We Skipped a Step

I started my career making coffee.

Not in some romantic, third-wave pour-over sense. I was a seventeen-year-old in Ohio who tried and tried to get hired at Starbucks, finally got the job, and thought I might work there forever. Just over twelve years and three states later, 2004 to 2016, I'd been a barista, a training store specialist, a field implementation manager, a multiunit manager. What I'd really been doing, in every one of those roles, was learning how to build trust with a stranger in about three minutes, line to beverage in hand. That was the actual goal we set for ourselves. Hundreds of times a day. For over a decade.

People started telling me that was a superpower before I believed it myself. It took years to accept what they saw plainly, that building instant trust at scale isn't a personality trait, it's a trained capacity. And it's the thing I carried with me when I left Starbucks and found my way into customer success, because CS at its best is the same skill at a different scale. You're still in the business of making someone feel seen, helping them solve something real, being the person they actually want to call when things get hard.

I miss that right now.

Not in a nostalgic way, I'm not trying to go back. But somewhere in the last couple of years, under the pressure to do more with less and the promise that AI would make that possible, I've ended up spending more time staring at a screen than sitting across from a customer. Building trackers. Sending recaps. Trying to get a language model to produce something that sounds like me so I can move to the next thing on a list that keeps getting longer.

A few weeks ago I was in a room with a group of CS leaders at an AI native CS event in San Francisco. Someone on the panel said something that stopped me: AI should be giving us back the thing we're actually good at. Less time in front of the screen, more time with real customers. I believed it immediately, because it's right. It's the whole point.

But I also thought: we are nowhere near there yet. And I'm not sure we're being honest about why.

The promise is right

The panelist was Ziv Peled, Chief AI Officer and Chief Customer Officer at AppsFlyer, and his thesis is one I'd bet most CS people would sign onto without hesitation: AI should free CS professionals from the operational grind so they can spend their time on the part of the job that actually matters, which is the human being on the other side of the account.

I believe that wholeheartedly, and not as an abstraction. Human connection is what carried me through twelve years at Starbucks and every year since. It's why I do my best work in a room with a customer, helping them solve a real problem, not behind a screen managing a queue. If AI can hand that time back to CS professionals, it's not a minor efficiency gain. It's a return to the actual job.

Here's where it gets complicated. Peled recently added Chief AI Officer to his title on top of Chief Customer Officer, and in a separate conversation he laid out how he actually drives adoption inside his own org. It isn't gradual. It isn't optional. He turns off the old workflow and forces the switch, because in his experience people don't build new habits around a tool they can still avoid.

I sat with that for a while. The same guy telling the industry that AI should hand CS people back their time with customers is also the guy who believes you get there by taking away the option to fall back on the old way. Both things can be true. Forcing the switch might be exactly what closes a fluency gap fast. It might also be exactly how you end up with slop, if people get shoved into a tool nobody taught them to use well. I don't think Peled is wrong. I think he's further along in a tradeoff most of us haven't even clocked yet.

The gap nobody wants to name

Here's the problem. That promise assumes a level of AI fluency that most CS orgs simply don't have yet, and almost nobody is saying so out loud.

Not because people are lazy or resistant. Because building real fluency with a new tool takes time, structure, and practice, and most of us have been handed the tool with none of those things. We've been told to use it and to move faster, at the same time, with no space carved out to actually get good at it. So people are doing the only thing they can: producing more, faster, without always fully understanding what the AI produced or whether it's any good.

And this isn't a fringe problem. Something like ninety five percent of CS teams are already using AI in some capacity, most of it exactly this kind of individual, ad hoc use. Someone drafting an email. Someone summarizing a call. That's not the edge case anymore. That's the whole profession, right now, mid transition, mostly on its own.

What that gap produces

When low fluency meets high pressure, you get volume without judgment. An email that sounds like no one. A recap that could belong to any account, any CSM, any company. I notice it the moment I see it, and it's not that I discredit the person who sent it. But I also don't hold it in the same regard as something written on their own letterhead, in their own voice, with their own judgment behind it.

That's the part that worries me most, because the thing that makes a CSM irreplaceable is exactly the thing generic AI output can't fake: real knowledge of a specific customer, earned trust, a sense of what matters to this person at this company right now. When everything gets AI smoothed into the same texture, that signal disappears. The promise was that AI would make us more present. Instead, in a lot of orgs, it's making us more generic.

I want to be clear about what this is and isn't. It isn't a character flaw. Nobody wakes up wanting to send a hollow recap. It's what happens when a real skills gap meets an unforgiving mandate to do more with less, and nobody addresses the first thing on the way to demanding the second.

There's a second gap sitting right underneath this one, and it deserves its own piece rather than a rushed paragraph here. Ask most CS orgs who actually owns the quality of what their AI produces, who's accountable when a recap is wrong or a health score flags the wrong account, and you'll get a shrug more often than a name. That's not a fluency problem, it's a governance problem, and it's the quieter, scarier cousin of everything I'm describing here. Consider this the trailer.

Closing the gap is the actual work

If the golden opportunity is real, and I think it is, then fluency has to become a real competency that CS leaders build on purpose, the same way they'd build product knowledge or negotiation skills. Not a webinar. Not a tool rollout with a Slack announcement. Actual training, actual standards, actual time to get good at it before being expected to move fast with it.

And the time AI is supposed to free up has to actually get protected. If it just gets absorbed back into more tickets and more accounts, we haven't returned anyone to the customer. We've just made the treadmill faster.

I keep coming back to Peled's forced switch, because it's actually a useful test case for what closing the gap on purpose could look like, done right. Forcing adoption isn't inherently the problem. Forcing adoption with no training, no standards, and no space to get good at the tool first is the problem. The lesson isn't don't force the switch. It's don't force the switch and skip the part where you teach people to drive.

I still believe the promise. Human connection is what this whole profession is built on, going back to a seventeen-year-old kid learning to read a stranger in three minutes over an espresso machine. The technology to protect that, and even amplify it, is genuinely here now. But we have to close the fluency gap first, or we'll end up using the most human-centered technology we've ever built to become less human, not more.

One more thing, since it's relevant to everything above: I used AI to help write this piece. The ideas, the story, and the judgment about what belongs are mine. AI helped me structure it and get it onto the page faster. I think that's worth saying plainly, because the difference between this and the slop I'm describing isn't whether AI touched it. It's whether the author stayed the author the whole way through.