Hear it call your own phone
Half an hour. We build a flow for your enquiries and run it against your number.
What AI agents can't do yet in customer service, and what to do instead
AI agents now handle routine conversations well. The work that still gets stuck is stuck for four specific reasons, and each one has a practical answer that involves a person.
By the Unibound team6 min read
Most writing about AI in customer service is either a sales pitch or a warning. This is neither. We build a platform where AI agents and people work the same conversations, so we have a commercial interest in being accurate about where each one belongs. If the AI is given work it can't finish, our customers' customers are the ones who suffer.
So here is the honest position at the end of 2026: what AI agents do well, the four places they still fall short, and what to build around each gap.
What AI agents now do well
The gloomy statistics that circulate are mostly out of date. When Sierra's researchers published the τ-bench benchmark in 2024, the best model completed fewer than half of its simulated retail and airline service tasks. On the current version of that benchmark, the leading models complete around 85 to 88%.
Real deployments show the same thing for routine work. Octopus Energy reported in July 2026 that emails answered by its AI scored 76% customer satisfaction against 72% for comparable replies written by people, across about 8,000 emails a week.
The detail matters, though. Octopus's AI handles routine queries only. Its emails are labelled as AI, a person is available on request, and complex or vulnerable cases go straight to staff. It is a good result for AI inside a design that has people in it.
Limit one: an agent can only act through the tools it has
This is the limit people underestimate most. An AI agent that can talk fluently about your refund policy cannot issue a refund unless it has a connection to the system that issues refunds. Without one it can explain, apologise and take notes, and that is all.
Most businesses are a long way from connected. MuleSoft's 2026 survey of 1,050 IT leaders found the average organisation runs 957 applications, of which 27% are connected to each other. In the same survey, 82% named data integration as one of their biggest obstacles to using AI.
Standards such as the Model Context Protocol (MCP) are making connections easier to build, but they don't conjure them. If your job-booking system, your quoting spreadsheet or your warranty database has no API and no MCP server, the agent cannot use it. Benchmarks that test agents against realistic business systems reflect this: on Salesforce's CRMArena-Pro, leading agents managed about 58% of single-step tasks and about 35% once the task took several turns.
What to do instead: decide, system by system, whether the agent acts or a person does. Where a connection exists, let the agent use it. Where it doesn't, the agent's job is to collect everything a person will need and pass the conversation over. We cover this in detail in what an AI agent can do without access to your systems.
Limit two: being right once is not being right every time
A demo is one conversation. Production is ten thousand. The original τ-bench paper measured this directly: agents that could solve a task once solved it fewer than one time in four when asked eight times running. Models have improved a great deal since, but consistency figures for the current generation are rarely published, and it is consistency your customers experience.
The failures that reach the news are usually confident inventions:
- Air Canada's chatbot told a customer he could claim a bereavement fare retroactively. A tribunal held the airline to it.
- Cursor's AI support agent invented a one-device policy to explain a bug, and customers cancelled before the company corrected it.
- In September 2026, United's chatbot told a customer a travel credit lasted five years. It lasted one.
What to do instead: keep the rules out of the prompt. Things that must always be true, such as who may be contacted, when, how often and at what cost, should be enforced by the system around the agent, not requested of it. Give the agent your own written material to answer from, and review a sample of real conversations continuously. In Unibound these are safeguards and monitoring; whatever you use, the principle is the same.
Limit three: some customers want a person, and they are entitled to
You can build an agent that resolves a query perfectly and still lose the customer, because they wanted to speak to someone. Gartner's 2026 survey of 3,566 customers found 87% say companies using generative AI for service must provide access to a human. Five9's survey of 3,000 consumers in the US, UK and Germany found two-thirds still prefer a person, and that trust in AI to resolve an issue rises from 26% to 55% when a human is reachable.
That last figure is the useful one. Offering a person doesn't only help the people who take up the offer. It makes everyone else more willing to let the AI try. We look at the full set of surveys in do customers want AI or a human?
Regulators are moving the same way. Ofcom said in September 2026 that it would be concerned if AI acted as a gatekeeper to support, and Spain's new customer service law requires a human on request from the end of 2026.
What to do instead: make "can I speak to someone?" a first-class outcome in every conversation, on every channel, with somewhere real for it to go: a transfer, a place in a call queue, or a reply from a named person in the morning.
Limit four: judgement, emotion and exceptions
The most rigorous recent evidence here is a 2026 field experiment at Alibaba's Taobao, where support workers supervised an AI agent that resolved eligible chats itself. Conversations got shorter, but customer ratings for the AI-handled chats fell substantially. When people stepped in on technical problems, quality was preserved. When they were brought in late, after the customer was already frustrated, it was much harder to recover.
Two lessons follow. An upset customer is not a routine contact, however routine their question. And a handover that arrives after the damage is done is worth far less than one that arrives early.
What to do instead: hand over on topic, not only on failure. Complaints, bereavement, money disputes and anything involving a vulnerable customer should go to a person by rule. See how to hand over between AI and people for the mechanics.
What happened to the companies that went AI-only
The record is mixed, and worth stating fairly.
- Klarna said in 2024 that its AI assistant was doing the work of 700 agents. In 2025 its chief executive said cost had weighed too heavily, that quality had dropped, and that a human would always be available to customers who want one. It remains heavily automated.
- Commonwealth Bank of Australia reversed 45 redundancies made after introducing a voice bot, and apologised, when call volumes rose instead of falling.
- Salesforce cut its support headcount from about 9,000 to about 5,000, citing its own AI agents, and has not reversed it.
Gartner's reading of the trend is that no Fortune 500 company will have eliminated human service by 2028. The companies doing well are not the ones that removed people. They are the ones that changed what people do.
Design for the limit
Put together, the four limits point to one design: a written-down journey in which the AI does what it is good at, and every place it might get stuck has a route to a person.
| Where the AI stops | What the journey should do |
|---|---|
| No connection to the system that holds the answer | Collect the details, tell the customer what happens next, pass to a person with a summary |
| A decision that needs discretion | Route to the person with authority; let the AI confirm the outcome afterwards |
| The customer asks for a person | Transfer or queue straight away, with the transcript, on any channel |
| Complaint, distress or vulnerability | Hand over by rule, early, before the AI tries to solve it |
| Rules that must never be broken | Enforce them in the system, not in the prompt |
Here is that design as a support flow. The AI answers first; a refund needs a person; the person hands the conversation back when the decision is made.
None of this is an argument against AI agents. It is an argument for being specific about the job. An agent with a clear task, the tools to do it and a person one step away will finish far more conversations than a cleverer agent left on its own. That is what AI and your team on one platform means in practice.
Questions people ask
Will AI replace call centre and customer service agents?
Not wholesale, on current evidence. Gartner predicts that no Fortune 500 company will have fully eliminated human customer service by 2028, and that half of the companies that cut service staff because of AI will rehire for similar work by 2027. What is changing is the mix: AI takes routine contacts, and people take the complicated, sensitive and exceptional ones.
What should AI never handle on its own?
Anything where being wrong is expensive or the customer is distressed: complaints, bereavement, vulnerable customers, disputes about money, and decisions that need discretion such as refunds outside policy. An AI agent can still take the first contact, gather the facts and pass the conversation to a person.
Why do AI agents fail in production when the demo looked good?
Usually because the demo was one conversation and production is thousands. Agents that succeed on a task once are less consistent across repeats, real customers go off script, and the agent often lacks access to the system that holds the answer. The fix is to narrow what the agent is asked to do, connect the tools it needs, and give it a clean route to a person.
Do customers trust AI customer service?
They trust it for simple things, and much more when a person is reachable. In Five9's 2026 survey of 3,000 consumers, trust in AI to resolve an issue was 26% with no human option and 55% when a human was available.
Sources
- Sierra Research, τ-bench leaderboard (2026). taubench.com
- Yao et al., τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains (2024). arxiv.org
- Octopus Energy, AI trial wins customer approval (July 2026). octopus.energy
- MuleSoft, 2026 Connectivity Benchmark Report. blogs.mulesoft.com
- Huang et al., CRMArena-Pro (Salesforce AI Research, 2025). arxiv.org
- Gartner, 87% of customers say companies using GenAI for customer service must provide access to a human agent (August 2026). gartner.com
- Five9, Consumer CX Report (October 2026). five9.com
- Wang et al., agentic AI in customer service: a field experiment (2026 preprint). arxiv.org
- The Register, Cursor AI support bot invents a policy (April 2025). theregister.com
- NBC Chicago, AI chatbots are taking over customer service, but how often are they wrong? (September 2026). nbcchicago.com
- Bloomberg, Commonwealth Bank reverses job cuts decision over AI chatbots (August 2025). bloomberg.com
- Gartner, none of the Fortune 500 will have fully eliminated human customer service by 2028 (September 2025). gartner.com
- Ofcom, consumers using AI to manage telecoms services, but human contact still valued (September 2026). ofcom.org.uk
