September 16, 2026, 13 min read

AI Augments, It Doesn't Replace

Anti-hype, not anti-AI: AI can't replace leaders, coaches, or the work of developing them — its job is to augment. From someone who builds with it daily.

AI has become a topic that is taking over the workspace. It feels like every other HBR article is about AI these days, every third post on LinkedIn is about AI (I hear the irony in all of this as I post it) and it has taken over social media, news lead lines amongst other things. In my world specifically there’s some pretty exciting things around AI.

What are the ways our work gets to be more efficient? What are the latest hacks around making work easier? How can we incorporate it to do X Y or Z

And — there’s the not so fun half of things. In my world specifically… Is AI going to flatten the org chart and cut out middle management? Will it replace the coach? It’s created fear and frankly noise.

Here’s my take, as someone who spends about 8 hours a day working with an LLM. AI cannot and should not replace leaders, leadership coaches or leadership development. Its role is to augment, not replace. I want to be clear about what kind of claim that is, because it’s easy to misread — this is anti-hype, not anti-AI. I build with these tools every day. The argument isn’t that AI is weak. It’s that we keep mis-naming what the work actually is.

The leaders leg

This isn’t just noise in the feed anymore. It’s showing up in the job itself: in what you’re being told to use, and increasingly in how you’re measured. So the worry isn’t coming from nowhere.

Start with the leaders, because that’s where the fear is loudest. The middle gets cut. AI flattens the org chart, and the people managers stuck between strategy and the actual work are first on the block.

I get why that story lands. Middle managers really are overloaded. HBR made exactly this case in June: AI hits unevenly. Juniors get faster, senior partners get more reach, and managers get buried under new work (overseeing AI, checking its output, coaching people through it), all piled on the delivery pressure they already had. Gartner thinks a fifth of companies will use AI to flatten structure in 2026. We’re already watching people get thinned out. So this is a real pressure. I’m not waving it away.

But here’s where it goes wrong. AI didn’t start this fire. It’s pouring fuel on a fire that was already burning. The fire is that we’ve measured managers wrong for years. We judge them on production numbers (the team’s output rolled up into an OKR) instead of on the actual work of leading people. That pushes the pressure back onto the manager to move a number, and calls that management. AI doesn’t fix that. It makes it worse, faster.

And here’s the newest version of the same mistake. We’ve started grading managers on AI adoption itself. Meta, Google, and Amazon are now scoring employees on AI use, some building it straight into performance reviews. Think about what that measures. Not whether you led your people well. Whether you used the tool. It’s the same production-metric mistake in a 2026 costume.

And the answer to a broken measurement isn’t to delete the thing you measured badly. Ask people what affects their mental health and they rate their manager as high as their spouse or partner, and higher than their doctor or therapist (UKG surveyed 3,400 people on this). Pretending we don’t need a strong bench of people leaders is a cute idea at best. As that HBR piece put it: the question isn’t whether to thin this layer. It’s how to reinforce it.

The fallacy

So why is “cut the managers” so easy to believe? Because there’s a real, smart idea underneath it that’s getting misread.

Capturing good judgment is genuinely useful. If your best people have hard-won instincts, writing them down so the whole team can use them is a good move, and AI is getting better at helping. I’m not against that. Where AI still struggles is memory and context, and a lot of the hype (the fully-automated-workflow talk) quietly assumes you can write a clean rule for when the work is “done.” Sometimes you can. Often you can’t.

Here’s the leap that doesn’t hold: “we can capture judgment, so we need fewer managers.” That only works if a manager is basically a work-checker and a rule-enforcer. If that’s the whole job (inspect the output, apply the policy, flag what’s wrong), then sure, automate it. AI is great at that.

But that’s a thin picture of the job, and it always was. In my own work I call managing three modes, not one: Do, Direct, and Develop (that’s 3D Management). Do is the hands-on work. Direct is running the work: setting direction, calling priorities, making decisions, and yes, checking quality. Develop is growing the person doing the work, building what they’ll need before the job demands it. The “work-checker” manager is one thin slice of one mode. AI can take that slice. It can take some of Do, it can’t take most of Direct, and it can’t touch Develop. That was the actual job all along.

The reach ladder

So let me get specific about where AI actually reaches, because “augment, not replace” isn’t a slogan. It’s a claim about altitude.

Picture three rungs.

The bottom rung is measuring whether the work is good and whether it got done. AI is genuinely capable here. It can check output against a standard, catch what’s off, confirm the thing shipped. This is the checking work, and it’s fine to hand it over.

The middle rung is improving how the work gets done, the efficiency layer. AI is weaker here, but climbing. Think automations, and making automation accessible to people who could never build it before. Memory and context are its real limits, and both are getting better, so I expect this rung to firm up over time. Give it that.

The top rung is developing the people and the conditions they work in: engagement, psychological safety, the sense that it’s safe to speak up and worth trying. This is where AI can’t climb, and not because the tech isn’t there yet. It’s a different kind of work. You can’t generate psychological safety with a model. You can’t make someone feel seen by automating the seeing. AI can remove a check-and-balance, sure. It can’t remove people leadership, because the top rung was never a checking problem.

Notice the ladder only runs one way for the replacement argument. Every honest version of “AI replaces the manager” quietly assumes the whole job lives on the bottom rung. It doesn’t. The higher you climb, the more the work is the human part, and the higher you climb, the less AI reaches.

The coaches leg

Now the coaches. This is the one people ask me about most, probably because I’m a coach saying it. Can AI replace the coach? Here the evidence is further along than you’d think.

The strongest thing we have is a randomized trial from de Haan, Terblanche, and Nowack this year. They gave leaders one of three things: a human coach, an AI coach, or nothing. The human coaching worked: real gains in goals and stress. The AI coaching did no better than nothing at all, and engagement with it was markedly weaker (dropout roughly ten times higher, and several participants asked to switch to a human coach). And the thing that actually predicted whether coaching worked was the relationship itself, which was far stronger with a person.

Now, I don’t want to overcorrect. AI coaching isn’t useless. For the structured stuff (tracking goals, checking in on what you said you’d do), it works fine. There’s even a solid study where an AI coach built around goal-setting kept pace with human coaches on hitting goals. That’s real. If “coaching” means “did you do your three things this week,” a bot can handle that.

But that’s not the real work. At least not the holistic leadership coaching that we do at Bricolas. The real work is the human part: helping a leader see what they keep missing, staying in the hard stuff with them, hearing what they’re not saying. That’s not something AI hasn’t caught up to yet. It’s a different thing entirely. The coaches I know already figured this out, and their phrase is the best I’ve heard: AI around the sessions, human in the sessions.

The leadership-development leg

Last leg: leadership development itself. Not the coach, not the manager. The whole work of growing leaders.

Here’s what people miss. You don’t learn to lead in a classroom (or the traditional lecture style classroom at least). You learn it by doing, under real pressure, with someone giving you honest feedback while you’re still in it. Reading about leadership builds about as much leadership as reading about swimming builds a swimmer.

That’s exactly what AI puts at risk, and not in the obvious way. The risk isn’t that AI teaches leadership badly. It’s that AI does the reps for you, so you never build the muscle. HBR has been tracking this: junior people now turn out polished work fast, but they’re not building the judgment underneath it. The path from good contributor to real leader is quietly getting hollowed out. You can hand someone a finished answer on the best way to give someone feedback. But that answer doesn’t necessarily translate to the skills involved with the experience of actually delivering the feedback.

I keep coming back to the trades to make sense of this. For a couple of decades we treated skilled trades as lesser, chased the prestige of a corner office and a fancy title, and thinned the pipeline. Now there’s a shortage and the trades are paying a premium. People-leadership is the skilled trade of the organization. We’re devaluing it again, this time with AI hype, and if we keep going we’ll pay to rebuild it later at a much higher price.

Where the real risk actually is

One more thing, because I don’t want this to read as “relax, nothing’s at stake.” Plenty is. The real risks just aren’t the one the headlines sell. Before I point at the new stuff, notice we’ve already walked past two of the biggest ones. The coaching trial where AI produced no benefit for leaders. That’s a risk: replace the relationship and you lose the part that works. The judgment junior leaders never build because AI did the reps. Same. Those aren’t footnotes to this piece. They’re the argument. The rest are quieter, and they’re mostly about what happens when we let AI run the parts of leadership that were never about output.

Start with psychological safety. When AI runs the monitoring, people get more anxious, not less. A recent study traced the whole chain: AI adoption eroded psychological safety, and eroded safety fed depression and disengagement. The one thing that buffered it was ethical leadership, a human actually paying attention. Hand the watching to the machine and the trust goes with it.

The advice all starts to sound the same. Feed the same models the same prompts and you get the same answers. One study ran fifteen thousand strategy simulations across seven AI systems, and they landed on the same buzzword playbook no matter the situation. AI advice regresses to the mean. Your judgment is the part that doesn’t.

Confident and wrong is the dangerous combination. These tools rarely say “I don’t know.” They’re built to sound sure even when they’re guessing. OpenAI’s own researchers found the training rewards confident guessing over admitting uncertainty. In an advisory role, a confident wrong answer does more damage than an honest “I’m not certain.”

The cuts may land in the wrong place. The loudest fear isn’t AI-as-your-boss. It’s AI used to justify cutting headcount. Worth talking about. But if we thin the people-leadership layer to save money, we’re cutting the exact thing the first half of this piece said we can’t afford to lose.

And someone decides who still gets a human. If AI coaching becomes the default for junior staff while senior leaders keep their human coaches, we haven’t democratized development. Bachkirova and Kemp argue exactly this risk plays out: a two-tier system where the people with the least power get the least human.

Actionable: what you actually do

If you’re a leader being told you must use AI, here’s how to say yes without gutting the thing that matters. One test runs through all of it: are you using AI to buy back time for the human work, or to do the human work? Take the toil off your plate so you have more attention for developing people and making the calls. Start where the stakes are low and the task repeats, and always own the last mile: the judgment and the voice.

Use it as a thinking partner.

  • First draft, never final word. If it’s going to handle your email, let it draft the email, the brief, the plan, then you make the judgment pass.
  • Read what you don’t have time to read. Summarize the long report, the thread, the market roundup, so you get to the real questions faster. Check anything important before you act on it.
  • Make it your devil’s advocate before a big call. Ask it to argue against your decision and name what you’re missing. That sharpens your judgment instead of replacing it.
  • Prep the hard conversation, then have it yourself. Draft the talking points, anticipate the reactions, then walk in and do it human.
  • Get up to speed fast in an unfamiliar area, so you can ask your experts better questions instead of replacing them.
  • Translate across audiences. Turn the technical version into plain language for the board, or the strategy into concrete terms for the front line.

Use it to automate the toil.

  • Automate the reports you produce on a cycle. If you generate the same summary or rollup regularly, automate the assembly and spend your time reviewing and adding judgment.
  • Kill the admin drag. Meeting notes into action items, status roll-ups, scheduling, inbox triage. Every hour of admin you automate is an hour back for the people work.
  • Templatize the repeatable. Onboarding docs, standard proposals, common responses, where AI fills the template and you customize the last ten percent.
  • Turn your team’s hard-won judgment into a shared, searchable resource, to speed people up rather than justify cutting them.
  • Let AI hold the data and the dashboards so you can spend your attention developing people, never as a surveillance tool. The test: does it free you to coach, or replace your coaching?

None of this replaces you. All of it gives you more room to do the part only you can do.

The close

So that’s the whole argument, across all three. AI can’t replace the leader, it can’t replace the coach, and it can’t replace the work of growing either one. Not because AI is weak (I said at the top, I build with these tools every day) but because we keep mis-naming what the work actually is. The parts we call “management” and “coaching” and “development” were never mostly checking and measuring and passing along information. They were the human parts. Augment those. Don’t pretend to replace them.

I try to hold myself to this in my own practice, and I’ve gotten specific about how: what I use AI for, where I won’t, and what I tell the people I work with. That’s its own piece, because “use it responsibly” is easy to say and harder to pin down. Here’s how I hold it →.

And if you want proof instead of a promise, look at how the paper behind this whole view got built. To find where AI actually defaults on questions of leadership, I pointed it at the problem three different ways. I asked four models to define leadership cold and watched where each one landed. I made one work through twenty-five bodies of the actual scholarship, from the mainstream taxonomy out to the critical and decolonial edges. And I rebuilt a real seven-year academic project, the General Theory of Leadership, as a council of twenty-three AI personas of the scholars who lived it, and set them deliberating.

The AI looped through all three experiments. I’m the one who designed them, and the one who made the claim. And here’s the part I actually care about: it was human-in-the-loop the whole way, but it let one person reach a breadth I could never get to alone — twenty-five literatures, twenty-three reconstructed scholars, four models, the kind of sweep that used to take a full research team and a couple of years. That’s the augment I keep coming back to. The machine didn’t do the thinking; it gave me the reach to do more of it, across more ground than I’d cover on my own.

It’s all documented, the method and the roster and the transcripts, on the paper itself →. Don’t take my word for the loop. Go look at what it built.

So here’s the question I’d leave you with. Not “will AI replace us.” I think that’s the wrong question. The real one is quieter: now that the machine can do the checking, are we finally going to do the part that was always the actual job?

send a note.