What an AI customer-service assistant can really do today, and what it still can't

Ask ten business owners what an AI customer-service assistant can do, and you will hear ten answers, from “it answers everything” to “it makes things up.” Both are partly true. These assistants are now good enough that customers trust what they say, and when one is wrong, the business owns every word. Here is what they can do, what they can’t, and what your customers think of them, from sources you can check.

What they do well

• Answer the questions you already answer every day. Hours, prices, policies, how-to, order and booking questions: anything your website or documents already explain.
• Answer at any hour, instantly, on every channel at once. Website chat, email, text and the phone, including the 9:40 p.m. question nobody is there to take.
• Speak your customer’s language. The better systems handle dozens of languages in writing; spoken languages are fewer.
• Make your people better. One of the largest studies so far followed 5,172 support agents at one company: with an AI assistant, they resolved 15 percent more issues per hour on average, and newer, less experienced agents 30 percent more. Customers were also more polite and less likely to ask for a manager (Brynjolfsson, Li and Raymond, Quarterly Journal of Economics, May 2025). That study measured AI helping human agents, not AI answering on its own.

How much can it handle on its own?

Every vendor publishes a number, and every number is the vendor’s own. Fin (formerly Intercom, bought by Salesforce in September 2026) says its AI resolves an average of 76 percent of conversations across more than 12,000 customers. Salesforce says its own AI resolved 70 percent of 4.3 million inquiries on its help site (June 2026). Ours is 85 percent or more of inbound phone calls, in a study by our largest customer, Curacao, in April 2026. Treat all of these, ours included, as claims, and ask one question about each: what counts as “resolved”? Some vendors count a conversation as resolved when the customer simply leaves without asking for more help. A customer who gave up looks exactly the same.

What they still can’t do

• Know what it was never told. An assistant is only as good as the information it reads. If your website is out of date, so are its answers.
• Be hallucination-proof. The real danger is not “I don’t know.” It is a confident answer that is wrong. In April 2025, the support bot of Cursor, a software company, told users about a log-in policy that did not exist; the co-founder had to apologize publicly. Even the biggest names get about one answer in ten wrong: in an analysis done for The New York Times, Google’s AI answers were correct about 91 percent of the time on a standard factual test (April 2026). Specialist tools that answer only from their own vetted content do no better: Stanford researchers found the legal AI tools from LexisNexis and Thomson Reuters gave false information 17 to 33 percent of the time (2025). Ask any vendor what its assistant does when it cannot find the answer, and how it checks an answer before your customer sees it.
• Replace humans. A furious customer, a billing dispute, a grieving family: these need a person. In a 2026 Gartner survey, 85 percent of service leaders said they are expanding their human agents’ responsibilities, and Gartner stresses human involvement in “complex, high-stakes or advisory interactions” (April 28, 2026).

What your customers think

This is the part most vendors skip. In Gartner’s survey of 3,566 customers in February and March 2026 (published August 4 and September 2, 2026):

• 87 percent say access to a human is essential when a company uses AI in customer service.
• Only 27 percent would try a chatbot again after a bad experience.
• 49 percent would have been willing to use a chatbot, but only 7 percent actually did in their most recent service contact.

Read together, those numbers say something hopeful: customers are willing, but they have been burned. The assistant that wins is the one that is reliable and makes reaching a person easy, not the one that traps them.

The rules that already apply

• You are responsible for what it says. In 2024, a Canadian tribunal ordered Air Canada to honor a fare discount its website chatbot had wrongly promised, rejecting the airline’s argument that the chatbot was responsible for its own actions. In California, since January 1, 2026, a business sued over harm cannot defend itself by saying the AI acted on its own (Civil Code 1714.46).
• Say it is AI. Since July 1, 2019, California has made it unlawful to use an online bot that hides that it is a bot in order to sell to someone; saying clearly that it is a bot avoids that (Business and Professions Code 17941). In the EU, since August 2, 2026, people must be told when they are talking to an AI unless it is obvious (AI Act, Article 50). Utah requires a business to say it is using AI when a customer clearly asks, and up front for high-risk advice in licensed professions.
• Outbound calls need consent. The FCC ruled in February 2024 that AI-generated voices count as “artificial” voices under federal law, so AI calls to customers generally need their prior express consent (written consent for marketing calls). Answering calls that come in to you is not affected.

How we handle these limits

We build one of these assistants, so here is how ours deals with each limit above, in our own words. It answers from your website and documents. Research from MIT and Google shows AI answers get more accurate when several models check one another (2024), so ours does not rely on one model: each step, from understanding the question to a separate check of the draft answer, runs on the best-suited model from competing AI vendors. In two separate studies by two of our largest customers, Curacao and Omie, of tens of thousands of real answers to their customers’ questions, over 97 percent of the answers to questions their content covers were correct. When it cannot find a source, it says so before answering. It does not improvise medical, legal or financial claims. It hands the conversation to a person on every channel: on the phone, by bringing your live agent into the call, during the hours you set; in website chat, email, text and WhatsApp, by sending a link to Live Connect, our free live-agent app, whenever one of your agents is signed in. It does so when it cannot answer, when a customer is frustrated, and on the topics you mark as human-only. It writes in over 100 languages and speaks six on the phone. For actions such as payments and account changes, it follows fixed business steps rather than letting the AI decide. The conversation is not perfect, but it is more accurate than most human agents: the same audits found it more accurate than practically all of Omie’s tier-1 support engineers and Curacao’s top customer-service agents. For a sense of how often people get it wrong, when U.S. Treasury auditors tested the IRS’s own walk-in assistance centers, they got the correct answer to their tax questions in only 33 of 61 visits (June 2026). The transactions are a different matter: every service, from a payment to an account change, runs as a fixed, deterministic script, so that layer is hallucination-proof by design. The AI can misread what a customer wants; it can never invent a step, an amount or an action.

Five questions to ask before you choose any assistant

1. Where do its answers come from, and what happens when it cannot find one?
2. What does “resolved” mean in your numbers?
3. How does a customer reach a person, and on which channels?
4. Does it tell customers it is an AI?
5. Who is responsible when it is wrong? (The answer is you. Ask what the vendor does to keep that from happening.)

My view: the question is no longer whether AI can answer your customers. It can, and well. The question is whether it knows when to stop. Pick the one that says “I don’t know, let me get you a person” over the one that always has an answer.