AI software development means two different things depending on who is saying it. Sometimes it means building software that has AI in it. More often, in 2026, it means building software with AI: developers using models to write, review and refactor code. This post is about the second one, because it is the one changing what you are quoted, and because the claims being made about it in sales conversations have run some distance ahead of the evidence.
The short version, and it is worth stating before the numbers: writing code got faster. Deciding what to build, connecting it to the systems you already run, checking that it is right, and owning it for the next five years did not. Those four have always been most of the cost of custom software. They still are.
What the evidence actually says
Adoption is essentially settled. Google's 2025 DORA research, surveying nearly 5,000 technology professionals, found around 90% now use AI in their work, up fourteen points in a year, with a median of about two hours a day spent working with it. Over 80% believed it had increased their productivity.
Belief is doing a lot of work in that sentence, and this is where it gets interesting.
The most cited counter-evidence is a randomised controlled trial published by METR in July 2025. Sixteen experienced open-source developers completed 246 real tasks in codebases they knew well, some with AI tools available and some without. They predicted AI would make them 24% faster. Afterwards, they believed it had made them about 20% faster. Measured, they were 19% slower. The gap between what they felt and what the clock recorded was the finding, not the slowdown itself.
Being fair to the evidence matters more than having a good line, so: METR itself now labels that result historical. A February 2026 follow-up showed some evidence of speed-up, but the organisation flagged that selection effects had made the central estimate unreliable, because developers who benefit most from AI were reluctant to participate in a study that might require them to work without it. The honest reading is not "AI makes developers slower". It is that self-reported productivity gains are not evidence, and almost every number you have been quoted is a self-reported productivity gain.
The most useful figure we have seen comes from GitClear's January 2026 analysis, which compared heavy AI users to their own past selves rather than to other developers. Heavy AI users out-produce non-users by four to ten times, but most of that gap pre-dated AI: these were already the high performers. Against their own prior output, the gain was a more modest 25%.
Twenty-five percent is a real number. It is also not the number in the pitch deck.
The other DORA finding worth carrying into a commercial conversation: AI acts as an amplifier. It magnifies the strengths of organisations that already work well and the dysfunctions of ones that do not. A team with clean architecture, a stable pipeline and real code review gets faster. A team without those gets more code, faster, in a worse state.
The bill moved, it did not disappear
This is the part a buyer needs and rarely hears.
GitClear's 2026 research, "The Maintainability Gap", tracked seven code-quality signals across 623 million changed lines from 2023 to 2026. The direction is consistent and it is not good. Duplicated code blocks are up 81% and at the highest level on record. Properly refactored "moved" code, the signature of code being consolidated and reused rather than copied, fell from 21% of changed lines in 2022 to 3.8% in the first half of 2026. Copy and paste rose from 9.4% to 15.7% over the same window. Cross-file function calls, a rough proxy for genuine reuse, are down 35%. Constructs that mask errors rather than handling them, code that swallows a failure instead of dealing with it, are up 47%. And work on legacy code, the unglamorous business of returning to something written a year ago and tidying it, has fallen 74% since 2023.
Translated out of engineering vocabulary: the code is being written faster and being made harder to change. Duplication is a tax you pay later, every time a rule changes and you have to find all eleven places it was copied to. Error-masking is worse, because it produces software that appears to work and behaves strangely, which is the most expensive category of defect there is.
DORA found something adjacent from the practitioner side: about 30% of respondents said they had little or no trust in AI-generated code, and Stack Overflow's 2025 developer survey put distrust of AI accuracy at 46%. The people using these tools most are not the ones making the boldest claims about them.
None of this is an argument against AI-assisted development. We use these tools daily and would not go back. It is an argument about where the cost went. It moved from writing to reviewing, and from this year's budget to the next three.
What to ask when you commission software in 2026
If you are buying custom software from anyone, including us, these are the questions that separate a fair quote from a cheap one.
- Where did the AI saving actually land? A firm that has genuinely got faster should be able to say where: prototypes, boilerplate, tests, migrations. A firm claiming a uniform discount across all work has not measured anything.
- Who reviews the code, and how? With AI in the loop, review is more important than it was, not less, and it is the first thing cut when a quote is squeezed. Ask what their review process is and whether it is a person or another model.
- What is the maintenance arrangement? Cheap to build and expensive to own is the specific failure mode the maintainability data predicts. Ask what year two costs, not just year one.
- What do you get at the end? Source, documentation, deployment access, and enough explanation that another firm could pick it up. If the answer is vague, the low build price is a deposit on a long dependency.
- How much of this is glue? Most business software value sits in connecting things that already exist. Ask what proportion of the estimate is integration, because that part has not got meaningfully cheaper and any quote that assumes it has will be revised upward later.
The last one matters most for a mid-sized business, because the AI integration work of connecting a system into your CRM, your accounting package and your existing processes is where the effort concentrates, and it is stubbornly resistant to being sped up by a model that cannot see your Sage instance.
Where the maths genuinely did change
Being even-handed: there are places where AI-assisted development has changed what is worth doing, and they are real.
Prototypes and throwaways. Building something to find out whether an idea is worth pursuing is now cheap enough that not doing it is the expensive choice. Where a proof of concept used to be a two-week decision, it is often a two-day one.
Small internal tools. The category of job that was always worth doing and never quite worth a developer's fortnight, the reconciliation script, the one-off data clean-up, the little internal dashboard, has moved firmly inside the economics. This is the quiet win most businesses have not collected yet.
Migrations and glue. Mechanical, high-volume, well-specified transformation work is exactly where the tools are strongest.
What these have in common is that the specification is clear and the cost of being wrong is low. That is the same shape as the AI workflows worth automating elsewhere in your business, which is not a coincidence.
The thing that did not get cheaper
None of this touches the actual bottleneck. The hardest and most valuable part of building software has never been the typing. It is working out what to build: which decision a person is trying to make, what would have to be true for the software to change that decision, and what happens to the surrounding process when it does. A model can now write a feature in an afternoon. It cannot tell you whether that feature should exist.
This is the same argument we make about AI projects generally, that you should start from the decision, not the data, and it applies with more force now that building is cheap. When the cost of building falls, the cost of building the wrong thing does not fall with it. It rises, because there is less friction stopping you.
So the honest answer to "has AI made custom software cheaper" is: somewhat, in specific places, and less than you have been told. What it has definitely done is raise the return on knowing exactly what you want before you start, and on working with people who will tell you when the answer is that you do not need it built at all. That is what our AI consulting work exists to establish, usually before anyone opens an editor.
FAQ
What is AI software development? It usually means one of two things: building software that includes AI features, or building software with the help of AI coding tools. In 2026 the second is the more common usage, and it describes the near-universal practice of developers using models to write, review and refactor code.
Has AI made custom software development cheaper? Partially. Writing code is faster, particularly for prototypes, small internal tools, boilerplate and migrations. Deciding what to build, integrating with existing systems, reviewing output and maintaining the result have not become meaningfully cheaper, and those have always been the larger share of the cost.
How much faster are developers with AI? Less than the marketing suggests. Around 90% of technology professionals use AI and over 80% believe it makes them more productive, but a METR randomised trial found experienced developers were measurably 19% slower while believing they were 20% faster. GitClear's comparison of heavy AI users against their own past output found a gain of around 25%.
Does AI-generated code create technical debt? The published evidence points that way. GitClear's 2026 analysis of 623 million changed lines found duplication up 81%, properly refactored code down from 21% to 3.8% of changed lines since 2022, and legacy maintenance work down 74% since 2023. The practical consequence is software that is cheaper to build and more expensive to change.
What should I ask a development firm about their use of AI? Where specifically the saving lands, how code is reviewed now that AI is writing some of it, what maintenance costs in year two, what artefacts you receive at the end, and what proportion of the estimate is integration work. Vague answers to those five are the warning sign.
References
- Google Cloud / DORA: 2025 DORA Report: State of AI-assisted Software Development (90% adoption, over 80% perceived productivity gain, 30% low trust, AI as amplifier). https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report
- Google: How are developers using AI? Inside Google's 2025 DORA report (the trust breakdown and median daily use). https://blog.google/innovation-and-ai/technology/developers-tools/dora-report-2025/
- METR: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (the 19% slowdown against a 20% perceived speed-up). https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
- METR: We are Changing our Developer Productivity Experiment Design (February 2026; the selection-effect caveat on the follow-up data). https://metr.org/blog/2026-02-24-uplift-update/
- GitClear: The Maintainability Gap: 2026 AI Code Quality Research (623 million changed lines, 2023 to 2026; duplication, refactoring, error-masking and legacy maintenance trends), and AI Coding Tools Attract Top Performers (January 2026; the ~25% gain against past self). https://www.gitclear.com/the_ai_code_quality_maintainability_gap
- LeadDev: Code maintainability plummets in the AI coding era (independent summary of the GitClear and GitKraken dataset). https://leaddev.com/ai/code-maintainability-plummets-in-the-ai-coding-era
- InfoQ: DORA Report Finds AI Is an Amplifier in Software Development, But Trust Remains Low (including the Stack Overflow 2025 developer survey trust figures). https://www.infoq.com/news/2025/09/dora-state-of-ai-in-dev-2025
Written by JP, Sixees Labs. Last reviewed August 2026.