For most small businesses, buying beats building: a subscription of around £29 a month is cheaper than any custom build until your usage, user count or process is unusual enough that no product fits.
AI coding tools have made building cheaper, and that has made the decision closer than it was. It has not made it a coin toss.
This guide works through the build vs buy software question with a three-year cost comparison, six tests that decide it, and the point where paying per user turns the answer round.
Is it cheaper to build or to buy?
Buying is cheaper in the first year almost every time, and the real question is where the lines cross over three to five years.
The reason is simple. A subscription spreads the cost of building, hosting, securing and updating a product across thousands of customers. A custom build asks you to carry all of it yourself. Even when AI tools halve the coding, the discovery, testing and upkeep remain, as we explain in the end of the £20k custom web app quote.
Evidence on AI projects points the same way. A 2025 MIT report on generative AI in business, covered by Fortune, found that tools bought from specialist vendors and built through partnerships succeeded about 67% of the time, while internal builds succeeded about one-third as often.
The study drew on 150 leader interviews, a survey of 350 employees and 300 public deployments, and it concerned larger organisations. Fortune's piece on the MIT report on AI pilots has the details. Treat it as a pointer, not a law, but it matches the common experience that internal builds underestimate the work.
What does the three-year maths look like?
Here is a comparison using typical UK small-business figures. Swap in your own numbers before you decide.
Suppose a tool does a job for your team of three. The subscription is £29 a month per user. A custom build costs £8,000, plus £30 a month for hosting and £150 a month for maintenance and fixes.
| Cost over 36 months | Buy | Build |
|---|---|---|
| Setup (10 hours at £50) | £500 | |
| Build | £8,000 | |
| Subscription, 3 users | £3,132 | |
| Hosting | £1,080 | |
| Maintenance | £5,400 | |
| Total | £3,632 | £14,480 |
The subscription wins by almost £11,000. To justify the build, it has to deliver about £300 a month more in value, through saved time, extra sales or a capability no product offers.
Now change one input: your team grows to 20. The subscription becomes £580 a month, or £20,880 over three years plus setup, while the build costs about the same as before, because most of it was fixed. The build now wins.
That is the point of the exercise. Per-seat pricing means buying is cheapest for small teams and most expensive at scale, whereas a custom system costs about the same for ten users or fifty. Both lines are in your hands, so work them out before you decide.
Which six tests decide it?
Six questions, answered honestly, settle most cases.
- Does a product do at least 80% of what we need? If yes, buy.
- Is this process how we beat competitors? If it is the heart of your business, building may be worth it.
- How many people will use it, now and in three years? Multiply by the per-seat price.
- How sensitive is the data, and who must hold it? This can rule options in or out.
- How often will the process change? Products change at their pace, and custom builds at yours.
- Can we leave? Check that a product lets you export your data in a usable form.
Test two comes from the old distinction between "core" and "context" work. Core is what makes customers choose you. Context is everything else a business has to do. Buy context and consider building core.
For most firms, almost everything on the AI list is context. Replying to enquiries, chasing invoices and booking appointments are things every business does, and products exist for them.
A dental practice is a good example: clinics and dental practices can nearly always buy practice management software with reminders and booking built in, and gain little by building their own.
What about the middle option?
Buy the product, then connect and extend it with a workflow tool and small pieces of custom code.
This is where most good AI automation work sits. You rent the model and the core software, and build only the glue: the rules that move information between systems and the checks before anything reaches a customer.
The tools that do this vary in cost and in who has to maintain them, and our comparison of Make, Zapier and n8n covers that choice.
It keeps your upfront cost low and your exit options open. If a vendor raises prices or closes, the glue can be pointed at another product. A full custom system is harder to move away from because you are the vendor.
Where do the AI models themselves fit?
You will almost always rent the model and never build one.
Training a language model from scratch costs far more than any small business will spend, and fine-tuning one rarely pays off at small scale. What you decide is whether to use a finished product with AI inside it, or to call a model directly from your own workflow and write the surrounding logic yourself.
The product route is quicker and the vendor takes care of updates. The direct route gives you control over prompts, data handling and cost, and it is where a tool like Make or a small piece of custom code does the work.
Pay-as-you-go model pricing usually favours the direct route at low volume and can favour a flat-rate product at high volume, so run the numbers at your expected usage and at three times that.
Whichever you choose, keep the prompts, templates and test examples in a document your company owns. They are the real asset. The model behind them will change several times a year.
What do people forget when they buy?
Price rises, per-seat creep, lock-in and data protection.
Subscription prices rise, often with a year's notice or less. Seats multiply as people join. Features you rely on move to a higher plan. Check the terms for how and when the vendor can change them.
Lock-in is quieter. If your customer records, templates and history live inside a product, leaving means rebuilding them elsewhere. Test the export before you commit, not on the day you want to leave.
Data protection applies to every tool that touches personal data. A bought AI product is a processor, and you need a contract, a clear picture of where the data goes and an answer on whether your inputs train its models.
What do people forget when they build?
The upkeep, the bus factor and the second year.
Softomate Solutions, a UK software supplier, claims post-launch costs make up 60% to 70% of five-year spend on a custom web application. That is one vendor's figure, but it fits what we see: the build is the start of the cost, not the whole of it.
The bus factor is the number of people who would have to disappear before nobody understands the system. For many small builds, it is one. Make sure the code is in your repository, documented and understood by someone besides its author.
The second year is when it matters. Libraries need updating, a service you depend on changes, and the person who built it has moved on. Budget for it from day one, as an explicit line.
Retailers feel this especially, since ecommerce brands and retailers often start with a small custom order or stock tool and then end up maintaining it through every platform change.
Security is the other cost that hides in a build. A bought product has a team watching for vulnerabilities and patching them. Your custom tool has whoever you hired, and if that person is busy, nobody. Ask who will apply security updates, how quickly, and what happens if the answer is nobody for three months.
Insurance and client contracts can add a further requirement. Some customers ask suppliers about their software and data handling, and a bespoke system is harder to answer for than a well-known product with a published security page.
How do you decide for your own business?
List the candidate jobs, run the six tests on each, and cost the top two over three years.
flowchart TD
A[Job to automate] --> B{Does a product fit at least 80%?}
B -- Yes --> C{Per-seat cost at three-year headcount acceptable?}
C -- Yes --> D[Buy]
C -- No --> E[Compare with a custom build]
B -- No --> F{Is it core to how you win?}
F -- Yes --> G[Build, with upkeep budgeted]
F -- No --> H[Buy and extend with workflow tools]
Finding and ranking the jobs is the first step: list them with the hours each one takes. After that, the cost table above takes about twenty minutes per candidate.
We do both sides of this. Some clients need a subscription configured and connected, and a few need a bespoke tool built, and we will tell you which.
Our AI automation services page describes how we scope each, our AI agency guide explains what to ask any supplier, and you can contact us with a short description of the job you are weighing up.
