The verdict in three sentences
With 1,800 emails a week sorted by hand, an insurer's customer service answers in 48 hours on average and lets urgent claims slip among certificate requests. An AI email classification system, at 15,000 to 30,000 USD in Nairobi (6 to 12 million FCFA for the same scope in Abidjan) plus 80 to 250 USD a month of API fees, routes 85% of messages automatically, extracts the policy number and answers simple requests on its own. First response falls to 6 hours, and the rollout takes 2 months.
What manual triage costs
Three agents spend their mornings opening the shared inbox, reading each message, looking up the policy in the core system and forwarding to the motor claims, medical, certificates or complaints teams. Attachments (accident reports, hospital invoices, photos) are renamed by hand. The Insurance Regulatory Authority (IRA) monitors complaint handling times, and an overdue complaint hurts the insurer's standing.
| Indicator (2026 order of magnitude) | Manual triage | AI triage |
|---|---|---|
| Emails received per week | 1,800 | 1,800 |
| Share routed automatically | 0% | 85% |
| First response time | 48 h | 6 h |
| Certificate requests handled without an agent | 0% | 60 to 70% |
| Urgent claims spotted same day | 55 to 65% | 95% |
| Agents assigned to triage | 3 | 0.5 (reviewing doubtful cases) |
| Routing errors | 8 to 12% | 3 to 5% |
Architecture and data confidentiality
The legal team's main concern: policyholder data must not be used to train a third-party model. The chosen architecture keeps emails and attachments inside the insurer's infrastructure. Only the text needed for classification, after masking ID and phone numbers, is sent to a language model API under a contractual zero-retention option. The setup complies with Kenya's Data Protection Act 2019 and is registered with the Office of the Data Protection Commissioner (ODPC).
| Component | Budget (USD) | Timeline |
|---|---|---|
| Mailbox connector (Microsoft 365, Google Workspace, Exchange) | 2,000 - 3,800 | 1 week |
| Personal data masking | 1,800 - 3,500 | 1 week |
| Classification and extraction engine (policy, claim, amount) | 4,500 - 8,800 | 3 weeks |
| Automated replies to simple requests | 2,200 - 5,000 | 2 weeks |
| Integration with the policy administration system | 3,000 - 6,200 | 2 weeks |
| Volume and response time dashboard | 1,500 - 2,700 | 1 week |
| Total | 15,000 - 30,000 | 2 months |
API cost scales with volume: 7,800 emails a month come to 80 to 250 USD depending on message length and model (50,000 to 150,000 FCFA in the Abidjan scenario). Hosting the connector and dashboard adds 70 to 150 USD a month, and maintenance 15% of the initial budget per year.
How the AI learns your categories
The project starts with 3,000 historical emails labelled in one week by two experienced agents: 12 categories (motor claim, medical claim, certificate, endorsement, cancellation, complaint, and so on) plus an urgency level. The system first runs in suggestion mode for 3 weeks: the agent approves or corrects. Once accuracy passes 90% on a category, its routing switches to automatic. Low-confidence messages stay in human review.
Mini case study
Grace, head of customer experience at an insurer in Nairobi, assigns 3 agents to triage at a loaded cost of 900 USD a month each. After rollout, 2.5 roles move to claims handling, worth 27,000 USD a year. She invests 24,000 USD, plus 350 USD a month of API and hosting (4,200 USD a year). Net annual gain: about 22,800 USD. Payback: about 13 months, before counting fewer complaints about delays.
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
FAQ
Can the AI read attachments like accident reports or invoices?
Yes, text recognition extracts key fields from a PDF or photo. Budget an extra 2,500 to 5,000 USD to cover accident reports and hospital invoices.
What happens when the AI is unsure?
Below a confidence threshold set at 80%, the message goes to human review. In steady state, that is 10 to 15% of emails.
Does policyholder data leave the country?
Emails stay in your infrastructure; only masked text goes to the API. A locally hosted model option is possible for an extra 20,000 to 38,000 USD.
Does it handle emails in Sheng, Swahili or with typos?
Yes, recent models handle English, Swahili and informal mixes well. In our tests, accuracy drops by only 2 to 4 points on such messages.
How long before triage is reliable?
After 3 weeks in suggestion mode, accuracy typically exceeds 90% on the 8 main categories. Full rollout fits in 2 months.
Let's scope your project. Share your email volume, categories and core system, and we will price an AI triage between 15,000 and 30,000 USD, live in 2 months. 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.
