids around 3,400 contacts a month. For a complex-bracket project around 18,000 EUR plus running costs, payback is measured in months.
FAQ
What accuracy can we expect at launch?
Expect 85 % on the main categories during the parallel phase, then 90 to 92 % after 2 to 3 months of feeding corrections back. Cases below the confidence threshold are always checked by a clerk.
Does health data leave the European Union?
No. The architecture uses certified hosting in the EU and an AI model hosted in the EU or on your own infrastructure.
Do we need to change our core system?
No. The AI drops classified documents and extracted data into existing queues through an API or scheduled import.
What happens to the sorting team?
In most projects, 2 to 3 people stay on quality control and the others join claims or member services, where workload grows with new contracts.
Let's scope your project. Classification, document extraction and routing into your claims queues, with compliant hosting: we define scope and timeline with your operations team. 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.
