AI Customer Service Agents Finish About Half the Questions. The ROI Lives in the Other Half.

Shop for an AI customer service agent and the brochures quote resolution rates of 70%, 80%, 85%. Two named companies published something more useful in the past year: their own split of the work. CMA CGM, one of the world's largest container shipping groups, said on 14 October 2026 that its AI agents now resolve more than 46% of the cargo-release enquiries they handle. Salesforce's chief executive, Marc Benioff, said in 2025 that about half of his support interactions now run through AI agents and about half through staff. Read side by side, they make the same point. The agent takes about half. Whether it pays depends on how well you run the other half.

CMA CGM: 54 million requests and one narrow job

CMA CGM says it handles more than 54 million customer requests a year. It did not point its AI voice agents, built with the start-up HappyRobot, at all of them. It started with pilots and then picked one family of questions in customer service: cargo release — release status, customs clearance and shipping documentation. The agents can also trigger emails from internal systems that carry the cargo release number, so the customer gets the document, not just an answer.

The reported result: more than 46% of those enquiries resolved directly by the AI agents. The rest go to people.

Two labels belong on those numbers. They come from CMA CGM's own announcement, published jointly with the vendor, and CMA CGM has also invested in HappyRobot. They are a company's account of its own project, not an audit. And 46% is a share of one narrow question type, not of all 54 million requests. That narrowness is the lesson. The company chose questions that are frequent, repetitive and answerable from data it already holds.

Salesforce: the vendor's 85% and the boss's 50%

Salesforce runs its own AI agent on its help site, and it reports two very different-looking numbers. In July 2025 the company said the agent had handled more than 1 million conversations, with resolution rates in the 85% range. The same company post noted that this was only a small fraction of the 3 to 4 million visits its help site receives each month. Weeks later, Benioff told The Logan Bartlett Show that he had reduced support headcount from about 9,000 to about 5,000, with interactions now split roughly half and half between agents and staff. A Salesforce spokesperson said support cases had declined, the company no longer needed to actively backfill support engineer roles, and hundreds of people had been redeployed into professional services, sales and customer success.

Both numbers can be true. The 85% is measured on the conversations the agent receives. The half is measured on the whole support operation. These are vendor claims about the vendor's own product, so hold them at arm's length. But the gap between them is the most useful thing a buyer can learn: a resolution rate means nothing until you know the denominator.

Why most service teams can't show a return yet

Gartner surveyed 1,303 senior leaders between January and April 2026. Service and support leaders put a median 12% of their 2025 budgets into AI, the highest of ten business functions assessed. Only 24% demonstrated positive financial returns across their AI use cases. A separate Gartner survey of 199 service leaders found AI spending up 38% while overall service budgets grew 2%, with money moving away from labour and overhead. And in a survey of 3,566 customers, 87% said it is essential that a company using generative AI for service offers a way to reach a person.

Spend is moving from people to software, customers still insist on a person, and three in four leaders can't yet show the money came back. We covered the sharp end of this in how Klarna and Commonwealth Bank rehired people after AI cuts: both removed staff before the work was proved covered, and the unanswered half came back as overtime, apologies and rehiring.

The other half is where the money is decided

An agent that resolves 46% of a question type has not saved 46% of the cost. It has moved the hard 54% to your staff, often in a worse state: the customer has already tried once, is less patient, and may be repeating themselves. If the handover loses the context, your people spend their attention rebuilding the case. If the agent answers when it should have handed over, you can end up where Air Canada did when a tribunal held it to its chatbot's answer.

That is why we treat customer service agents as a division of labour, not a headcount line. The machine takes the routine questions — status, documents, opening hours, simple changes — so your people's attention stays on judgment: the exception, the upset customer, the decision involving money or risk. Done well, the half that reaches a person arrives with the case already summarised. Done badly, it arrives angrier.

How a small firm should scope an AI customer service agent

1. Pick one question type, not "customer service". CMA CGM chose cargo release. You might choose order status, appointment changes or document requests. Pull fifty recent examples.
2. Check the answer already exists. If the answer lives in someone's head rather than a system or a written policy, write it down first.
3. Decide what the agent may do alone, what it drafts for approval, and what it must hand over immediately.
4. Design the handover before the bot. A person must get the full conversation and a one-line summary, and the customer must always be able to ask for a person.
5. Measure finished resolutions against all questions of that type, plus the ones staff had to repair. That is your real rate, and the basis of any cost per finished task.

If you want help scoping the agent, the handover and the measurement, that is the work we do in AI agents for business and through our automation service. If phones are the bottleneck, our guide to the best AI receptionist for small business covers the voice side.

Frequently asked questions

What is an AI customer service agent?

Software that answers customer questions and takes simple actions, such as checking an order status or resending a document, using your own systems and policies. It still needs written policies, limits on what it may do alone, and a clear handover to a person.

What resolution rate should I expect from an AI customer service agent?

Ask what is being counted. CMA CGM reports more than 46% of cargo-release enquiries resolved directly by its AI agents. Salesforce has reported resolution rates around 85% for conversations that reach its help-site agent, but that agent saw only a small fraction of 3 to 4 million monthly help-site visits. Measure resolution against all the questions of one type you receive, not only the ones the agent chose to take.

Will an AI customer service agent replace my support staff?

It can absorb routine questions, which may mean you stop backfilling roles, as Salesforce says it did. Klarna and Commonwealth Bank both brought people back after cutting too early. Measure what the agent finishes first, then decide what your people do with the time released.

How do I measure the ROI of an AI customer service agent?

Count finished resolutions that met your standard, then price the rest: questions handed to a person, answers someone had to repair, and the subscription and setup cost. Compare that with the same questions answered by staff alone.

Tell Ada the one customer question your team answers most often. She will sort which half an agent could finish and which half needs a person, and put it in your free report. Talk to Ada and request your free report.

Sources: CMA CGM Group, "CMA CGM and HappyRobot scale up a tailored voice-based AI solution to support CMA CGM's customers and employees", 14 October 2026 (company announcement made with the vendor; CMA CGM has invested in HappyRobot). Salesforce, "Salesforce's Learnings with a Customer Service AI Agent", July 2025, and "Salesforce Exec Shares Agentforce Customer Support Lessons Learned" (vendor figures about its own product). Marc Benioff on The Logan Bartlett Show and a Salesforce spokesperson statement, as reported by Fortune and Business Insider, September 2025. Gartner press releases of 8 July, 4 August and 26 August 2026.