Pick the smaller number, then earn the bigger one

Forrester expects customer service quality to dip in 2026 as companies scale AI and bump into operational realities they hadn't planned for. They go further: they think a third of brands will erode customer trust with self-service AI that shipped before it was ready.

That's a strange forecast for a year when the technology has never been better. The models aren't the problem. What I see, over and over, is something simpler and more human: ambition applied in the wrong order.

The 70% problem

There’s a pattern. A team comes in and says we're going to hit 70% resolution right out of the gates. Full volume, every channel, day one.

I understand the instinct. There's real pressure from above to show a number. But look at what's being proposed. Before you have any real conversation data, you're pointing an agent at your entire customer base and asking it to be excellent immediately. You don't know what your customers actually ask, in what words, in what order, with what edge cases. You don't know where the agent will struggle, because it hasn't struggled yet.

The benchmarks show how upside down that is. Most deployments start somewhere around 20% to 40% containment and climb into the 70% to 90% range as they mature. A 70% target on launch day is someone else's endpoint, borrowed and moved to the front of the timeline. You've skipped the part where you learn.

What's actually at risk

Support raises the stakes on this more than almost any other function, because the thing you're experimenting on is the customer relationship itself.

When a support experience goes badly, you don't just lose the ticket. You spend down trust built over years, in seconds. You may well be chasing a cost reduction, but cost savings paid for out of revenue is a bad trade, and it's a trade most cost-focused business cases never price.

Klarna is the case study everyone in this industry now knows. They went early and hard, replaced an enormous amount of human support capacity, and posted remarkable efficiency numbers. Then satisfaction slipped, and they walked a good deal of it back.

That's a sequencing failure, not an indictment of AI. They turned the agent loose on everything at once, including disputes and high-stakes money questions, before the system had a reliable way to know what it shouldn't handle.

Know what your agent shouldn't do yet

That's the discipline I'd put at the center of any AI CX plan: knowing what your agent can do, and what it shouldn't do right now.

The customers I've watched succeed almost never launch the full plan in month one. They don't go live on five channels. They pick the highest-volume, most predictable tasks, the questions they already know the answers to, and they do those genuinely well. Then they iterate, and each expansion is earned by evidence from the last.

It looks slower on a roadmap. It is dramatically faster in practice, because every step compounds instead of backfiring. And notice the standard at each stage: did the customer get a faster answer and a better one? If a new use case can't clear that bar, it isn't ready, and shipping it anyway is borrowing against your brand.

The bounded start is also what makes the data flywheel spin. Real conversations tell you where to point the agent next. You cannot reason your way to that list in a planning doc. You can only learn it from customers, which means being in production, narrowly, doing something well.

Why this requires ownership

An iterative approach only works if you can actually iterate. Find a gap on Tuesday and you need to fix it on Tuesday: see the conversation, understand the decision the agent made, change the logic, ship. When your agent is a black box that someone else operates, that loop stretches from hours into weeks of change requests, and "start narrow, expand on evidence" stops being available to you. So you do the opposite. You launch big, because you only get a few swings at it.

Ownership is what makes restraint viable. It's what lets you start small without being stuck small.

If Forrester is right and 2026 is the year service quality dips, I'd bet the companies that avoid the dip will be the ones that were honest about what their agent could do on day one, kept a tight grip on the loop between seeing a problem and fixing it, and let customers, rather than targets, decide what to automate next.

Pick the smaller number. Earn the bigger one.