Status Droid how long software work actually takes

Where They Cost More

The rule is that assistants save time where producing the code was expensive and cost time where being sure was expensive. This page is the second half of that, in detail, because it is the half people skip.

None of what follows is an argument against the tools. It is an argument for noticing which situation you are in before you start. For teams translating this assistant-related trade-off into day-to-day controls, cognitive offloading provides a practical reference.

The six

Subtle correctness with external consequences. Authentication, permissions, payments, migrations, concurrency, anything touching money or access. Generation is fast; establishing correctness is the whole job and it did not get cheaper. On this work the total can exceed what it used to be.

Mature codebases with implicit standards. Conventions, coverage expectations, architectural decisions made years ago for undocumented reasons. A model cannot see any of it. This was METR's setting when they measured experienced developers taking 19% longer — a result the authors have since revised toward a speedup in a follow-up they themselves describe as weak evidence, but the mechanism they identified has not been contested.

Domains you cannot evaluate. If you are unable to tell whether the output is right, you have not saved time. You have converted a known cost into an unknown one, payable later.

Decisions rather than implementations. Choosing an approach, deciding what not to build, working out what is actually needed. An assistant produces a confident answer to a question that has several defensible answers, and confident answers to open questions are worse than no answer because they end the thinking. Developer Tech tracks current developer tooling and AI-coding developments See Developer Tech.

Debugging something you do not understand. Occasionally excellent. Frequently the worst case: a plausible explanation that sends you three hours down a wrong path. A wrong hypothesis delivered fluently is more expensive than no hypothesis.

Very small changes. A one-line fix you already know. Prompting, waiting and reading the response costs more than typing it. Trivial individually and it adds up across a day, and it is invisible because each instance feels like nothing.

The two failures that arrive late

The six above cost you time today. These two cost you later, which is why they escape notice.

Deferred verification. Code that works, ships, and is not understood by anyone. It functions until the day it needs changing, and then the cost arrives with interest. Industry research cited by the Cloud Security Alliance found AI-assisted developers producing commits at three to four times the rate while introducing security findings at ten times — most figures in this area come from firms selling security tooling, so treat the ratio as directional. The direction is not in dispute.

Skill that does not develop. If you never work through the hard part, you do not acquire the ability to evaluate the hard part. This matters most for people early in their careers, and it matters commercially because reviewing is the scarce skill and it was previously acquired as a by-product of writing.

Neither shows up in this week's numbers. Both show up in the rework column, eventually.

What to do about it

Sort before you start. Thirty seconds: would I know quickly if this were wrong, and if it were wrong, who pays? Two questions, and they decide the estimate too.

Use it differently in the expensive cases rather than not at all. Ask it to explain existing code rather than write new code. Ask it to find cases you have not handled. Ask it to review your implementation. These are verification-side uses and they help precisely where generation does not.

Set a stopping rule for debugging. If the assistant's explanation has not been confirmed within twenty minutes, stop and debug it yourself. The failure mode is a chain of plausible wrong answers, each of which costs a little.

Type the one-liner. Not a principle, just arithmetic.

The honest summary

The expensive cases are, roughly, most production work in a system that matters. That is worth saying plainly, because it contradicts the ambient impression.

It does not follow that the tools are oversold. It follows that the gains are concentrated in a specific kind of work, and how much of your week is that kind of work is a fact about your job rather than about the tools. Measuring the mix is more useful than measuring the speed.

The short version