TL;DR: in 2026, a well-scoped AI customer service agent resolves 40-60% of requests from day one, and the best deployments exceed 75%; the non-negotiable condition is keeping a visible path to a human. For a small business, a first agent on one well-defined process starts as a sprint from €2,500 (excl. VAT), the rest on written quote.
The paradox of the year: AI agent resolution rates have never been higher, yet Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027. Both facts are true at the same time, and that is exactly what makes the question interesting for a business owner: what works is known and documented, but rarely applied.
Key takeaways
- The average resolution rate of an AI customer service agent sits between 40% and 60% at initial deployment, and climbs above 60% after 6 to 12 months of optimization, according to the 2026 benchmarks published by Fin, Intercom's AI agent.
- Fin reports a 76% average resolution rate across 12,000 customers, at $0.99 per resolution, while a human-handled conversation costs $6 to $12.
- Klarna's AI assistant handled 2.3 million conversations as early as February 2024, the equivalent of 700 full-time agents, then the company started rehiring humans in May 2025, reports CX Dive.
- Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls, per its June 25, 2025 press release.
What does an AI customer service agent actually resolve in 2026?
Between 40% and 60% of requests at initial deployment: that is the average range given by Fin's 2026 benchmarks for an AI customer service agent properly connected to the company's documentation. After 6 to 12 months of tuning against real customer questions, the average moves above 60%. Fin, Intercom's agent, claims a 76% average resolution rate across a base of 12,000 customers, with its best deployments above 84%.
One piece of vocabulary is worth settling upfront, because vendors keep the confusion alive: deflection counts the tickets that never reach a human; resolution counts the requests actually settled, from the customer's point of view. An agent that discourages a customer without answering their question is deflecting, not resolving. The only number that matters for your P&L is the second one.
The economics are easy to lay out. Fin charges $0.99 per resolution; a human-handled conversation costs between $6 and $12 according to the same source. On a volume of a few hundred repetitive tickets per month, the gap shows quickly. But that math only holds if resolved requests are genuinely resolved: a resolution rate inflated by deflection gets paid for later, in lost customers.
The Klarna lesson: why all-in AI walked it back
Klarna, the Swedish buy-now-pay-later fintech, remains the most instructive case study on the market. In February 2024, the company announced that its AI assistant, built with OpenAI, had handled 2.3 million conversations in one month, the workload equivalent of 700 full-time agents. Response times improved by 82%, average resolution dropped below 2 minutes. On paper, the demonstration was total.
Fifteen months later, in May 2025, Klarna was rehiring humans for its customer service, reports CX Dive. CEO Sebastian Siemiatkowski publicly acknowledged that quality had degraded on complex cases, and now considers it essential that every customer knows they can always reach a human if they want to. Forbes summed up the episode in one line: customers like talking to people.
The lazy reading of this episode concludes that AI does not work. The useful reading notices something else: Klarna's AI still handles two-thirds of requests. What failed was not the agent, it was the replacement. Using AI instead of humans degrades quality; using it alongside humans, on the repetitive volume, with a visible escalation path to a person, is exactly the model that holds. That is the distinction every small business should keep in mind before signing anything.
What works for a small or mid-sized business?
Four conditions come up in every deployment that lasts, and they depend neither on company size nor on the tool you pick.
A tight scope. The agent starts with the 20 or 30 questions that come up most: opening hours, delivery times, order tracking, return conditions, pricing. Not with the promise of handling everything. A narrow, well-covered scope beats a broad, mediocre one, on satisfaction as well as on the numbers.
Answers drawn from your real data. A useful agent pulls from your actual documentation (the technique known as RAG, retrieval-augmented generation), cites its sources and says when it does not know. An often-forgotten corollary: if your documentation is wrong or outdated, the agent will recite it with confidence. The first workstream of an agent project is often updating the knowledge base.
A visible human escalation path. That is the Klarna lesson. The customer must be able to reach a person, and know it. Sensitive cases (complaints, disputes, cancellations) go to a human by default, with the context already assembled by the agent.
Honest measurement. You track real resolution, confirmed on the customer's side, not the number of deflected tickets. That figure is what decides whether to extend the scope or fix things.
The table below sums up the three approaches available to a small business, with entry costs and the right use case for each.
| Approach | Entry cost | What it handles well | The right choice when |
|---|---|---|---|
| Scripted chatbot | A few dozen euros per month | Closed questions: hours, tracking, returns | Low volume, highly repetitive questions |
| AI module of a support platform (Intercom, Zendesk...) | Subscription, then about $1 per resolution | Rich FAQ, high volume of written tickets | You already use the platform and your answers are documented |
| Custom AI agent | Sprint from €2,500 (excl. VAT), then on quote | Requests that require your business data (CRM, orders, contracts) and actions in your tools | Heterogeneous tools, strong GDPR constraints, need to act and not just answer |
The middle row deserves honesty: if you already use Intercom or Zendesk and your canned answers are clean, your platform's AI module is probably your best first step. Custom becomes relevant when the agent must query your CRM, check an order, trigger an action, or when your data constraints rule out US platforms.
When an AI agent is a bad idea
This needs saying as clearly as the rest, because this is where the 40% of canceled projects announced by Gartner get manufactured: runaway costs, business value never defined, poorly controlled risks.
If your support desk receives a handful of requests per day, a well-written FAQ page and three canned replies will do better than an agent, at zero cost. If your internal documentation is scattered or outdated, fix that first: an agent plugged into wrong sources industrializes the error. And if most of your requests carry a heavy emotional or regulatory load, AI must stay in an assist role behind your team, never on the front line. In those three situations, the honest answer is: keep your current setup, or improve your documents before automating anything.
What I see in the field
Two concrete examples of this logic, from my own projects. On mathieuhaye.fr, the Chat with my AI page is a public agent connected to the site's content: it answers within a bounded scope, cites its source pages, and points to a call as soon as a question falls outside its field. It is exactly the architecture described above, at a smaller scale. And in the Emma CRM of France's 3018 hotline, the national number against cyberbullying, the integrated sovereign AI assists the 12 counselors and managers handling 5 channels; it never answers in their place. On sensitive human situations, AI prepares, the human decides. The full story is in the 3018 case study.
Frequently asked questions
What resolution rate should you expect from an AI customer service agent?
The 2026 benchmarks put the average at 40-60% of requests resolved at initial deployment, and above 60% after 6 to 12 months of optimization. The best deployments exceed 75%. The right metric is real resolution, confirmed by the customer, not mere ticket deflection.
Can an AI agent replace an entire support team?
No, and Klarna proved it: after handing two-thirds of its conversations to AI in 2024, the company started rehiring humans in May 2025, as quality had degraded on complex cases. The model that holds up is hybrid: AI absorbs the repetitive volume, a human stays reachable for sensitive cases.
How much does an AI customer service agent cost for a small business?
The AI module of a mainstream support platform is often billed per resolution, around $1 per resolved request with the major vendors. A custom agent, connected to your data and tools, starts as a sprint from €2,500 (excl. VAT) for a first process, the rest on written quote, plus model usage billed on consumption.
Is an AI customer service agent compatible with GDPR?
Yes, provided it is designed that way: hosting in France or the European Union, an agent limited to strictly necessary data, explicitly authorized sources, auditable logs and human validation on sensitive cases. These choices are made at scoping time, not after going live.
The question to ask is therefore not "can AI run my customer service?" but "which 20 questions is my team tired of repeating?". If the list fills up in five minutes, you have the scope of your first agent. You can check it against what the AI agent for SMEs page describes, or talk it through directly: a 30-minute call is enough to tell whether your case justifies an agent, a module of your current platform, or simply a better-written FAQ.