A few years ago, a website chatbot was a menu in disguise: pick an option, get a canned answer, and give up looking for a phone number. Today’s AI assistants can read your help docs, understand a customer’s question in plain language, look up an order, and write a reply that sounds human. For a small company that can’t staff support around the clock, the appeal is obvious.

But the story isn’t all success. Customers remain wary, some of the most publicized AI service rollouts have quietly reversed course, and an AI that confidently invents an answer can cost you a customer, or a court case. Here’s what’s changed, what customers actually want, and how to put an AI assistant to work without the backlash.

From scripted bots to AI agents

There are really three generations of chatbot in use today:

GenerationHow it worksStrengthsWeaknesses
Rule-based botsDecision trees and keyword matchingPredictable, cheap, never makes things upRigid; frustrates anyone whose question doesn’t fit the menu
Generative AI assistantsA large language model answers in natural language, ideally grounded in your own contentHandles varied questions; available 24/7Can produce confident, wrong answers if poorly grounded
AI agentsAn AI assistant that can also take actions: check an order, reschedule, issue a refundResolves issues end to endMistakes have real consequences; needs tight permissions and oversight

The jump from the second to the third generation is the important one. An assistant that only talks can mislead a customer. An agent that acts can also process a refund you didn’t authorize. The more a system can do, the more carefully its boundaries need to be drawn.

The numbers behind the rise

Adoption is climbing fast. In Salesforce’s 2025 State of Service survey of 6,500 service professionals in 39 countries, AI was already handling about 30% of customer service cases, and respondents expected that to reach 50% by 2027. AI jumped from tenth to second on service leaders’ priority lists in a single year.

The efficiency gains are real. The most rigorous evidence comes from a study by economists Erik Brynjolfsson, Danielle Li and Lindsey Raymond, who tracked 5,179 customer support agents as a company rolled out an AI assistant. Agents with the tool resolved 14% more issues per hour on average, and newer, less experienced agents improved by 34%, as the AI effectively passed along the know-how of the company’s best performers. Customer sentiment improved and fewer agents quit. Salesforce’s survey points the same way: service reps using AI reported spending 20% less time on routine cases, freeing roughly four hours a week for harder problems.

Who’s getting it right

  • Bank of America’s Erica. Launched in 2018, Erica had passed 3 billion client interactions and nearly 50 million users by August 2025, averaging more than 58 million interactions a month. The bank says Erica helps clients find what they need 98% of the time. Versions of it now handle more than 40% of interactions in the bank’s business-client chat and have cut employees’ IT help desk calls in half. Notably, Erica was built up steadily over seven years, not launched as a big-bang replacement for people.
  • Klarna’s first month. When the payments company launched its AI assistant in early 2024, it reported handling 2.3 million conversations in the first month, two-thirds of all its service chats, with average resolution time dropping from 11 minutes to 2 and repeat inquiries down 25%. Klarna projected a $40 million profit improvement that year. (The story has a second chapter, below.)
  • Octopus Energy. The UK energy supplier began using AI to draft replies to customer emails in 2023, with staff reviewing each draft before it’s sent. Its CEO, Greg Jackson, reported that the AI-assisted emails earned 80% customer satisfaction, compared with 65% for emails written by people alone, and that the AI was doing work equivalent to about 250 employees.
  • AI for the agents themselves. In the study above, the AI didn’t talk to customers directly. It suggested responses to human agents in real time. Some of the biggest gains come from helping people, not replacing them.

The common thread: the most successful programs pair AI with people, grow it step by step, and measure whether customers’ problems actually get solved.

What customers actually want

Customers are more skeptical than the technology’s boosters suggest. In a late-2023 Gartner survey of nearly 5,800 customers, 64% said they would prefer companies didn’t use AI for customer service, and 53% said they’d consider switching to a competitor over it. Their biggest worry wasn’t the technology itself. It was that AI would make it harder to reach a person.

Attitudes are softening, but with conditions. In a June 2026 survey of 1,000 U.S. consumers by the customer-support company Kim.cc, 45% said they welcome AI for simple tasks like order tracking, FAQs and store hours. But most still wanted a person for complaints (60%) and billing disputes (50%). Customers aren’t rejecting AI; they’re rejecting bad service, and they want a blend of both.

Some early adopters learned this firsthand. After Klarna’s impressive first month, its AI assistant took on more and more of the company’s support work. By May 2025, it had resumed hiring human support staff, with its CEO acknowledging that the AI had cut costs but fallen short of the company’s standards for customer experience.

The economics may also shift. In January 2026, Gartner predicted that by 2030 the cost per resolution for generative AI in customer service will exceed $3, more than many offshore human agents, as AI vendors move toward profitability. Gartner’s advice: use AI to improve engagement and loyalty, not just to cut costs.

The lesson isn’t to avoid AI. It’s that customers judge the outcome: was the problem solved, quickly, and could they reach a human when it mattered?

When execution fails: hopeless bots in the wild

The failures are instructive because they’re so avoidable:

  • DPD (January 2024). After a system update, the UK delivery company’s chatbot could be coaxed into swearing at a customer and writing a poem calling DPD the worst delivery firm in the world. The exchange went viral, and DPD switched off the AI feature. The fix: test after every update, and set firm limits on tone and topics.
  • A Chevrolet dealership (December 2023). A visitor manipulated a dealer’s website chatbot into agreeing to sell a new Tahoe, worth more than $70,000, for $1, calling it “a legally binding offer.” The fix: block bots from negotiating prices or making commitments.
  • New York City’s MyCity chatbot (2024). Built to help small businesses, it told owners they could take a cut of workers’ tips and refuse cash payments, both illegal in the city. The city added disclaimers and narrowed its scope, and in January 2026 the new mayor announced he would shut it down, calling it functionally unusable. The fix: keep bots away from legal advice, and fix known errors fast.

Notice the pattern. In every case, the technology did what it was built to do: generate a plausible answer. The failures came from what the companies didn’t build around it, such as grounding in real policy, limits on what the bot could promise, testing, and a clear path to a person. (For more on why AI makes things up, see Botshit in the Boardroom.)

Where AI assistants work well, and where they backfire

Good fits:

  • Frequent, simple questions: hours, pricing basics, shipping times, how-to steps from your help center.
  • After-hours coverage: answering what it can and collecting details so a person can follow up first thing.
  • Order and account lookups: status checks and simple changes, with secure verification.
  • Triage: understanding what a customer needs and routing them to the right person with the context already captured.
  • Helping your team: drafting replies, summarizing long threads, and suggesting help articles for a human to review and send.

Poor fits, or proceed with care:

  • Anything involving money, policy or legal commitments, such as refunds, pricing exceptions or contract terms. Air Canada learned this when a tribunal made it honor a refund policy its chatbot invented (see Botshit in the Boardroom).
  • Upset or vulnerable customers, who need empathy and judgment.
  • Complex or unusual problems that don’t match your documentation.
  • Regulated advice: health, financial or legal guidance.

What it costs

The pricing model has changed along with the technology. Many customer service platforms now charge for AI by the conversation it resolves, not just by the seat. Intercom, for example, charges $0.99 per resolution for its Fin AI agent, on top of seats that start around $29 a month, or with a small monthly minimum if you run it alongside another help desk. Others bundle a pool of AI resolutions into each pricing tier.

For a small business, that makes it easy to start small. If an AI assistant fully resolves 500 routine questions a month, that’s roughly $500, which you can weigh against the staff time those questions take today. Two things to watch: how the vendor defines a “resolution” (some count a conversation as resolved if the customer simply doesn’t reply), and Gartner’s warning, above, that AI costs per conversation are likely to rise.

How to deploy one well

  1. Start with the problem, not the tool. Pull the last few months of support requests and find the questions that come up most. Those are your first use cases.
  2. Fix your knowledge base first. An AI assistant is only as accurate as the content it draws from. Update help articles, policies and FAQs, and remove outdated ones.
  3. Ground it and fence it. Configure the assistant to answer only from your approved content, to say “I don’t know” rather than guess, and to stay out of topics you’ve ruled off-limits.
  4. Make the human exit obvious. A clear “talk to a person” option at every step, with the conversation history handed over so customers don’t repeat themselves.
  5. Be upfront. Tell customers they’re talking with an AI assistant. Pretending otherwise erodes trust when they figure it out, and it’s increasingly the law. California has required bots to identify themselves in certain commercial interactions since 2019, Maine now requires disclosure whenever a customer couldn’t otherwise tell, Utah requires it upfront in regulated fields like finance, and the EU’s AI Act transparency rules apply from August 2026 (overview).
  6. Limit what it can do. If it can take actions, start with read-only lookups and add permissions gradually, with limits and approvals for anything involving money.
  7. Review and improve. Read a sample of conversations every week, track the numbers below, and fix the gaps you find.

Most companies are better off starting small: one channel, a handful of question types, and a team that reviews the results. Expand once the assistant has proven it helps.

The numbers to watch:

  • Resolution rate: the share of conversations the AI fully solves, confirmed by the customer, not just closed.
  • Handoff rate: how often it passes customers to a person, and whether those handoffs are fast and include the history.
  • Customer satisfaction: for AI conversations and human ones, side by side.
  • Repeat contacts: customers who come back about the same issue within a week, a telltale sign the AI’s “resolution” didn’t stick.
  • Cost per resolution: total AI and staff cost divided by issues actually solved.

The bottom line

The chatbots have risen. The question now is whether they make customers’ lives easier or just make your company harder to reach. Used for the right jobs, grounded in accurate content and backed by real people, AI assistants can deliver faster answers and better service at a lower cost. Used as a wall between customers and your team, they cost more in trust than they save in payroll.

If you’d like help choosing and setting up AI customer service tools that fit your business, Market Tactics can help.

Sources