AI Agents for Customer Service: What They Can Do and How to Deploy One Safely

AI agents for customer service can resolve high-volume, low-judgment contacts on their own today: order status, policy questions, password resets, returns inside policy, and small refunds under a cap. They still fail on exceptions, complaints, and anything where a wrong answer binds the company.

The work is choosing which contacts the agent owns, capping its actions, and measuring real resolution instead of silence.

What can AI agents for customer service actually do today?

Two products share the label. Agent assist drafts answers for a human. A customer-facing agent talks to the customer and acts in your systems. The chatbot vs agent distinction covers definitions.

Agent assist has the strongest evidence. A 2025 Quarterly Journal of Economics study followed the staggered rollout of an AI assistant that suggested responses to customer-support agents. It covers 5,172 agents, 1,636 of whom were observed after getting access. Issues resolved per hour rose 15% on average and 30% for less skilled and less experienced workers. The most experienced and highest-skilled agents saw small speed gains and small declines in quality. Customers were less likely to ask for a manager.

Customer-facing autonomy has weaker evidence. In a Gartner survey of 5,728 customers fielded in December 2023, only 14% of service issues were fully resolved in self-service, and only 36% of issues customers called “very simple.” That fieldwork predates most agentic tools, so treat it as the baseline agents must beat.

A Gartner survey of 3,566 B2B and B2C customers, run in February and March 2026, found 49% would have been willing to use a chatbot, but only 7% used one in their most recent service interaction. Only 27% would try a chatbot again after a bad experience.

That last number matters most. A failed conversation shrinks the pool of customers who will try the agent next time.

Which contact types should an AI agent handle?

Score every contact type on three axes before you automate it.

Volume is the share of total contacts.

Policy complexity is how often the right answer depends on an exception, a judgment call, or a changing policy. In the same 2024 survey, 45% of customers who started in self-service said the company did not understand what they were trying to do. Exceptions are where that happens.

Action risk is what happens if the agent acts wrongly. Reading an order is safe. Moving money is not.

Decision rule: automate first where volume is high, policy complexity is low, and the action is read-only or easy to reverse. Where the action moves money or cannot be undone, the agent proposes and a person approves, or a hard cap applies. Where the policy needs judgment, the agent gathers facts and routes.

The limits below are my defaults. Set dollar caps at or below what your tier-1 humans can approve without a supervisor.

Identity fields (email, phone, payout bank account) are missing on purpose. A chat agent should never change the details used to verify the customer.

Resolution vs deflection: what is the difference?

Deflection means the contact did not reach a human. It says nothing about whether the problem was solved. A customer who gives up counts as deflected.

Resolution means the problem was solved. My test has two parts: no repeat contact about the same issue on any channel within 7 days, and a QA reviewer agreeing the answer and action were correct. The 7-day window is my choice.

Vendor “resolution” numbers usually sit between the two. The billing meter shows why.

Intercom bills Fin at $0.99 per outcome. A resolution counts when the customer confirms Fin’s last answer helped, or “exits the conversation without requesting further assistance,” which Intercom calls an assumed resolution. Its FAQ says a frustrated customer who just leaves is billed that way. Intercom deducts the charge if the customer returns to the same conversation. Its rules do not cover a customer who gives up and calls your phone line.

Zendesk’s resolution tiers, introduced May 18, 2026, are stricter. An LLM reviews each conversation, and only “Verified” resolutions draw down your allowance. But the check runs at conversation end, which for messaging defaults to 2 hours after the first message. A repeat contact the next day falls outside it.

Salesforce Agentforce charges $2 per conversation for customer-facing agents. That meter counts conversations, not outcomes. The build vs buy guide compares the full set of meters.

Worked example (hypothetical). An agent handles 10,000 conversations a month. The vendor reports 6,500 resolutions: 4,000 confirmed and 2,500 assumed. You join contact records across channels and find 30% of assumed resolutions (750) and 5% of confirmed ones (200) came back within 7 days. Assume every repeat came through a new conversation or another channel, so Intercom’s same-conversation deduction never applies and all 6,500 stay billed. This example applies no QA check.

• Resolutions after excluding repeat contacts: 6,500 minus 950 is 5,550, or 55.5%, not 65%.
• Billing at $0.99: 6,500 times $0.99 is $6,435, or $1.16 per remaining resolution.
• If each repeat contact costs $6 of human time, that adds $5,700. Total cost per remaining resolution is $12,135 divided by 5,550, or $2.19.

If only 15% of assumed resolutions come back, remaining resolutions are 5,925 (59.3%) and the cost is $9,885 divided by 5,925, or $1.67. The dashboard overstates the result either way.

At $2 per conversation, the same month costs $20,000, or $3.60 per remaining resolution at 5,550. The outcome-based pricing guide covers contract terms that tie the meter to a stricter definition.

What the public record shows

Klarna. In February 2024, Klarna said its OpenAI-powered assistant handled 2.3 million conversations, two-thirds of its customer service chats, in its first month. It claimed the work of 700 full-time agents, CSAT on par with humans, and 25% fewer repeat inquiries. By May 2025 the CEO told Bloomberg that “cost unfortunately seems to have been a too predominant evaluation factor” and the result was lower quality, as CX Dive reported. Klarna renewed its investment in human support and reaffirmed that customers can always reach a person. Its February 2024 release had already said customers could choose live agents. A spokesperson said the assistant still handled two-thirds of inquiries.

Commonwealth Bank of Australia. In July 2025, CBA said 45 customer service roles would go after it introduced an AI voice bot. In August it reversed the decision, saying its assessment “did not adequately consider all relevant business considerations.” The Finance Sector Union said call volumes were rising and managers were offering overtime. My read: the plan assumed demand the bot would remove, and the queue said otherwise.

Air Canada. In Moffatt v. Air Canada, decided February 14, 2024, the airline’s chatbot told a grieving customer they could apply for a bereavement fare after travel. The linked policy page said the opposite. Air Canada argued it could not be liable for its chatbot. The tribunal called that “a remarkable submission” and said the customer had no reason to know which part of the website was accurate. It ordered C$812.02 in total. The amount was small; the rule is not: a policy statement from your agent is a statement from your company.

What actions should an AI agent be allowed to take?

Actions save the most time and do the most damage. Five limits:

Compute eligibility, do not infer it. The agent may refund only when a rule over system data says yes: order date inside the window, item marked returnable, no prior refund. If the model has to interpret the policy, the case is not eligible for automation.

Cap per action and per day. Say 400 refund-eligible contacts arrive a day, each capped at $50. A broken rule could pay out $20,000 a day. A $5,000 daily aggregate cap stops it after 100 maximum-size refunds.

Quote policy, do not paraphrase it. Moffatt is a case of incorrect policy information. For any answer that creates an entitlement, such as a fare, warranty, or refund window, the agent quotes current policy text and links the source. Re-run your policy test questions whenever the page changes. The hallucination budget guide covers how much error each workflow can take.

Confirm before irreversible steps. Cancellations and plan changes need an explicit yes after the agent states the price and effect in plain words.

Give the agent its own narrow credentials. An agent that cannot change an email address cannot be talked into doing it. The least-privilege guide shows how to cut scopes. Log every action with the policy version and the data behind it.

How should the handoff to a human work?

Customers judge an AI agent by how easily it lets them leave. A Gartner release from August 2026, drawing on a February to March 2026 survey of 3,566 customers, found 87% say access to a human agent is essential when a company uses GenAI for service. Gartner’s analyst advises against making GenAI “a mandatory first step for every issue.”

Four rules for the customer-facing side:

1. Honor “human” on the first ask. Only 27% of customers would try a chatbot again after a bad experience, so a forced retry is expensive.
2. Pass the context. The human sees the transcript, the facts collected, and any action taken. The customer never repeats their order number.
3. Say what happens next when no one is available. At 2 a.m., the agent says when a person will reply and opens the case.
4. Disclose the AI at the start. In the EU, Article 50 of the AI Act requires that people are told they are interacting with an AI system, unless it is obvious, at the latest at the first interaction. The transparency rules have applied since August 2, 2026. I am not a lawyer, so check your own jurisdictions.

The trigger, payload, owner, and response-time contract behind these rules is in the escalation design guide. Voice adds latency and consent rules, covered in the voice agent guide.

A measurement plan with thresholds

Measure each contact type separately, against human-handled contacts of the same type and period. The thresholds are mine; tighten them for regulated work.

Read CSAT next to repeat contacts, never alone. CSAT can suffer from nonresponse bias, so compare survey response rates for agent and human contacts before trusting the gap.

The QA sample needs real math. If the true critical error rate is 1%, a random sample of 100 conversations has a 63% chance of showing at least one (1 minus 0.99 to the 100th power). A sample of 300 raises that to 95%. So I would review about 300 agent-closed conversations per contact type before expanding it, scoring answer, action, and handoff.

Roll out each contact type in three steps: shadow mode, then live with every refund approved, then no review below the cap. Expand only when the table says so.

Bottom Line

AI agents for customer service work on contacts that are frequent, rule-based, and cheap to get wrong.

Pick contact types with the triage table. Cap every action that moves money. Measure resolution by what the customer does next, not by what the dashboard counts.

Klarna kept its agent and reinvested in human support. Start with both.

FAQ

Do I have to tell customers they are talking to an AI agent?

In the EU, yes, since August 2, 2026, unless it is obvious to a reasonably well-informed person. Article 50(5) says the information must be clear and given by the first interaction. Elsewhere, I would disclose anyway.

Is a company liable for what its customer service AI says?

In Moffatt v. Air Canada, a British Columbia small-claims tribunal held the airline responsible for its chatbot’s wrong answer. An Air Canada representative had already admitted in email that the chatbot used “misleading words.” Treat it as a signal of how tribunals reason, not settled law everywhere.

Will AI agents replace customer service teams?

Not so far. In a Gartner survey of 321 service leaders run in October 2025, only 20% had reduced agent staffing due to AI. Gartner predicts that by 2027, half of companies that attributed headcount cuts to AI will rehire for similar work under different titles.

How much do customer service AI agents cost?

List prices I checked on October 15, 2026: Intercom Fin is $0.99 per outcome, and Agentforce is $2 per conversation or $500 per 100,000 Flex Credits. The price that matters is cost per resolution after excluding repeat contacts, which ran 1.7 to 2.2 times the list price per outcome in the worked example.

How long before I know whether the agent works?

My estimate is at least a month per contact type. Each week’s 7-day repeat window closes a week later, and you need about 300 reviewed conversations to catch a 1% critical error rate with 95% confidence.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect