AI coding tools have lowered the cost of writing software, but they have not lowered it as far as the hype suggests, and a custom web app quote should now be cheaper and more detailed than it was two years ago.
The £20,000 figure used to be the entry ticket for a small bespoke system, such as a client portal or an internal tool. For well-defined projects, that floor has dropped. For anything messy, it has not.
This article looks at what custom software development costs in the UK now, what the research says about AI-assisted coding, which parts of the bill do not shrink, and how to read the next quote you receive.
What does custom software development cost in the UK now?
Published UK ranges are wide, and they are mostly written by development firms, so treat them as one side of the market.
Magora, a software company, put a first release of a custom web or mobile product at £11,000 to £34,000, with a median of about £25,000, built in six to eight weeks by two or three developers.
A fuller web platform with a back end and admin area ran from £25,000 to £60,000 over three to six months. Softomate Solutions, another UK supplier, said most UK SME projects land between £30,000 and £80,000, with internal tools at £15,000 to £30,000 and customer portals at £30,000 to £80,000. Both published in 2026.
Neither says whether the estimates assume AI coding tools. Quotes still tend to be built on hours of human coding, and the evidence below suggests some of those hours are going.
Do AI coding tools really make developers faster?
On well-defined tasks, yes, and by a lot. On large, mature codebases with experienced developers, the picture is much less clear.
In an early controlled experiment with GitHub Copilot, programmers asked to write an HTTP server in JavaScript finished 55.8% faster than those without the tool, with a wide confidence interval of 21% to 89%. The paper, The Impact of AI on Developer Productivity, recruited 95 professional programmers, though only 35 completed the task and survey.
A randomised trial by the research group METR in 2025 found the opposite for a different group. Sixteen experienced developers worked on 246 tasks in open-source projects they knew well, and with AI tools they took 19% longer.
They had expected to be about 24% faster. The authors said the result might not hold for newer models, and eWeek's report on the METR study is a fair summary of the caveats.
Put those together and a sensible reading emerges. A fresh project with a clear specification, built on a popular stack, is where AI helps most. Changing an old system that one team has maintained for years is where it helps least.
Most small business projects look like the first kind.
Where does the money in a quote actually go?
Mostly into building, then testing, then product management, and AI only touches the first slice directly.
Magora's breakdown of a developer-day puts about 75% on development, 13% on testing and 12% on product ownership. That is one firm's split, but it is a useful picture of where a bill goes.
pie showData
title "Where a developer-day goes (Magora, 2026 estimate, %)"
"Development" : 75
"Testing and QA" : 13
"Product ownership" : 12
Suppose AI halves the development slice. Using those proportions, the total falls by about 37%, not 50%, and only if testing and product work stay the same. In practice they often grow, because more code produced faster has to be checked by someone, and the METR report notes that AI tools can introduce errors and security flaws.
That arithmetic is illustrative, not a measurement. It explains why "AI cuts software costs by 90%" claims are wrong, and why "nothing has changed" claims are also wrong.
Which costs does AI not remove?
Five of them: working out what to build, design decisions, security and testing, hosting and running costs, and the ongoing maintenance.
Working out what to build is the most expensive mistake in software. A tool can produce a feature in an hour that nobody needed. A good discovery phase, where someone talks to the people who will use the system and writes down the rules, still takes days.
Design and usability need judgement about real users. Testing needs someone who understands what could go wrong, particularly where money or personal data is involved.
Then there are running costs. Hosting, databases, email and SMS, payment fees and monitoring come monthly whatever wrote the code. Softomate's guide claims post-launch costs make up 60% to 70% of five-year spend, which is a vendor's figure but matches the general experience that the build is the smaller part.
Finally, the thing is a living system. Libraries need updating, a third-party service changes its API, and someone has to be around when it breaks. A software subscription includes all of this. A custom build does not, unless you pay for it.
How do we build now?
Our standard stack is Next.js, Supabase, Stripe and Vercel, with Claude Code doing much of the first-draft coding under a developer's direction and review.
That choice affects what you pay for. The framework and hosting are mainstream, so any competent developer can pick the project up. The database and authentication are managed services, so there is less to build and less to secure.
The AI handles repetitive code such as forms, tables and API calls, and a person decides the structure, reviews every change and tests the result.
This works best for projects such as internal dashboards, client portals, quoting tools and booking systems, where the rules are clear and the user group is small. Professional services firms often fit this, with onboarding portals, document trackers or time-and-fee reporting that off-the-shelf software handles poorly.
It works less well where the system must connect to old software without a proper API, handle very high volumes, or meet a strict regulatory standard from day one. Those projects still need more time and more senior people, and we say so at the start.
How should you read a custom software quote now?
Look for itemised phases, a stated assumption about AI tooling, and a clear answer to who owns the code and the accounts.
Here are the questions we would ask any supplier, including ourselves:
- Which parts of the build use AI coding tools, and how much of the saving is passed on?
- How is the price split between discovery, build, testing and handover?
- Who reviews AI-written code, and what testing happens before launch?
- Will the code sit in a repository and hosting account in my company's name?
- What are the monthly running costs, and what do you charge to maintain it?
- What happens to personal data, and which services see it? (Our UK GDPR and AI tools guide covers this.)
A quote that cannot answer the first question is probably the old pricing model with a new label. A quote that promises a complete system for a tiny sum, without a discovery phase, deserves similar suspicion.
For retail and ecommerce businesses, where ecommerce brands and retailers often want custom inventory or order tools, the same logic applies. Check whether your platform's existing integrations cover 80% of the need before paying for a build.
What does a smaller quote not mean?
It does not mean a smaller commitment, and the cheaper first release can hide a larger long-term bill.
A system that costs less to build still needs a person responsible for it. If the quote has no line for maintenance, the cost has not disappeared; it has moved to the day something breaks.
Put a monthly retainer or a defined support arrangement in the contract, and ask what it covers: security updates, dependency upgrades, backups and a response time.
It also does not mean you should skip discovery to save money. Faster building makes it cheaper to make the wrong thing, which is a reason to spend the first week on the question.
A short, paid discovery that produces a written specification is worth having even if you then build elsewhere, because it forces the decisions that otherwise surface halfway through the project.
Finally, a smaller quote does not mean lower standards for testing. If anything, code written quickly needs more checking, so ask to see how the supplier tests before you sign.
Should you build, or just buy a tool?
If a subscription covers most of what you need, buy it; build when the process is your competitive edge or no tool fits.
That decision is now more finely balanced than it was, because a custom build costs less than it used to. We lay out the test in build vs buy for AI tools. A list of your repeated tasks, with the hours each one takes, is the best first step.
If you want an itemised view of what a specific project would involve, with the AI assumptions stated, our AI automation services page explains how we scope custom work.
You can also contact us with a short description of the system you have in mind, or read what we think a good supplier looks like in what people mean by "AI agency".
