The verdict in three sentences
When a customer service team spends 40% of its time reading, classifying and routing 1,200 emails a week, generative AI delivers a fast, measurable return. An AI triage agent at 8,000 to 20,000 EUR excl. VAT, with 50 to 300 EUR a month in API costs, reaches around 92% classification accuracy in practice and drafts replies that an agent approves before sending. With EU hosting and a clear scope, the project ships in 6 weeks and pays back in under 6 months.
What the agent really does, and what it does not
The agent reads every incoming email, extracts the intent (order tracking, return, invoice, complaint, quote request), identifies the customer and order in your ERP or CRM, assigns a priority and routes the message to the right queue. For simple requests it drafts a reply using real order data.
| Task | Automatable | Confidence level 2026 | Human validation |
|---|---|---|---|
| Classification by reason | Yes | 90 to 95% | Sample checks |
| Order and customer number extraction | Yes | 95 to 98% | No |
| Prioritisation (urgent, dispute, VIP) | Yes | 85 to 92% | Yes for disputes |
| Order tracking draft | Yes | 90% usable without edits | Yes, 1 click |
| Return or refund draft | Partly | 70 to 80% | Yes, mandatory |
| Sensitive or legal complaint | No | not relevant | Senior agent |
| Automatic sending without review | Not advised | not relevant | Always |
The rule we recommend: the agent classifies and proposes, the human decides and sends. That keeps service quality high and limits risk when the model gets it wrong.
Price and recurring costs in 2026
| Item | Content | Indicative cost (excl. VAT) |
|---|---|---|
| Scoping and taxonomy | Reasons, priority rules, 500 labelled emails | 1,500 to 3,000 EUR |
| Mailbox connector | Microsoft 365, Gmail, Zendesk or Freshdesk | 1,500 to 3,500 EUR |
| ERP or CRM integration | Read orders, invoices, customer history | 2,000 to 6,000 EUR |
| Classification and drafting agent | Prompts, guardrails, tests on history | 2,500 to 5,500 EUR |
| Monitoring dashboard | Volumes, accuracy, response times | 500 to 2,000 EUR |
| Model API | Depending on volume and model | 50 to 300 EUR a month |
| EU hosting and maintenance | Server, logs, updates | 150 to 400 EUR a month |
A project limited to classification on a Microsoft 365 mailbox starts at 8,000 EUR excl. VAT. With ERP integration, drafts and several languages (Dutch, English, German), budget 15,000 to 20,000 EUR.
Limits and GDPR
Emails contain personal data: name, address, sometimes bank details. The model must be called through a provider offering processing in the EU, with no reuse of data for training and a data processing agreement compliant with GDPR article 28. Logs are kept 30 to 90 days and sensitive attachments are masked before reaching the model. The Dutch regulator, the Autoriteit Persoonsgegevens, expects this to be documented.
Other limits: accuracy drops on very short or poorly written emails, long threads and rare languages. Plan 2 to 4 weeks of tuning after go-live and a monthly review of 100 randomly sampled emails.
Mini case study
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Sarah, customer service manager at an electrical equipment distributor in Amsterdam, leads 4 agents who receive 1,200 emails a week. Triage takes 40% of their time, or 1.6 full-time equivalents. The AI agent cuts that time by 70%, freeing 1.1 FTE, about 44,000 EUR a year at a loaded cost of 40,000 EUR. Project: 14,000 EUR excl. VAT plus 450 EUR a month for API, hosting and maintenance (5,400 EUR a year). Counting only half the gain (22,000 EUR), the investment pays back in 10 months; average first response time falls from 26 hours to 6 hours.
FAQ
How much does an AI email triage agent cost for an SME?
Budget 8,000 to 20,000 EUR excl. VAT for setup depending on integrations, then 200 to 700 EUR a month for API, hosting and maintenance. Timeline is 6 weeks.
What classification accuracy should we expect?
Around 92% after tuning, on a taxonomy of 8 to 15 reasons. Above 25 reasons accuracy drops and it is better to merge categories.
Can the agent reply to customers on its own?
Technically yes, but we advise against it beyond acknowledgements. One-click approved drafts keep 80 to 90% of the time saved without the risk of a wrong answer.
Is it GDPR compliant?
Yes if the model is hosted or processed in the EU, with no reuse for training and a processing agreement in place. Sensitive data is masked before processing and logs are purged within 90 days.
Do we need to change our ticketing tool?
No, the agent plugs into Microsoft 365, Gmail, Zendesk or Freshdesk. It adds tags and drafts inside the tool your agents already use.
Let's scope your project. Share your email volume, contact reasons and tools and we will price the scope, the budget and a 6-week go-live. 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.
