Your Customers Are Asking ChatGPT for Support. What Is It Telling Them?

Picture a hardware store that spent three years remodeling its service counter. New signage, better lighting, a friendly kiosk that can answer questions around the clock. The managers are proud of it. Then somebody finally walks out to the parking lot and notices where the customers are.

They're standing around a guy leaning on his truck. He's never worked at the store. He read most of the manuals once, a while ago, and he's happy to answer anything. He's right most of the time. When he's wrong, he's wrong with total confidence, and nobody inside the store ever hears about it.

That guy is ChatGPT. Or Claude, or Gemini, or Copilot, or whatever assistant your customer already has open in another tab. And for a lot of B2B companies in 2026, he's become the first stop for support.

For Customer Success Leaders, Support and Service Ops Teams, and B2B Executives who own retention

The number Gartner just put on it

In July, Gartner published a survey of 3,566 B2B and B2C customers, conducted in February and March 2026. The headline: customers are about three times more likely to use third-party GenAI tools than company-provided chatbots when they're trying to resolve a service issue.

The trend lines behind that number are the part I'd show my CEO. Use of third-party GenAI during service interactions nearly doubled in a year. Use of company-provided chatbots has been, in Gartner's words, "statistically unchanged since 2022."

Four years of chatbot investment, and the usage line is flat.

Eric Keller, the Gartner analyst who led the research, put it plainly: "GenAI is shifting some service interactions outside of company-owned channels." The service conversation is still happening. It just isn't happening on your property.

This didn't come out of nowhere, either. A year earlier, in July 2025, Gartner surveyed 5,801 customers and found that 51% of service journeys already started on third-party platforms like Google, YouTube, and ChatGPT. Only 22% of customers started, stayed, and resolved their issue entirely inside company-owned channels. And customers reported a 62% average success rate on those outside platforms, which is a big part of why they keep going back.

Meanwhile, the money is going the other way

This is the uncomfortable part.

Gartner's August 2026 survey of 199 service and support leaders found that AI spending in service rose 38% while total service budgets grew just 2%. Leaders are pulling money out of labor and overhead to fund it. And looking two years out, those same leaders expect GenAI chatbots to become their single most valuable service channel.

So the plan is to pour more money into the channel customers have mostly ignored for four years.

The returns so far don't make the case either. In a separate Gartner survey of 1,303 senior leaders, service and support teams put a median of 12% of their 2025 budget into AI, the highest share of any of the 10 business functions measured. Only 24% of service leaders could show a positive financial return across their AI use cases.

I don't think the lesson is "stop investing in AI for support." Keller's own read is that the disappointing results have "less to do with technology limitations and more to do with misalignment with customer expectations." The tech works. It's sitting in the wrong place, waiting for customers who already found another door.

Why B2B should care more than anyone

It's tempting to file this under consumer behavior. Someone asks ChatGPT how to cancel a gym membership. Not your problem.

But look at the B2B number buried in the July release. Among customers who use GenAI, 58% said they'd used it to complete a task on their behalf, not just answer a question. In B2B, that figure was 74%.

Think about what that means for a software company. Your customer's admin isn't asking an assistant "what does this setting do?" They're asking it to write the API call, build the workflow, map the fields for the integration, or draft the SSO configuration. Then they paste the result into your product.

If the assistant learned your product from a two-year-old forum thread and a deprecated docs page, that configuration is going to break. And when it breaks, guess whose product looks broken.

The developer world is already living in the extreme version of this. Mintlify, which hosts documentation for a lot of software companies, reported in late July that AI agents made up 66% of measured traffic across the docs it powers. In July alone it logged more than 213 million agent requests against 105 million human page loads. At the start of 2026, agents were 15.2% of that traffic.

Mintlify is a docs vendor with an obvious interest in this story, and it's upfront that attribution is messy (some agents look like humans, and some humans hide behind tools). Even so, the direction is hard to argue with. For a growing share of your customers, the first reader of your help content is a machine, and the human only sees what the machine passes along.

The four ways this hurts retention

Most CS teams I talk to see support moving off-property as a deflection win. Fewer tickets, lower cost per customer, happy CFO. Some of that is real. But the costs are quieter and they land in places CS owns.

1. Your health score goes blind

Ticket volume has been a health signal for years. A customer who files thoughtful tickets is engaged. A customer who goes quiet might be fine, or might be halfway out the door.

When the customer's admin is getting answers from an assistant, silence stops meaning anything. You can't tell the account that's self-sufficient from the account that's been fighting a misconfigured integration for six weeks and has quietly decided you're the problem.

2. Wrong answers become your product's reputation

When an outside assistant gives a confident wrong answer about your product, nobody files a bug. The customer just concludes the feature doesn't work, or the workaround is ugly, or your product can't do the thing it does perfectly well.

You never see the conversation. You only see the downstream effect: lower adoption of the feature, a mention in the renewal call that "the reporting never really worked for us," a competitor eval nobody told you about.

3. You lose the voice of the customer

Gartner's Keith McIntosh made a point in the 2025 research that's easy to skip past. Third-party resolution is fine, he said, as long as companies aren't "missing out on valuable customer information."

Every question that goes to an outside assistant is a question your product team never hears. The confusing setting, the missing integration, the onboarding step everyone gets stuck on. That used to show up in your ticket queue as a pattern. Now it shows up nowhere.

4. Expansion signals vanish too

Some of the best expansion leads come from support. A customer asks whether they can connect a second business unit, or how to handle more volume, or whether there's a way to automate a workflow your premium tier covers.

When that question goes to ChatGPT, the answer might be a clever workaround on the plan they already have. Fine for them. Invisible to you.

What to do about it

You can't drag customers back to your chatbot, and I'd stop trying. The better move is to accept that the support conversation happens in places you don't own and make sure the outside guy is reading your best material.

Treat your knowledge base like a product, because it's now an API

McIntosh's line from the 2025 research is the one I'd tape to the wall: "Third-party platforms, including those that use AI, still rely on the knowledge content that organizations produce."

Your help center is the training data and the retrieval source for every assistant your customers use. That changes how you should run it:

• Give every article an owner and a review date. Stale content is the main way assistants end up confidently wrong.
• Mark deprecations loudly and in plain text. "This setting was removed in version 4.2. Use X instead" is something a model can pick up. A grayed-out banner image is not.
• Write answers, not tours. An article that opens with the answer and then explains it gets quoted correctly far more often than one that buries the answer in step nine.
• Think hard before putting help content behind a login. If the assistant can't read it, it'll fill the gap from forums and old blog posts.

Make your docs readable by machines on purpose

Mintlify ran a 2,400-run benchmark with coding agents across 20 docs sites. With plain HTML docs, agents averaged 2.23 dead-URL errors per retrieval request. Adding a link to an llms.txt file (a plain-text map of your docs built for AI tools) dropped that to 0.11.

You don't need to be a developer tools company to borrow the idea. Offer clean markdown versions of key articles, publish a sitemap built for AI tools, and keep your URLs stable. When the customer's assistant can find the right page on the first try, the customer gets the right answer.

Let the customer's AI do the task, safely

That 74% figure is the opportunity. B2B customers want their assistant to get things done. So give it a sanctioned way to act: documented APIs for common admin tasks, and where it makes sense an MCP server or similar connector, with clear permission scopes.

Gartner's own advice points the same direction. Keller argues for "conversational, action-oriented experiences, rather than treating GenAI as a standalone chatbot." For a B2B vendor, the most action-oriented experience you can build might be one that works inside the assistant your customer already uses.

Run a monthly "what does the AI say about us" audit

This is cheap and a little humbling. Pull your top 25 support questions from the last quarter. Once a month, ask each one in ChatGPT, Claude, Gemini, and Copilot. Score the answers: correct, outdated, wrong, or made up.

Every wrong answer becomes a content ticket. Track the score over time like you'd track CSAT. If you're only going to do one thing from this article, do this one.

Add one question to every ticket and every CSM call

You can measure the off-property journey if you ask. Add a field to your ticket form: "Did you try anything before contacting us? (search, an AI assistant, the community, a colleague)." Train CSMs to ask in check-ins: "When you get stuck, where do you go first?"

Within a quarter you'll know how much of your customers' support life you can't see, and which accounts lean hardest on outside help.

Rebuild the health score around outcomes, not contact

If tickets no longer mean what they used to, stop weighting them so heavily. Lean on product signals that hold up no matter where the customer got help: adoption of key features, integration health, error rates on customer-built workflows, and whether the admin who set things up is still there.

A spike in API errors on custom integrations is a decent proxy for "an assistant wrote some bad configuration." It's a signal you already have and probably aren't watching.

Re-check the chatbot budget before you renew it

None of this means your in-product assistant is worthless. It can do things outside tools can't, like see the customer's account, configuration, and history. That's the case for it: context the outside guy will never have.

But if your bot mostly answers general questions that ChatGPT answers just as well, you're paying to compete with a free tool your customer already prefers. Before the 2027 budget locks, ask what your assistant knows that a public model doesn't. If the honest answer is "not much," move some of that spend into content quality and machine-readable docs.

The honest limits

A few caveats, so nobody oversells this internally.

The Gartner customer data covers both B2B and B2C, and most of the headline numbers blend the two. The 74% task-completion figure is B2B-specific, but the 3x ratio isn't broken out. Your customer base might skew more or less toward outside tools than the average.

The Mintlify numbers come from developer documentation, which is the most AI-saturated corner of support there is. An HR platform or a logistics tool will see less agent traffic. Probably still more than you'd guess, though.

And some of this shift is good news. Customers who solve their own problems quickly are happier customers. You don't need to win every interaction back. You need the answers they get elsewhere to be right, and enough visibility to still know which accounts are struggling.

What to do this quarter

1. Pull your top 25 support questions and run the first AI audit this month. Share the scorecard with product and docs.
2. Add the "what did you try first?" field to your ticket form.
3. Assign owners and review dates to your 50 most-visited help articles, starting with any that cover deprecated features.
4. Ask your docs team what it would take to publish markdown versions and an llms.txt file.
5. Review how heavily your health score leans on ticket volume, and add at least one product-outcome signal.
6. Before 2027 planning closes, write one paragraph on what your support chatbot knows that ChatGPT doesn't. If you can't fill the paragraph, that's your answer.

Your customers didn't stop needing help. They just stopped walking up to the counter. The companies that hold on to their renewals will be the ones that go out to the parking lot, find out what the guy by the truck is saying, and make sure he's reading from the right manual.