Recently, a startup leader I was talking to went silent.

10 minutes into her exciting AI updates, she paused in a moment of real vulnerability.

“Are we genuinely speeding things up & making ourselves more efficient? Or is there a bit of smoke & mirrors around this…”

“We're just creating different types of work to get to the same end result — but maybe less good quality in some cases.”

She had deeper worries about their culture,

“The fun has kind of been zapped out… People leave each day feeling more exhausted, maybe less fulfilled”

This is very normal.

Almost every company that has adopted AI in the last few years is facing these confusing realities:

These threads of personal discombobulation are confirmed by the data.

An Upwork 2026 survey found that although 3/4 of SMB leaders saw increased productivity with AI, most saw gains below 25%.

It seems these leaders were "largely operating on conviction while they wait."

Worse, a WebMD study revealed that people who strongly believe that AI makes them more productive were 4.5X likelier to burn out than their peers. Meaning the people most convinced on AI are the ones breaking.

So why is that?

Why is our great AI rollout giving us meh results & making our people feel uggh?

As our Gen Z employees would ask, why is this so mid?

It comes down to 3 reasons.

We suck at measuring our own AI productivity.

A 2025 METR randomized controlled trial proved this. They had open-source developers work on their own repositories that they had maintained for years. Half could use AI, half couldn’t.

By the end, the AI-assisted developers estimated they’d been 20% faster.

In reality, their tasks took 19% longer.

Successive studies showed scores all over the map - ranging from 18% speedup to 4%.

One of the reasons for this discrepancy is silent task distribution.

Whenever an employee gets faster at their AI tasks, a huge chunk of that gain quietly spreads into other parts of their work — the tidy-ups, the extra drafts, the nice-to-haves that never used to be worth doing — they’re doing more of it & not noticing.

All of this proves one thing: even the most sophisticated analyses of self-reported AI productivity can’t be trusted.

Why? Because human beings are bad judges of their own AI-driven impact. Self-assessed competence is a misleading indicator in all these AI engagement surveys.

Add to that the fact that the person who measures AI impact in a company is often the person rolling it out, and this just further muddies the results.

1st Rule of AI Roll-out Club: You do not ask about AI directly

Every new AI initiative suffers from 3 types of leaks, but that’s not the real problem.

The real challenge is our inability to identify the right one at the right time.

Like learning any new complex tool, AI adoption has its own J-curve.

The risk here is that after a leader watches their team’s quality drop & struggle to get better, they render it a failed verdict and start over. Or just straight up cancel it.

Thus losing any potential gains they might have had, if they’d just stuck with it a little longer & helped their people wrestle this LLM beast into submission.

A recent developer study showed that AI cut time-to-push code by over 50%, while simultaneously increasing the reviewer’s time to review that code by 4.6X.

In the end, the whole process took more than twice as long.

This same dynamic plays out in so many parts of the company.

Content marketer & CMO. Recruiter & Hiring Manager.

The bottleneck that AI resolved just gets pushed to another part of the firm, usually downstream. And the further downstream a bottleneck gets pushed, the more expensive it becomes.

The reason this keeps happening is because almost every process in our company is designed for human-rate output. And AI changes the input rate, not the capacity downstream.

So if we want to fix this specific problem, nip it in the bud further upstream, redesign the pipeline and for God’s sake, don’t throw more people at it.

If the car industry gave birth to the seatbelt laws & airplanes gave rise to the FAA, then the AI verification tax is the cost of doing business in this brave new world.

We can’t escape it.

First, AI workslop is an ongoing nuisance. On average, it takes an employee 2 hours to vet & re-do someone else’s AI output while slowly resenting that person as well.

Next, knowledge decay sits silently in every team. This foundational AI work in the company’s operations is quietly pushing polished but wrong output throughout the org & that’s dangerous.

For example, an agent that generates a customer research deck for every new CRM prospect has a small flaw that no one notices. The agent only checks for the most common 2 pushbacks from its training, and misses all the customer’s actual biggest redlines. Now each customer roadmap is built on faulty assumptions.

The real questions here aren’t on how to avoid checking AI output. It’s

As speed increases, our genius slowly decreases.

Although many employees are discovering they’re doing their old work much faster, they’re also doing a lot of different work to maintain this new baseline, and that’s having 2 profound effects.

Employees are performing more & more tasks outside their zone of genius.

When an editor is spending more time reviewing other people’s writing than penning their own, they’ve lost the connection to the core activity that brought them happiness.

The skill that made them damn good at their job is now more automated & replaced with a task that they’re mediocre at, thus stealing their sense of competence & pleasure.

The same startup executive told me, “The joy that people used to find in their work has diminished.”

This doesn’t necessarily burn them out, but it does make them more checked out.

When correcting AI output, we’re often editing foreign work & trying to claim it as our own. And the deeper we go down that rabbit hole, the farther we go from knowing what’s good and what’s not.

After a while, we’ve not only lost sight of what success looks like.

Worse, we’ve dulled our instinct — the thing we were hired for.

So the end result is many employees hide their AI work to avoid getting asked about it.

Because if they actually had to explain it, they wouldn’t know where to begin & how it’s supposed to end.

When leaders keep misfiring on their AI initiatives & slowly gutting their people’s best work, while assuring everyone that the numbers are headed in the right direction

— that is systemic gaslighting.

If an organization continues down this path of drinking the AI Kool-Aid, it could lead them off a cliff.

It could not only lead to costly rollbacks of AI strategies & runaway vendor costs, but in the worst case scenario, this untempered computational FOMO energy could lead to 1 of 3 scenarios:

Each of these outcomes will leave irreversible damage on our brand. It will be pretty painful to come back from this.

And yet, this is a very real possibility for any of us if we don’t pay attention to the signs & admit there’s a problem.

So step 1 of fixing our AI strategy:

Admit we’re in a toxic relationship.