A few days into working with a group of our corporate clients, one of the participants said: “I don’t want AI to just agree with me. I want it to push back on my ideas and act as a thought partner.”

That comment wasn’t a one-off. As the group built AI agents tuned to their specific roles, the same want kept surfacing: not a faster way to crank out output, but a critical thinking partner that would interrogate their reasoning instead of waving it through. Everyone in those rooms had already felt the alternative, an AI that calls your half-formed idea excellent and your thin reasoning sound. It feels great for about four seconds, until you ship the work and find out the AI was just telling you what you wanted to hear.

That instinct points at something bigger than how one AI talks to you. Underneath it is a more basic question: is AI making people sharper, or just leaving them where they were? A sycophantic AI is what that looks like for one person: it feels like help, but it never challenges your thinking, so you come away no better than you started. Companies answer that same question at full scale, in the choice between augmenting their people and automating their work.

Automating means handing the task off and taking the output. Augmenting means using the tool to build up the people doing the work, so they make better calls than they would have alone.

Researchers from Oxford and Stanford writing in Harvard Business Review put the fork plainly: leaders can chase the bottom line by automating and reducing headcount, or grow the top line by augmenting what their people can do. Most companies treat which one they get as a feature of the technology. But they’re choosing it, one task at a time, without stepping back to ask the bigger question: are we using AI to replace the work, or to develop the people who do it? Anthropic’s Economic Index, which tracks how AI actually gets used rather than how people say they use it, found augmentation ahead of automation in early 2025, roughly 55 to 41. By mid-2025 automation had pulled ahead for the first time; by early 2026 augmentation had edged back to just over half. The balance moved meaningfully inside twelve months, which means the habits are still forming. The companies that set deliberate norms now are the ones most likely to come out ahead.

When an organization defaults to handing the thinking off, the cost doesn’t land on the task. It lands on the people. For a while the evidence was soft, mostly workers reporting they thought less when they leaned on AI. A Microsoft and Carnegie Mellon study put numbers to it: the more workers trusted the AI, the less critical thinking they did. The 2026 research made it causal. Controlled experiments found that AI assistance reduced people’s persistence and degraded their independent performance, and the researchers spelled out the cause: an AI optimized to be immediately helpful can erode the long-term capability of the person it’s helping.

This isn’t new in principle. In 1983, Lisanne Bainbridge described the “irony of automation”: automate the routine parts of a job and you strip away the everyday practice that keeps human judgment sharp, so when judgment is finally needed, the edge is gone. Companies are now running that experiment on knowledge work. EY’s workforce survey found that while about 88 percent of employees use AI at work, only around 5 percent use it in any advanced way. The rest treat a thinking partner like a vending machine, which EY estimates leaves up to 40 percent of the available productivity on the table. The tool is deployed; the capability to use it well never got built. That’s the cost of automation-by-default: you pay for the tools and risk your people growing more dependent and less capable, all while capturing only a fraction of the value.

I put a related question to that same cohort: should the people they hire next year still learn to do the work they’re now handing to AI? It set off more debate than we could settle, with good arguments coming from every direction. But one thread ran underneath most of them: if nobody coming up still knows how to do the task, nobody can tell when the AI gets it wrong.

So what does the other path require? An AI that challenges you instead of agreeing. Go back to the person who wanted pushback. The reason that instinct is right isn’t only that an agreeable AI is unhelpful. It’s that an agreeable AI is convincing. A study presented at the CHI conference looked at people consulting AI privately on contested questions. When the AI mirrored their existing view, they came away more confident and less open to other perspectives. Agreement didn’t refine their thinking; it locked in what they already believed. The same research makes the constructive case too: disagreement isn’t a flaw to design out, it’s a resource to design in. Set the AI up to take a real opposing position, and people stay open instead of digging in.

This is what we built with those clients, and it’s the part any team can copy. For anyone working inside Claude, we set up a skill, a reusable instruction set, that turns the model into a challenger by default: its job is to find what’s weak, argue the opposite, and name the assumption you’re leaning on without noticing. For anyone on another tool, the same thing works as a prompt you paste into any conversation: don’t validate, challenge; argue the position opposite to mine; tell me where my reasoning is thin. The wording matters less than the permission. You’re telling the tool that disagreement is expected. That’s augmentation made concrete: not a vaguer, gentler automation, but a deliberate decision to make the AI improve the person instead of replace the thinking.

Here’s what it looks like when it lands. One of the leaders in that group got visibly excited about a possibility he hadn’t expected. The agents his team built could take over the grinding parts of data analysis, which meant he could move the people who’d been producing those reports into roles where they spend their time on strategy, interpreting the data instead of just generating it. He wasn’t looking at AI as a way to shed those analysts. He was looking at it as a way to promote them. That’s the whole distinction in a single decision. Yes, a task got automated, but the point of automating it was to free a person for work where their judgment matters more. Automating a task to eliminate a person shrinks the company. Automating it to elevate a person grows it.

Most leaders watch adoption climb and read it as progress. The number underneath says otherwise. ManpowerGroup found that even as regular AI use rose 13 percent over the year, workers’ confidence in the technology fell 18 percent, because people were handed powerful tools with no training and no context. Usage is easy to count and easy to celebrate. It says nothing about whether anyone is getting better at the work, and a workforce leaning on AI more while trusting it less isn’t a workforce growing sharper.

That gap doesn’t close by buying more tools. It closes by teaching people to use the ones they have as thinking partners that sharpen judgment instead of softening it. We call the principle Author Before Tool. Your people stay the authors. The AI is how they raise the level of the work, not the thing they hand the work to.

None of this makes automation the enemy. Plenty of routine work can be handed off, and the smartest companies will automate it where the work genuinely doesn’t need a person’s judgment. The difference is what they do with the time it frees: the companies that win the next few years won’t be the ones that automated fastest, but the ones that turned the savings into sharper people, because someone decided, on purpose, to stop handing the thinking away.