Keep One Thing Manual

A woman I've known for years runs a small commercial insurance brokerage. Eleven people. She's good at it in a way that's hard to describe until you watch her work.

About eighteen months ago she started routing her renewal reviews through AI. Pull the policy, summarize the coverage, flag the gaps, draft the client memo. It worked. It cut a two hour job to about twenty five minutes and the memos were arguably better, because they were more consistent than the ones she wrote at 9pm on a Thursday.

Last spring she was on a call with a manufacturing client and something felt off about an exclusion buried in their property policy. Not wrong exactly. Off. She couldn't articulate why for about ten seconds, and then she could, and it turned out to be a genuine problem that would have cost that client somewhere north of four hundred thousand dollars if it had gone unnoticed into a claim.

She told me afterward that the thing that bothered her wasn't the near miss. It was that she wasn't sure she'd have caught it a year later. Because the ten second gap between feeling off and knowing why had gotten longer, and she knew exactly why it had gotten longer. She hadn't read a full policy end to end in months.

That's the thing I want to talk about, and it's not a warning about AI. It's a question about what you're building.

Expertise has two layers and only one of them is visible.

The visible layer is output. The memo, the diagnosis, the estimate, the design, the recommendation. That's what you're paid for and it's what your clients see, so naturally that's what everybody automates first. It's also the layer where automation delivers the fastest, most obvious win.

The invisible layer is pattern recognition, and it's built almost entirely as a byproduct of producing the visible layer the slow way. You don't learn what a bad policy feels like by reading summaries of policies. You learn it by reading four hundred policies, most of them fine, until the shape of fine is so deeply installed that not fine sets off an alarm before you can explain it.

Here's the uncomfortable structure. The slow work builds the judgment. The judgment is what makes you worth more than the tool. And when you automate the slow work, you keep the output and you quietly stop paying the tuition on the judgment.

Not immediately. That's what makes it hard to see. Your judgment doesn't fall off a cliff in month three. It plateaus, and then it very slowly gets stale, and stale judgment feels identical from the inside to sharp judgment right up until the moment it doesn't.

I want to be clear about that, because the version of this argument you usually hear is romantic nonsense about craft, and it's usually written by someone who doesn't have payroll.

She should absolutely keep automating the renewal summaries. Twenty five minutes instead of two hours across eleven people is not a marginal gain, it's the difference between a business that scales and one that's capped by her personal hours. Giving that up to preserve some abstract notion of skill would be a genuinely terrible trade.

The move isn't to automate less. It's to be deliberate about which single capability you refuse to hand over, and then to actually protect it, on purpose, on a schedule, at real cost.

Because right now most people are making that choice by accident. Whatever was easiest to automate first got automated first, and nobody stopped to ask whether the easy one was load bearing.

There's one question that gets you most of the way there.

What do clients pay you for that they could not tell was wrong?

Sit with that one. Most of what you do, a client can evaluate. They can tell if the report is late, if the copy reads badly, if the site is ugly, if the numbers don't add up. That's the visible layer and it's the layer that's being commoditized fastest.

But somewhere in your work there's a judgment call your client is structurally unable to assess. They hired you specifically because they can't. The insurance client cannot evaluate whether an exclusion is dangerous. That's the entire reason the broker exists.

That capability is your actual product. Everything else is packaging. And it's the one you should think very hard about before routing through a machine that will do it faster and about eighty five percent as well, because the fifteen percent is where your whole pricing power lives.

Three more questions that sharpen it.

What did you used to do weekly that you now do monthly or never? Frequency decay is the earliest signal and it precedes skill decay by a long way.

Where do you still get surprised? Surprise means you're operating at the edge of your model of the world, which is the only place learning happens. If nothing has surprised you in your core work in six months, you've stopped building.

What would you have to relearn if the tools went away for a quarter? Not "would you struggle." Everyone would struggle. Would you have to genuinely relearn it?

Not much, honestly, which is why people don't do it. It's easy to skip precisely because it's small.

Pick a cadence and a volume. She landed on two full policies a month, read end to end, cold, before any tooling touches them. Roughly three hours. She writes her own read first, then runs the automated pass, then compares the two and writes down every place they diverged.

That comparison step is the whole thing. Reading slowly is nostalgia. Reading slowly and then measuring yourself against the machine is calibration, and calibration is what tells you whether the gap is widening or closing.

Track one number. She tracks divergences per policy. When it trends up, either she's drifting or the tool is, and either answer is one she needs. In her first four months it went from one to five, which alarmed her, and then she looked at the divergences and found that three of the five were her catching things the summary flattened. So the number going up was good news. That's the kind of thing you only learn by actually keeping the log.

Put it on the calendar as a recurring block and treat it as unmovable. If it's the thing that gets bumped when the week gets busy, it will be bumped every week, because every week gets busy. If it helps, block it in the same tool you use for everything else and let it fight for the slot like real work does. I track mine in Rize alongside billable time specifically so I can see when I've quietly stopped.

Three hours a month against a capability that's the reason your business charges what it charges. That's not a hard case to make. It's just an easy one to forget.

Two failure modes here and they're both common enough to name.

The first is choosing something nostalgic instead of something load bearing. Plenty of people pick the part of the work they simply enjoy most, which is a fine reason to do a thing and a bad reason to call it strategy. If you love writing proposals and you protect proposal writing, that's a hobby with a business justification stapled to it. The test is not what you'd miss. It's what your clients cannot evaluate.

The second is picking too many. Everyone's first instinct is to protect three or four capabilities, which sounds thorough and is actually a way of protecting none of them. Three hours a month on one thing is a practice. Three hours a month split across four things is a gesture. Pick one. You can revisit it next year.

There's also a version of this that applies to your team, and it's the harder conversation. If you have people whose job used to be the slow work and is now reviewing machine output, they are not building the judgment you built. They're being paid to approve things they don't yet have the experience to evaluate, which is a fine arrangement for eighteen months and a real problem at year four when you need one of them to be able to do what you do. The same fix applies. Somebody's calendar needs a recurring block where they do it the long way.

Here's the part that decides whether this matters.

The tools are getting better on a quarterly cadence. Prices are falling. The gap between what you can do and what someone with no experience plus good tooling can do is compressing in every direction that can be described in a process document.

What doesn't compress is the judgment that's built from volume and can't be written down. That's not because machines can't get there. It's because your specific version of it, built from your specific four hundred policies and your particular market and the six times you got burned, isn't in anyone's training data.

Three years out, the operators who are still expensive will not be the ones who automated the least. They'll be the ones who automated aggressively everywhere except one place, and who can still do that one thing at a level nobody can rent.

The operators who are in trouble won't have made a bad decision. They'll have made a hundred good small ones without ever asking which one was load bearing.

Twenty minutes. Write down the three things you're best at professionally.

For each one, answer honestly: could a competent person with good tools and no experience produce something a client would accept?

Cross off every yes. That's the packaging.

Whatever survives is your product. Look at it and ask when you last did it slowly, by hand, all the way through.

If the answer is longer ago than you're comfortable saying out loud, put three hours on next month's calendar before you close the laptop.

Not because the machines are coming for you. Because the version of you that's worth hiring in 2029 is being built or not built right now, in the hours you're not thinking about.

See you tomorrow.

Jordan Hale
The AI Newsroom