I use AI and agentic workflows in every part of how I operate. I genuinely do not remember how I worked before them, and I am not interested in a conversation about slowing them down.

Which is exactly why this needs saying: the fastest way to lose a customer is now available to everybody, at scale, on a schedule.

If the player above has not loaded yet, the talk is here on YouTube.

That is the talk I gave at CSX Week 2026. The argument is written out below, so you do not have to watch it to use it.

The Mechanism

Customer success is when your customers reach their Desired Outcome through their relationship with your company. Desired Outcome is two things, not one: the goal they need to hit, and the Appropriate Experience for them specifically.

Most people hear that and weight the goal. The experience half does more work than it gets credit for. Deliver the right experience without the result and they will stay a while longer, and give you more chances to get there. Deliver the result through an experience that is wrong for them and they will leave anyway, holding the outcome they wanted, because getting it cost them something they were not willing to keep paying.

AI changes the scale of this in both directions. It can deliver the appropriate experience to more customers than any team could reach before. It can deliver the wrong one at exactly the same speed.

With AI you can build a machine that proves to your customers you never understood them, and you can run that machine at scale.

Trap One: You Delivered Before You Listened

A company I work with runs a done-for-you service that only became possible because of agentic workflows. It starts with an intake form. Thirteen questions.

Everyone there already knew the form was a problem. Customers often skip it. When they do fill it in, they fill it in thinly, because people edit themselves when they write. The kickoff call was where the gaps actually closed, and where things nobody thought to put in a form surfaced.

Then they automated. Time to first value went from three weeks to before the kickoff call had even happened, which sounds like an unambiguous win until you see what the customer receives: work built on a form they know they answered badly.

The reactions write themselves. Why are we having this call if the work is done. Why did we have the call if you were not going to listen to me. Why did you let me fill that out so carelessly if you were going to build on it.

The fix was not to slow down to three weeks. It was to wait a few days. Take the intake form, the sales call, and the kickoff conversation, combine them, then produce. Three weeks to a couple of days is still an enormous improvement, and now the work reflects something the customer actually said.

Why "this looks like AI" is usually not about the AI

A related thing happens when customers see automated output and call it AI slop.

They are rarely pointing at em dashes or formatting. Those are fixable, and you should fix them. What they are reacting to is that the work reads generic, and generic is what thin context produces. Strip every tell you like; if the input was one half-answered form, the output will still read like it was written by someone who has never met them.

The AI did not fail you. The missing context did.

Trap Two: A Tool for Everything, a System for Nothing

I would rather hand a customer a tool than teach them a process, and building tools is close to free now. That is genuinely good, and it is where the second trap lives.

One CSM builds something in a no-code app. Another builds an artifact in Claude. Someone in ops builds a proper agentic workflow. Same underlying logic, three interfaces, three places it can change.

Then it drifts. Nobody can say which version is current, so some customers are running last quarter's logic. Then someone leaves, and the tool their customers depend on is still sitting in a personal account you cannot reach, update, or switch off.

The thing built for efficiency is now consuming time, and it has become a dependency you do not control. This is the concrete argument for a customer success organization having an ops function: not to slow building down, but to make sure what gets built belongs to the company.

Trap Three: Automating on Top of Garbage

A multiplier multiplies whatever you feed it.

Take a workflow that audits a campaign, builds a plan from the audit, and executes the plan automatically. Clean loop, real efficiency. And it never ingests anything that happened outside itself.

Meanwhile the customer said something on a strategy call about the leads not qualifying. Support has a pattern. Sales heard something on an expansion conversation. None of it reaches the audit, so the audit reads the campaign metrics, decides things went fine, and optimizes toward more of the same.

Trace it back far enough and the whole loop is running on the intake form nobody filled in properly.

Fast garbage is still garbage. It just arrives more often.

The correction is not complicated. If you record your calls, and you should, that intel is available. Email, support threads, QBR notes, the strategy conversation from three weeks ago. Feed the audit everything the customer has told you, not just what your own system produced last time.

Two Things to Take Into Work

If you wouldn't sign it, don't scale it. If you have reservations about what a workflow is producing, scale is the wrong response to those reservations. It does not improve the output. It delivers the same output to more people, faster.

If you can't say no, you can always say something. Plenty of people reading this cannot halt a process. Fine. You can still escalate, and escalation works when it carries proof instead of a feeling. Not "I think this customer is having a bad experience." Instead: this customer is threatening to cancel because we built their deliverable from a form they told us they did not complete, we ignored everything from the kickoff call, and here is what the account is worth. Put a number on what the inappropriate experience costs and the conversation changes.

The Actual Risk

None of this is an argument against AI. Go use it. Scale your delivery of the appropriate experience as far as it will go.

Just be clear about what is now possible in both directions. We have never before been able to prove to a customer that we do not understand them this efficiently, this consistently, or to this many of them at once.

So slow down to speed up. A few days instead of zero days. One conversation before the deliverable. One person who owns the tools. One workflow that reads everything the customer has told you.

Going Deeper on Each

Each of these has more in it than a keynote allows for, so they are written out separately.

If you want to find out where your own operation is exposed, I built a seven-question audit of exactly these failure points. Under two minutes, no email.