We build with AI. So does every competent studio you are talking to. It has not made our invoices smaller, and any agency promising you a discount because of it is either mispricing the work or has not measured what actually changed.
Here is the argument in one line: AI compressed the cheapest part of building software and expanded the most expensive part. What you pay for is review, judgement and someone carrying the risk when it goes wrong, and there is more of that work now, not less.
Does AI reduce software development cost?
Not in the way the pitch implies. AI reduces the time spent producing a first draft of code, which was never the majority of a project's cost. It adds time to verification, and the evidence on that is now reasonably consistent.
The METR result is the one people quote, and it deserves its caveats: 16 developers working on mature open-source repositories they knew well, using early-2025 tools, and METR is explicit that it does not generalise to all developers or all settings. Treat it as one careful measurement rather than a law.
The pattern across all four holds up better than any single number. Generation got faster. Trust went down. Verification went up.
What are you actually paying an agency for?
Judgement about what to build, and accountability when it breaks. Neither has been automated, and both got more expensive.
Consider what the Veracode figure means in practice. If close to half of generated code introduces a security flaw when nobody specifies otherwise, then every line has to be reviewed by someone who knows what an injection vulnerability looks like in your particular stack. That is senior time, and senior time is the expensive part of any invoice. The cheap part, typing out a standard controller, is the part the model does.
Stack Overflow's 2025 survey puts a number on the specific pain: "AI solutions that are almost right, but not quite" is the largest single frustration, affecting 66 per cent of developers, with 45.2 per cent reporting that debugging AI-generated code takes more time. Almost right is the worst possible failure mode, because it passes a casual read and fails in production.
So where does the AI gain actually go?
Into scope, quality and speed of iteration, rather than into a lower price. On our own work the accounting is roughly this: we prototype faster, we throw more options away before committing, we write more tests than we used to, and we catch things in review that would previously have reached staging.
The client-visible result is a better product on a similar budget, or the same product sooner. It is not a 30 per cent discount, and an agency offering one is telling you they have removed the review step. That is the step you are paying for.
Note the difference between this and the question of whether AI belongs in your product. Building with AI and shipping AI features are separate decisions with separate business cases, and we argued the second one in should you add AI to your app.
Should you pay less for AI-generated code?
You should pay for outcomes and warranties, not for hours or keystrokes. The questions worth asking a supplier are about what happens after the code exists:
- Who reviews AI-generated code, and what is their seniority?
- What automated security testing runs before anything reaches your environment?
- Who is liable if a generated dependency carries a licence you cannot use commercially?
- What happens when a model-written component fails in production six months after handover?
An agency with real answers has costs to cover. An agency with no answers is cheaper because it is passing you the risk, which is the same trade we described in the real price of cheap offshore development, arriving by a different route.
Will AI make software cheaper eventually?
Probably, for the commodity end, and slowly. Standard builds that follow a well-trodden pattern will keep getting cheaper, in the same way that website costs collapsed once templates matured. That has already happened at the bottom of the market.
The work that does not get cheaper is the work where the requirements are unclear, the data is messy or the domain carries obligations. DORA's finding that AI amplifies instability points the same way: teams with weak testing and slow feedback got worse outcomes from AI, not better. The gap between good and poor delivery widened.
If your project is genuinely standard, expect prices to fall and pick a supplier accordingly. If it is not, the AI discount you are being offered is coming out of the review budget. For how the underlying numbers break down, see what it actually costs to build an app.