When Polished Output Stops Being Proof of Good Thinking

Last year, I spent a lot of time designing a new Leadership Alignment System at GitHub. The goal was relatively straightforward: create a better way to keep leaders aligned and prepared across the moments that mattered most without building an enormous communications machine around every single thing happening in the company.

AI was involved from the beginning. I used it to brainstorm structures, map workflows, think through different artifacts, pressure-test how information should move, and imagine all the ways the system might need to work across different leaders and scenarios. And AI was very good at helping me design a system.

The problem was that it was also very good at helping me design a much more complicated system than we actually needed.

The early versions had more process, more artifacts, more ways to account for every possible use case. None of it was obviously bad. Most of it was logical, thoughtful, and defensible. If I had dropped the whole thing into a document, polished it up, and presented it as a finished operating model, it probably would have looked like a lot of work had gone into it.

But looking like a lot of work had gone into it wasn't the goal. The goal was to make leadership alignment easier. So I kept taking things out.

What actually deserved a higher level of support? Where did leaders genuinely need more visibility? What needed an owner? What could become a default instead of another decision? Which artifacts helped people act, and which ones mostly proved that we had thought through every possible scenario?

Eventually, the system became much simpler: clearer tiers, clearer ownership, predictable readiness moments, and a smaller set of repeatable artifacts.

AI helped me build it. But the judgment was in knowing what to remove.

I've been thinking about that a lot lately because AI is changing some of the signals we have historically used to decide what good work looks like.

For a long time, clean, polished, complete output was at least a reasonable proxy for quality. A twenty-page strategy document represented real effort. Someone had to gather the information, organize the thinking, write the thing, make the connections, edit it, and probably turn it into a reasonably attractive deck before anyone else saw it.

Today, you can take a manager's raw notes, drop them into an AI tool, and have something that looks remarkably close to a finished strategy before your coffee gets cold. That is an incredible capability. I use it constantly.

It also means that the deliverable itself tells us a lot less than it used to.

The twenty pages are cheap now

I once heard someone talking about executive communications say that, by the time a senior leader opens your proposal, they will often understand 90 percent of what you're trying to tell them before you've even started talking.

That idea has stuck with me because it runs counter to one of the easiest traps to fall into: assuming that more information demonstrates more thinking.

It usually doesn't.

If you're trying to anticipate and answer every possible question in the document, there is a good chance you've made the real conversation harder. Someone's eye catches on a detail that doesn't actually matter. They start debating a mechanism when what you need is a decision about direction. Or they simply glaze over before you ever reach the tradeoff that matters.

Completeness is a bit of a myth anyway. There will always be another edge case, another question, another data point you could add.

The harder skill is knowing what the audience actually needs in order to understand the problem and make the next decision. That's why I think a twenty-page proposal is cheap now, while the right one-page proposal is becoming more valuable. The one-page version requires you to have an opinion.

You have to understand the audience, decide which context matters, make a tradeoff, know what you're not recommending, and have enough conviction to leave perfectly reasonable information out.

Those are judgment calls, and AI can make them look deceptively easy because it is very good at producing something that feels complete. And that's where I think people can get into trouble.

You can take raw notes from a leader, ask AI for a strategy, an internal event series, an operating model, a customer marketing plan, or almost anything else, and get a thoughtful-looking answer. It may have goals and pillars and workstreams and success measures and a very respectable-looking table.

But a strategy isn't good because it has all the expected parts; it's good because someone made the right calls about what matters.

The human steer

I've started thinking about my own AI-enabled work in four stages: Steer, Generate, Challenge, Execute.

The steer comes first. Before I ask AI to help me build something, I need to have some conviction about the bumpers around the problem:

• What are we actually trying to solve?
• Who really matters here?
• What are we optimizing for?
• What tradeoff are we willing to make?
• What are we explicitly not doing?
• What would good actually look like?

AI can absolutely help me sharpen those answers. In fact, I use it as a thought partner long before I ever ask it to produce a deliverable. I want it to challenge assumptions, give me alternatives, surface context I might be missing, and help me see connections I haven't made yet.

But that is different from asking it to decide what matters.

Once I have a steer, I want AI to generate aggressively. Give me options. Expand the idea. Find gaps. Show me what this could look like. Help me move much faster than I could on my own.

Then comes the part I think is easy to skip: challenge.

I tell AI no all the time. "No, that's too complicated." "That's not really the problem we're solving." "We're giving this too much weight." "Come back to the strategy." "Anchor this to something that's already been decided." "That recommendation makes sense logically, but it isn't going to work here."

And once the direction is right, I want AI involved in as much of the execution as possible. Build the workflow. Automate the repetitive pieces. Analyze the inputs. Connect the systems.

This isn't an argument that humans should think while AI executes. I don't believe that model holds up very well. AI can contribute across the entire process, including ideation and critique.

The distinction is whether you're using AI to improve your judgment or using it in place of judgment. Those are very different things.

The people with judgment are about to get a lot more powerful

There is another side to this that I find much more interesting than the endless debate over whether AI is going to make everyone lazy.

There are things I can build by myself today that I simply would not have attempted a few years ago. Recently, I built a Power BI dashboard that brings together engagement data from across an organization so we can see it in one place. I'm an English lit major and have literally never seen the backend of a Power BI dashboard.

A few years ago, the gap between “I think this view would be useful” and “this thing actually exists” would probably have required technical support, more time, or enough friction that I might have decided it wasn't worth building at all. AI dramatically reduced that gap.

That's where this gets really interesting to me. Someone with good judgment and strong AI capability can now carry an idea much further than they could before. They can research more broadly, pressure-test their thinking, produce multiple options, cross skill boundaries, and then automate large pieces of the execution.

One capable person has an extraordinary amount of leverage, and I suspect we're still thinking too narrowly about what that means because most of us are using AI to improve ways of working that were designed before AI existed.

I'm especially curious about the people entering the workforce now. I don't really subscribe to the idea that an AI-native generation is doomed to stop thinking for itself. Some people will absolutely use these tools to avoid doing hard thinking – there were people looking for shortcuts long before ChatGPT existed – but the curious, ambitious, high-agency people growing up with these tools will develop instincts that those of us who learned how to work before AI probably won't fully understand.

They won't approach a problem by asking, “How can AI help me do this old task faster?” They may not think of it as an old task at all. They'll build ways of working we wouldn't have imagined because the constraints that shaped our own instincts won't exist for them.

That feels much more exciting to me than scary.

We need better signals for good work

There is, however, a management problem hiding inside all of this. If polished output is becoming cheap, leaders need to be careful about continuing to reward polish as though it proves exceptional thinking.

The same is true of volume, completeness, and even speed. If one person produces ten polished plans, that does not necessarily mean they are ten times more valuable than the person who produced one. We have to look underneath the artifact.

• Did they identify the right problem?
• Did they understand the audience?
• Did they make a real tradeoff?
• Did they know what not to do?
• Did they recognize when the sensible recommendation wouldn't actually work in this organization?
• Did they change the direction of the work in a way that moved the business forward?
• And, importantly, could they explain the steers behind the output?

That is where I think AI may end up being surprisingly clarifying. For years, a lot of knowledge work has been bundled together. Thinking, research, synthesis, writing, formatting, coordination, and execution all showed up in one final artifact, so it was difficult to see which part actually created the value.

AI is starting to unbundle them, and when the execution layer becomes dramatically easier, the judgment underneath it becomes much harder to hide. The best AI users won't simply be the best prompters. They'll be the people who have something to say, something to decide, or something worth building – and who know how to use these tools to bring that idea to life.

That is also why I think leaders will need to change what they reward.

Stop rewarding the deliverable. Start rewarding the steer.

AI isn't making good judgment less important; it's making it much easier to see who had it in the first place.