Customers use outside AI three times as often as your chatbot. Now what?

What did Gartner find?

Gartner found that customers prefer someone else's AI to yours. In a survey of 3,566 B2B and B2C customers in February and March 2026, people were about three times as likely to use a third-party GenAI tool, such as ChatGPT, as a company's own chatbot to sort out a service problem.¹ Gartner published the survey on 8 July 2026.

Two trend lines sit behind the headline. Use of outside GenAI tools during service interactions has nearly doubled in a year, according to the same report. Use of company chatbots has barely moved since 2022.¹ Gartner did not state the geography of the sample, so we read it as a broad customer sample rather than one country's.

The second finding matters more for anyone scoping an agent. Among GenAI users, 58% said they had used it to act on their behalf, and in B2B settings that rose to 74%.¹ Customers no longer want a bot that explains the returns policy. They want one that books the collection.

Why company bots stalled while spending rose

Company bots stalled because most of them answer questions that an outside assistant answers just as well. A FAQ bot on your site draws on your help centre. ChatGPT draws on your help centre plus forums, reviews and everything else it has read, and the customer already has it open.

The spending did not follow the usage. A separate Gartner survey of 1,303 senior leaders across industries, run from January to April 2026, found that service and support teams put a median 12% of their 2025 budget into AI, the most of the ten functions Gartner assessed. Only 24% of service and support leaders showed a positive financial return across their AI use cases.²

Our view: the stalled bot is a scoping failure. Teams launched an answer engine into a market that already had better ones, then measured it on deflection. The customers who stayed were the ones with problems the bot could not touch.

What an outside assistant cannot do

An outside assistant cannot see the customer's account, so the useful company agent starts there. ChatGPT can explain how returns usually work. It cannot confirm that this customer's order qualifies, create the return label and tell the customer when the refund lands. That needs authenticated access to your order system, and that is where your agent has no competition.

• Account actions. Order status, delivery changes, returns, bookings, subscription changes and address updates, each done against the system of record.
• Verified identity. The agent knows who it is talking to, which an outside assistant does not.
• Policy applied, not described. The agent applies your refund rule to this order instead of paraphrasing the rule.
• A path to a person. When the rule does not fit, the conversation reaches someone who can decide, with the history attached.

This is the same test we use in our scoring model for a first agent: volume, cost per transaction, exception rate and reversibility. Account actions score well when the action can be held for approval or undone. Refunds above a threshold and anything touching a contract still go to a person.

Your policies are now answered somewhere else

If customers ask an outside assistant first, the first answer about your returns window or warranty terms is written by that assistant. It may be right. It may quote a forum post from three years ago, or a competitor's policy. Either way, the customer arrives at your support channel with an expectation you did not set.

That makes public policy pages part of support design. Write returns, warranty, cancellation and delivery rules as plain text on crawlable pages, one rule per heading, with the date each rule took effect. When an assistant gets a policy wrong, trace the source the way our guide to fixing wrong AI answers sets out, and fix it where the engine reads.

Our view: the support team should own a short list of prompts customers actually ask outside assistants about the company, and check the answers monthly. That is AI visibility work, and it belongs next to the help centre, not only in marketing.

What this does to the headcount case

The survey weakens any business case built on deflecting volume from people to a bot. If customers route around the company bot, the deflection never arrives. Earlier Gartner research points the same way. In a survey of 321 customer service leaders in October 2025, 20% had reduced staffing because of AI, while 55% kept staffing stable and handled more volume.³

Gartner also predicted in February 2026 that half of the companies that cut service staff because of AI will rehire by 2027, under different job titles.⁴ That is a forecast, not an outcome, but it fits the pattern: work moves to handling the cases the bot cannot, rather than disappearing.

Make the case on resolved actions and cost per resolved contact instead. If cost visibility is the gap, our note on KPMG's Q2 survey covers why most large firms cannot yet see what their AI costs to run.

Three changes to make before the next build

Three changes follow from the survey, and none needs a new platform.

1. Rescope from answers to actions. List the ten requests that reach your team most often, mark which ones need account data, and build the agent for those first.
2. Design the handover before the happy path. Decide which requests always go to a person, what the person sees and who owns the queue.
3. Measure resolution, not containment. A conversation that ends without escalation but brings the customer back the next day is a failure the containment rate counts as a win.

For the build itself, our AI automation service puts one workflow live on the systems you already run, with acceptance criteria agreed before the build. Start with the action your customers ask for most, and keep the person in the loop until the exception rate earns trust.