The Estimation Problem: Why AI Makes You Terrible at Guessing How Long Things Take

You used to be decent at estimation. Not perfect — nobody is — but your predictions were in the ballpark. You'd factor in edge cases, spot the hard parts early, and your stakeholders trusted your numbers. Then you started using AI coding tools. Now you ship twice as fast, and your estimates are three times as wrong.

What Estimation Actually Is

Before we can understand what's breaking, we need to understand what estimation actually requires. When you estimate a software task, you're doing several things simultaneously:

• Decomposing the problem — breaking it into sub-components you can reason about individually
• Accessing experiential memory — retrieving similar past experiences and adjusting for differences
• Spotting traps — noticing where similar tasks went wrong before, where edge cases hide, where unfamiliar territory starts
• Adjusting for context — accounting for code you don't know, dependencies you don't control, test coverage you need, review cycles you'll go through
• Communicating uncertainty — translating your probabilistic mental model into a number a stakeholder can use

The critical word is experiential. Your estimation ability is not a cognitive skill you've learned — it's a calibration you've built through repeated contact with reality. Every time you estimated, then watched what actually happened, then updated your model, your estimates got a little better. This is the same mechanism that makes a physician recognize a serious condition in seconds, or a mechanic hear what's wrong with an engine.

The calibration is the product of the feedback loop: estimate → experience → compare → update. This loop is what AI tools are quietly breaking.

The Three Ways AI Breaks Your Calibration

When you use an AI coding tool, several things happen that collectively destroy your estimation ability:

1. You stop feeling the friction

The hard parts of coding — the logic traps, the edge cases, the unfamiliar APIs, the subtle bugs — are also the parts that teach you the most. When AI handles them for you, you don't just avoid the work. You avoid the learning. Your brain's model of what constitutes "hard" starts to drift because the hard parts are no longer reaching you.

Consider a task you've done a dozen times: implementing a new endpoint with authentication. With AI, you might say "that's 2 hours." But what you know from experience is that the friction is in the edge cases — the rate limiting you need to handle, the auth token refresh race condition you hit last quarter, the logging requirement that nobody mentioned until the security review. Those edge cases are where the time actually goes. And now you're not touching them.

Your sense of how long things take is built on feeling the friction. Remove the friction, and your estimates drift toward whatever the AI-generated code "looks like" — smooth, clean, simple. Which it never is.

2. You lose the feedback signal

Here's the estimation feedback loop for a task you've done yourself: You estimate 3 days. It takes 4 days. You update your model. The next similar task, you estimate 4.5 days. Over time, your estimates converge toward reality.

With AI, something different happens. You estimate 2 days for a feature. The AI generates most of it in 20 minutes. You spend the next 1.5 days debugging edge cases and integration issues. It took 2 days — your estimate was "right." But you learned almost nothing from it, because the ratio of time spent was 95% debugging and 5% AI generation. Your brain filed this under "2 days, fine." It did not file it under "authentication endpoint edge cases are still hard and I need to plan for them."

The signal is scrambled. You hit the difficulty but you didn't earn the understanding of where the difficulty came from.

3. The reference class changes

All estimation is comparative. You estimate based on similar past work. But when AI handles the work, the reference class changes. You're no longer estimating based on "what it felt like to build this feature by hand." You're estimating based on "what it looked like when AI built it and I debugged it."

These are different tasks with different difficulty profiles. And the problem is that your brain is bad at distinguishing them. It just sees "took 2 days, shipped" and files it in the same category as previous 2-day tasks. But the actual cognitive effort profile is completely different — and that's what your next estimate depends on.

The Signs Your Calibration Is Drifting

You might be experiencing estimation calibration drift if:

Why Senior Engineers Are Most At Risk

You might expect this problem to hit junior engineers hardest — they haven't built calibration yet, and AI is preventing them from building it. And that's true.

But senior engineers are arguably in more danger. Here's why: your calibration was hard-won over a decade of direct contact with complexity. You built your judgment by spending hours on bugs that looked simple, by getting burned by edge cases you didn't anticipate, by learning the hard way where the traps are. You earned that calibration through pain.

And now you're watching AI sidestep that pain — for you and for everyone coming up behind you

But senior engineers are arguably in more danger. Here's why: your calibration was hard-won over a decade of direct contact with complexity. You built your judgment by spending hours on bugs that looked simple, by getting burned by edge cases you didn't anticipate, by learning the hard way where the traps are. You earned that calibration through pain.

And now you're watching AI sidestep that pain — for you and for everyone coming up behind you. The expertise you've spent years cultivating isn't just being augmented. It's being bypassed. And when your calibration goes, so does something harder to name: the confidence that comes from knowing you can look at an unfamiliar problem and have a defensible guess about what's involved.

Seniors are also more likely to be the ones giving estimates to stakeholders, managing sprint planning, and making architectural decisions. Broken calibration at the senior level has organizational gravity.

The Organizational Cost

This isn't just a personal problem. Broken estimation at scale creates organizational dysfunction:

The Research Behind This

This isn't just intuition. Several research threads converge on the same conclusion:

Klein's Recognition-Primed Decision model shows that experts make fast, accurate decisions by recognizing patterns from prior experience. When AI handles the edge cases that would have created those patterns, the recognition capability atrophies. You can't recognize what you've never seen.

原子弾 research on skill decay demonstrates that motor and cognitive skills decay without practice — and that decay accelerates when the skill is only partially maintained. Using AI for 80% of implementation while maintaining "oversight" still decays the 20% — because the oversight role is fundamentally different from the implementation role.

Gloria Mark's research on attention at work (University of California, Irvine) found that interruptions take an average of 23 minutes to recover from. But more relevant here: her work on how workers develop situational awareness — the intuitive sense of where problems are likely to emerge — requires sustained, uninterrupted contact with the work. AI, by smoothing over frictional moments, removes the data points that build situational awareness.

The Perlis paradox: a programmer who has learned from debugging AI-generated code has not learned what they think they've learned. The debugging felt productive. It was. But it built a different skill than writing the code would have — a diagnostic skill rather than a constructive one. Both matter for estimation, in different ways.

How to Know If Your Calibration Is Gone

Try this: estimate a feature you're about to work on. Then, without using AI, build just the hard part — the component that involves the most unfamiliar logic or the most integration complexity. Time yourself honestly.

If your actual time was more than 2x your estimate, your calibration is drifting. If it was 3x or more, you have a significant problem.

The reason this works: the gap between your estimate and your direct experience on the hardest part will reveal exactly where the drift is happening. The AI is probably handling the parts you would have handled quickly anyway. The gap is in the friction.

Recovering Your Calibration

You can rebuild your estimation calibration — but it requires deliberate practice, not just passive AI use. Here's what actually works:

The Question You Should Be Asking

Not "how do I ship faster?" You already know the answer to that.

The question is: "Do I still know what I'm doing, and can I explain it?"

If the answer is yes, your estimation is probably okay. You're maintaining enough direct contact with the work to keep your model calibrated. The AI is augmenting you, not replacing your judgment.

If the answer is "I can ship it but I'm not sure I could explain why that approach worked" — that's a calibration warning. And it's not just an estimation problem. It's the broader problem of ownership, authorship, and the ability to reason about your own work.

Estimation is just the canary in the coal mine. The drift you're seeing in your estimates is the same drift that's happening in your architectural judgment, your debugging intuition, and your ability to anticipate where complexity hides. They're all products of the same feedback loop — and that loop needs contact with friction to function.

"The single biggest problem in communication is the illusion that it has taken place." — George Bernard Shaw

Estimation is communication. And the illusion that AI has preserved our ability to estimate — when it's actually eroded it — is one of the most consequential blind spots in software engineering right now.