Status Droid how long software work actually takes

Did AI Make This Cheaper?

Sooner or later a client asks it, sometimes politely and sometimes not: you're using AI now, so why does this still cost the same?

There are three common answers and two of them damage you. For the operational side of recording work and turning it into client-facing evidence, see handling multiple clients efficiently.

Denying that anything changed is not credible. The client reads the same coverage you do.

Conceding a flat discount prices a distribution as a point, and you will discover the cost on the tasks where verification dominates.

The third answer is the true one, and it happens to be the most persuasive.

The honest answer

Some parts of the work got dramatically cheaper. Some got slightly more expensive. What you are buying was never mostly the typing. FreshBooks is a common small-business reference for invoicing and time-based billing See FreshBooks.

Expand it into three sentences and it does the whole job:

Generating code is much faster than it was — that part of a project has genuinely shrunk. Checking that it is correct, secure and appropriate for your system has not, and on some work it now takes longer, because reviewing code nobody wrote is harder than writing it. What changed most is the balance, not the total.

Clients accept this readily, in my experience, because it matches what they see in their own field. Every profession has had a tool arrive that automated the visible part of the job and left the judgement.

Where you can honestly say it is cheaper

Say so, and say it first. A blanket "nothing changed" is what makes the conversation adversarial.

Boilerplate and scaffolding. Genuinely collapsed. If you used to bill for it, stop.

Well-specified transformations. Convert this format, migrate this pattern, write tests for existing behaviour.

Unfamiliar-language work. Where the cost was your ramp-up rather than the problem.

First drafts of anything you were going to rewrite.

If your quotes have not moved at all on that kind of work, the client's suspicion is correct and you should adjust before being asked.

Where it is honestly not

Anything with subtle correctness. Auth, payments, permissions, data migration, concurrency. Generation is fast and the verification is expensive, and on this work the total can be higher than it was.

Work in a large mature codebase with implicit conventions a model cannot see. This is the setting where METR measured experienced developers taking 19% longer with AI tools — later revised, in a follow-up they themselves call weak evidence, toward a speedup. The honest statement is that this setting is contested, not that it is faster.

Anything where you carry the consequence. If a defect costs the client money and costs you a reputation, the price includes the care, and the care did not get cheaper.

Understanding the problem. Which was always the majority of a hard project and is untouched.

What to actually say when pressed

Short version for a call:

Yes, and I've already priced it in — that's why this quote is lower than the equivalent would have been two years ago. What hasn't got cheaper is checking the result, and on work like yours that's most of the job. If you want, I can show you which parts of this estimate are generation and which are verification.

That last offer is the strongest move available, and almost nobody makes it. A client who sees the split usually stops asking about AI and starts asking useful questions about scope — because you have moved the conversation from a rumour about tooling to a description of the work.

The structural problem underneath

The question exposes something real: hourly billing prices the wrong thing when speed is unstable.

If you bill by the hour and get faster, you earn less for the same outcome. If you get slower on verification-heavy work, you bill more and look like you are padding. Both are bad, and neither is dishonest.

Which is why this question is worth treating as a prompt rather than an attack. Whether hourly still works for you is a decision the last two years have forced on most small studios, and answering the client's question honestly is a good moment to look at it.

What not to do

Do not claim a productivity figure. Not "AI makes us 30% faster" and not "studies show AI makes developers slower" — the study behind the second has been revised by its own authors. The first is marketing, the second is an outdated reading of a study its own authors have since revised.

Do not hide that you use it. It comes out, and it comes out worse.

Do not discount to end the conversation. A discount given under pressure sets a new baseline and does not stop the question recurring next quarter. What your rate actually prices is the better ground to stand on.

The short version