Most Companies Use AI to Avoid Their Customers

In March 2025, Gartner predicted that agentic AI would autonomously resolve 80% of common customer service issues by 2029. One quarter later the same firm predicted that more than 40% of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Same analyst house, same technology, three months apart. Whatever you want to call the current moment, that’s what it sounds like from the inside.

Is the AI bubble about to burst?

Nobody honest knows, and a support team shouldn’t plan around the answer either way. The question you can act on is smaller: if AI spending corrects, what did you build that still works afterward? For most support organisations it isn’t the model, because you don’t control when that gets replaced. It’s whether a customer can get their job done with you. Most of the AI money in support has gone toward making that harder rather than easier.

The argument everyone is having

The spending is real and it’s enormous. Goldman Sachs’ baseline model puts AI capital expenditure at roughly $765 billion for 2026, and the largest hyperscalers are increasingly funding their share with debt rather than cash flow. Whether that adds up to a bubble is an argument for people with more market exposure than a support team has.

The more useful question is whether the projects are working, and that one has an answer. S&P Global asked 1,006 IT and line-of-business professionals how their AI work was going and published the results in October 2025. In a single year, the share of companies abandoning most of their AI initiatives went from 17% to 42%. Not trimming a project here and there. Walking away from most of what they had started.

None of which means the technology doesn’t work, and Gartner’s reversal didn’t claim it did. What it said was that a large share of deployments were “early stage experiments or proof of concepts mostly driven by hype and often misapplied.” Misapplied is the part that matters for a support team. It doesn’t mean AI failed. It means a lot of people pointed it at the wrong problem.

The bill your customers already got

While the industry argued about capex, customers were fairly clear about how they felt regarding the way AI was showing up in customer service.

Gartner found that 64% of customers would prefer companies didn’t use AI for customer service at all, and that 53% would consider switching to a competitor if they learned a company was going to use AI for support. Nobody in that survey was asked to evaluate a benchmark. They were describing an experience: a fast, confident, unhelpful answer from something that wouldn’t let them reach a person.

The pressure to ship more of it keeps climbing anyway. A Gartner survey of 321 service leaders fielded in October 2025 found 91% reporting pressure from their own executives to implement AI. So the people closest to the customer are being pushed hardest toward the thing customers said they didn’t want.

If a correction is coming, that’s the part that doesn’t get written off. Capex gets impaired on a balance sheet. A customer who learned that contacting you is a maze stays taught.

Two opposite ways to spend AI on a customer

Nearly all of the AI money in support has gone to one job: reduce the number of humans a customer reaches. Deflect the chat, defer the escalation, push contact rate down. Measured on cost per contact, it works. Measured on whether anyone got helped, it produced the 64%.

There’s a second job that barely gets funded, which is using the same technology to answer the question properly the first time, in the place the customer already looked.

The difference isn’t philosophical. It shows up in what you actually build. The first job puts a gate in front of your help center. The second one fixes what’s inside it: the article written once in 2023, with four screenshots of a UI you’ve since redesigned, that has never shown anyone where to click.

Gartner’s 2024 report on self-service (a measurement, not a prediction) drew on responses from 5,728 customers and found 73% try self-service but only 14% fully resolve their issue there. In 43% of the failures, the customer couldn’t find content relevant to their problem. That number is worth reading twice, because it isn’t a complaint about tone or hallucination. The content wasn’t there. No amount of agentic sophistication fixes an empty shelf. It just gets faster at telling you the shelf is empty.

Stocking that shelf is the unglamorous half of the job, so it’s worth being concrete about what one of those answers looks like. Below is a single help-center task, start to finish, generated from one walkthrough. It’s the whole video, not a highlight reel.

What actually survives a correction

Set the speculation aside and ask what a support org owns at the end of this cycle that it didn’t own at the start.

It isn’t the model, which gets replaced on a schedule you don’t set. It isn’t your vendor’s resolution rate, which is their number calculated on their definition. What you own is the content: an accurate, current, watchable explanation of how your product works. That keeps its value whether inference costs rise or fall, whether your agent vendor gets acquired, whether the category reprices.

Service leaders seem to have worked this out already. In that same 2026 Gartner survey, 58% said they aim to upskill agents into knowledge management specialists, explicitly citing the need for “accurate, continually updated content to support both AI systems and customer self-service interactions.” That’s a bet on the shelf rather than the gate, and it’s probably the least exciting line item in anyone’s AI budget.

Production cost was always the catch. Teams didn’t skip tutorial video because they doubted the format. They skipped it because a polished one cost thousands of dollars and weeks of calendar time, then went stale the day the UI changed.

That’s the catch the last couple of years of text and video AI progress quietly removed. Turning that into something a support team can actually use took real engineering, and we aimed it at the help center, because that’s where the gap is widest: the most questions to answer, and the least budget to answer them properly. A polished, narrated walkthrough now takes minutes instead of a production budget, which puts it in reach of the teams that were never going to have one.

What we’re not claiming

We generate video with AI, so a post about AI being misapplied owes you some precision about where we think we sit.

We aren’t claiming AI is more human than a person, because it isn’t. TechSmith’s 2024 research found that 87% of viewers prefer a real person over an animated character or AI avatar, and 90% have concerns about video content created with AI. We believe both of those numbers. It’s why the on-screen presenter in our product is optional and off by default. The narration is the product; the face is a setting.

The same research found something people skip past, which is that 75% are receptive to AI-integrated instructional video. Those findings sit together fine. People don’t object to a machine helping make something useful. They object to a machine standing between them and help.

It’s also worth being straight about what we replace. The alternative to an AI-generated tutorial was never a human-made tutorial. It was a text article nobody had time to write well. We aren’t displacing a person, we’re filling a gap where nothing was. And video doesn’t win at everything: a 2021 meta-analysis of randomized trials (Noetel et al.) found that adding video to existing material produces strong learning gains, which is a narrower claim than the one this category usually makes. Keep your reference tables. Film the multi-step tasks.

Frequently asked questions

Is the AI bubble going to affect customer service teams?

The macro question is unresolved, but there is measured data on how AI projects are actually going. S&P Global Market Intelligence surveyed 1,006 IT and line-of-business professionals and found the share of companies abandoning most of their AI initiatives rose from 17% to 42% in a year, with reported positive impact falling across every objective measured. Teams whose support strategy depends on a vendor’s resolution rate carry that exposure. Teams that invested in their own help-center content do not.

Why do customers dislike AI in customer service?

Gartner found 64% of customers would prefer companies didn’t use AI for customer service at all, and 53% would consider switching to a competitor over it. Wanting fewer tickets is a reasonable goal and we want the same thing. The problem is the method: most AI spending in support went toward making a person harder to reach, which lowers contact rate without resolving anything. Customers don’t experience that as self-service. They experience it as a wall.

Can AI make customer support more human rather than less?

It depends on what you point it at. Used as a gate, it puts distance between you and the customer. Used to produce content, it fills the gap that made self-service fail in the first place. Gartner surveyed 5,728 customers and found 73% try self-service but only 14% fully resolve there, with 43% of failures caused by not finding relevant content. Generating clear tutorials for questions that previously had none closes distance instead of adding it.

What should a support team invest in if AI spending corrects?

Assets you own rather than capabilities you rent. Models get replaced on schedules you don’t control, and vendor resolution rates are defined by the vendor. An accurate, current explanation of how your product works keeps its value regardless. In a 2026 Gartner survey of 321 service leaders, 58% said they aim to upskill agents into knowledge management specialists, citing the need for accurate, continually updated content.

The bottom line

If the AI trade corrects, the companies that come through it won’t be the ones that timed it correctly. They’ll be the ones whose customers could still get an answer.

Spending money there is deeply unfashionable right now, which is most of the argument for doing it. A help center where the top forty questions each have a clear, current, ninety-second answer is not an AI strategy anyone writes a headline about. It’s just what your customers were asking for the whole time.

That’s what Brevid does. You walk through the task once in your own product, and we turn it into the tutorial video, with narration, cursor, zooms and captions. Your walkthrough stays saved, so when the product changes you hit regenerate and Brevid replays it against the updated UI. Refreshing an article’s video costs you a click rather than a reshoot, and it happens when you decide it should.

When the steps themselves have changed enough that the old path no longer applies, you walk through the new flow once and that becomes the video. A few minutes of your own time, with no crew to book, no editing suite, and nothing to film twice.

We’ve written more about why the articles themselves stopped working in why your help center articles aren’t working, and about the format question in video vs. text.