# Augment, Don't Replace

*The "replace your team with AI" pitch is wrong twice — and your P&L agrees with your conscience.*

Sreenath Pillai, Founder & CEO, Vertikals. October 2026 · 3 min read.

Part of [Field Notes](https://vertikals.ai/blog).

Canonical page: https://vertikals.ai/blog/augment-dont-replace

---

"Replace your whole team with AI" is the pitch of the moment. It's wrong twice. It's the wrong thing to do to people — you already feel that, and I won't belabor it. The part that gets left out of the pitch is that it also produces *worse results at higher cost.* The ethics and the economics point the same way, which happens rarely enough that it's worth stopping when it does.

## Augmented beats autonomous on quality

A skilled person steering AI catches the mistake on draft one, not draft nine. They bring the constraints that keep the model from wandering — "no, we don't support that, use the existing auth layer" — that an unsupervised agent would never know to apply. Most importantly, they can look at output that is fluent, confident, and wrong and *know* it's wrong.

That last skill is the whole game, because confidently-wrong is the one failure mode AI cannot detect in itself. Autonomy removes the exact check you need most, right at the point where it's hardest to notice it's missing.

## Augmented beats autonomous on cost — and nobody runs these numbers

The "replace them" pitch quietly assumes humans are the expensive line item and compute is free. Run it out and it's backwards.

Turn a fleet of agents loose with no one steering and watch where the tokens actually go: retries, dead-end paths, regenerated work, agents spawning more agents to fix what the first ones broke — mountains of fluent output, much of it thrown away. The compute was never free. It's just invisible until the bill arrives.

Now put one capable person in the loop. A single sentence of human steering kills a thousand tokens of wrong-direction generation before they're ever spent. The cheapest token is the one you never had to spend going the wrong way. The human in the loop isn't overhead sitting *on top of* your AI cost — the human is what makes the AI cost efficient in the first place.

Then there's rework — the line item that never makes it into the "look how cheap AI is" math. Output that's 80% right is often *more* expensive than doing it properly, because someone now has to find the 20% that's wrong, buried inside something that reads perfectly. Hunting a flaw in fluent, confident work is slow, senior, costly labor. You didn't remove that cost. You pushed it downstream and made it bigger.

## The unit of value is the pair

Stop treating "AI" and "humans" as two columns you trade off against each other. The unit that actually produces value is the human-plus-AI *pair*. Design for the pair: give every person real leverage, budget for the pair, measure the pair. A good operator with AI leverage beats both a lone human and a swarm of unsupervised agents — on quality and on total cost, with salaries, compute, and rework all counted.

## What you throw away when you throw away people

The team you'd replace is also the team that knows *why* things are the way they are — the context, the history, where the bodies are buried. That institutional knowledge is exactly what turns AI output from plausible into *useful*. Fire it and you don't just lose headcount. You lose the thing that was going to make your AI any good.

Don't replace your people. Hand them leverage. You get better outcomes, lower burn, and a team that's still standing when it actually matters. For once, the right thing and the smart thing are the same thing.
