Tails Rule Everything Around Me

Been talking to a lot of folks about how they’re actually using AI. Not the deck version. The real version.

Two patterns show up everywhere.

Path one: the cost of chasing an opportunity dropped, so they’re chasing more. Proposals, pitches, customer segments they used to ignore, deals they used to pass on.

Path two: the cost of automating internal stuff dropped, so they’re automating more. Reports, reconciliations, onboarding flows, status updates.

Both feel obviously good. More opportunities, fewer manual tasks. Productivity has to go up, right?

Except the productivity numbers refuse to cooperate. Q1 2026 was 0.8%. Adoption is something like 91% of US businesses. Either the gains are coming and we just need patience, or something about how these paths play out is worse than the pitch.

I think its the second thing.

Path One: More Shots on Goal

The implicit assumption: each new shot has positive expected value, so more shots = more wins. Classic lottery ticket thinking.

It only works if the opportunity distribution is symmetric or right-skewed. In most mature markets, it isnt. It’s heavily left-skewed. The good opportunities aren’t hidden, they’re already being pursued by the people closest to them. What’s in the “discarded” pile got discarded for reasons. Bad unit economics. Adverse selection. Regulatory landmines. Channel conflict.

So when AI drops sampling cost to zero, you’re not sampling uniformly from “opportunities.” You’re specifically sampling from the set that was previously filtered out. That set is, by construction, enriched for bad bets.

This is adverse selection on your own pipeline.

Three things make it worse:

The cost asymmetry. Writing the proposal got cheap. Pursuing the opportunity didn’t. Onboarding, support, the operational tax of a bad customer, the management attention pulled away from your good accounts. None of that moved. You cheapened the trigger for incurring all the expensive downstream costs, which is exactly the wrong thing to cheapen. In lending terms: you’d never want origination cost to drop without underwriting cost dropping proportionally. You’d just originate more bad loans faster.

The slop tax. If proposals are free to write, they’re free to send. Everyone gets 10x more. Buyers build filters. Sales cycles lengthen. Win rate per proposal collapses. Every channel AI touches first gets cheaper to use and then becomes worthless because everyone uses it. Cold email is already there. Job applications are already there.

Polluted training data. When you pursue more marginal opportunities your win/loss signal gets noisy. You can no longer learn from your pipeline because its composition shifted under you.

The question to ask before expanding: what was the old filter actually filtering on? If it was filtering on writing cost being too high, great, expand. If it was filtering on the opportunity being bad, AI doesnt fix that. It just lets you fool yourself faster.

Path Two: Automating Internal Stuff

This one is sneakier. Manual work bad, automated work good. Right?

But here’s the question nobody asks before they automate something: why was this task being done at all.

A lot of what’s getting automated right now is some flavor of bullshit work. It existed because the org needed someone to be seen doing it, or because liability required a human signature on something performative, or because an old workflow ossified into a quarterly report that nobody reads.

You automate a task like that, here’s what happens. You save time. You don’t save money structurally because the headcount stays. You don’t produce more value because the task’s outputs were never the bottleneck for anything.

Worse: automating bullshit work legitimizes it.

The task now has an AI workflow, an integration, a dashboard, a Slack notification. It becomes load-bearing infrastructure. Harder to kill than when a person was doing it manually, because now killing it means deprecating a system. You took a soft target and hardened it.

Drucker had a line: there is nothing so useless as doing efficiently that which should not be done at all. AI is an incredibly powerful efficiency tool, which means its an incredibly powerful tool for entrenching things that shouldn’t be done at all.

A useful test before automating anything: if this task stopped happening tomorrow, would anything actually break. Not “would someone notice.” Would the business outcome it supposedly serves degrade in a measurable way. For a huge fraction of what’s being automated right now, the honest answer is no.

“But What If I’m Wrong About All This”

Worth taking seriously. Two real objections.

Measurement is genuinely weak. GDP and productivity stats were built for an economy of physical goods. They don’t capture free consumer surplus (you got a free tutor, a free coding partner, a free therapist), or quality improvements (your emails are better, your code has fewer bugs), or headcount you avoided hiring during growth. The Solow paradox of the 80s partly resolved because measurement caught up, not because firms got better at using PCs. The current paradox might be the same story. If so the gains are real and just invisible.

The old filters maybe weren’t that good. I assumed discarded opportunities were negative-EV bets that someone with judgment passed on. Honest version: most corporate “no” decisions aren’t crisp judgments. They’re “not worth the writeup,” “we don’t have bandwidth,” “my boss won’t like this.” If the old filter was filtering on evaluation effort rather than opportunity quality, then expanding is a genuine unlock, not adverse selection. AI doesn’t just lower sampling cost. It lowers evaluation cost too. Which kind of “no” was it? Most companies haven’t done the work to know.

And the base rate is real. Every general-purpose technology gets the “is it really productive though” treatment in its first decade. Electricity. PCs. The internet. The skeptics are usually wrong eventually. Betting against general-purpose tech on early aggregate stats has a roughly 100% loss rate.

So Where Does That Leave You

Maybe the productivity story resolves the way every prior one did. But the resolution always required the same thing: firms reorganizing around the new tech rather than bolting it onto existing workflows. The companies that won the IT productivity wave asked “what does our business look like if we redesigned it around this,” not “how do we use this to do more of what we already do.”

Both of the paths I described are versions of doing more of what you already do.

The question I’d ask any team, including my own, is two questions:

What were we not doing that we should have been, and is AI the thing that changes that math.

What are we doing that we shouldn’t be, and is AI giving us an excuse to keep doing it.

The companies that get real productivity out of this stuff will be the ones willing to answer both honestly.

My guess is that’s a much smaller set than the adoption numbers suggest.

Get the right tail.