With the advantages of generative AI, some teams ship fast and present cleanly, then go quiet the moment someone pushes on the plan. It’s not because anyone is hiding something, but because the plan came from a prompt, and prompting your way around a hard decision is now easier than making one.

Most delegation advice says to evaluate what someone achieved, not how they achieved it. “State the destination and trust the route.” When delegating to an AI model, that gets a bit more challenging. Let’s dig into why and some questions you can ask to help you and your team better leverage these tools.

The First Pass Isn’t Good Enough

Agentic workflows sever the link delegation has always leaned on: that if the output holds up under scrutiny, the thinking behind it probably held up too. Today, the tools make polish cheap, so a finished-looking artifact stops being reliable evidence that anyone worked through the tradeoffs, and that gap isn’t visible in the artifact itself.

Microsoft’s 2026 Work Trend Index found that 86% of AI users say they treat AI output as a starting point rather than a final answer, which is exactly the right instinct to have. The challenge is that instinct has to survive contact with a deadline, and when a team is behind schedule and a draft looks finished, treating it as a starting point takes discipline most people were never taught to have.

So what can leaders do to resolve this? That same Microsoft research found that when managers actively model their own AI use in front of their teams, employees reported a 22-point lift in critical thinking during usage of generative AI tooling.

• walking through where they’d expect a review gap
• showing the prompting guardrails they rely on
• pointing to the source of truth they checked the output against instead of the model’s first citation

These new types of modeling and evaluation skills aren’t innate, especially when the initial output looks this polished. Someone has to demonstrate what “don’t stop at the first plausible answer because…” looks like under real deadline pressure, and most leaders are learning just-in-time too.

I wrote about a related version of this problem in Our Dwindling Talent Pipeline: junior engineers losing the reps that used to build judgment. This is the same mechanism, one level up. Judgment doesn’t disappear, but the artifacts produced without it become indistinguishable from the artifacts produced with it, at least until someone asks the right question and there’s no good answer waiting. The problem is, asking that question takes time most managers don’t have.

The Squeeze on Judgment

These tools didn’t just make first-pass output look more finished, they increased the volume of it landing on a manager’s desk. Evaluation time didn’t grow to match. DDI’s 2026 Leadership Trends report found flatter organizations are handing each manager a wider, more complex span of control, and it’s costing them:

I just asked you to spend more time modeling and evaluating. The stat above says most of you don’t have it to spend. Both are true, and the way through isn’t finding more hours in a week that doesn’t have them, it’s changing what “looking closer” actually costs.

That’s the difference between a full review and a deliberate reasoning check. Catching a faked understanding doesn’t require re-deriving someone’s analysis from scratch, it requires a couple of pointed questions that a polished guess can’t survive.

The tell shows up as a gap, not a person: specs get more polished while the answer to “what would you cut if scope halved tomorrow” gets thinner. The spec is real work. The judgment underneath it is what didn’t transfer, and a ten-minute Q&A with the team catches it in a way a status update never will. Here’s what those ten minutes actually look like.

The Deliberate Check, In Practice

First, start with the rejected options, not just the chosen one. What did you rule out, and why? That question is close to impossible to prompt your way past convincingly, because answering it well requires having actually weighed tradeoffs rather than generated one defensible-looking option.

Second, validate the reasoning itself by requiring it alongside the answer, not instead of it. Many of the latest thinking models and subagents can document their “thought process,” cite sources, and show the research path they took. You just have to ask for it explicitly, the same way your high school algebra teacher wanted the work shown, not just the final number. The answer itself might be completely right — that’s not what’s in question. What’s at risk without the justification is confidence in it, and confidence is the thing that actually matters: whether you understand it well enough to implement it well, adapt it when reality pushes back, or catch the worst case before it lands.

This isn’t an argument against using AI, and I’d be lying if I claimed I don’t lean on it the same way everyone else does. It’s an argument for updating the checks and balances when delegating to AI.

None of this is unique to AI, either. If I handed a brand-new engineer the design for a piece of enterprise architecture, I wouldn’t assume something was wrong just because I didn’t yet understand their reasoning, but I’d still want to see it before trusting their understanding of the design under real load. Delegating without understanding has always been a way to get things wrong. AI doesn’t introduce that risk, it just removes the more obvious checkpoints that used to come along for free — a new engineer’s design takes days to produce, and those days are full of chances to ask questions along the way. A model’s output takes minutes, so the same conversation only happens if you deliberately build it back in.

The “what, not how” rule assumed the how was expensive enough that faking it wasn’t worth the effort. Is AI raising your team’s bar? covered the same shift from the individual contributor’s side. From the delegation side, the cost of faking comprehension just dropped to nearly zero until something goes awry.

Growing Into It

I don’t have the full replacement for “what, not how” yet, though there are signals beyond “this is pretty.” What I have is two narrower habits: ask about the path not taken and ask for the reasoning behind the path taken… before trusting either one.

This push catches more than I expected, and it’s uncomfortable in exactly the way good diligence should be, for me as much as for the people I’m evaluating. In a human sense, it’s the same process I’d take with another person or a new team. For an AI model, “prove it” becomes part of the prompt output rather than a polished guess you wave through and forget.

What’s the one question you could ask tomorrow that a polished guess couldn’t survive?