Build vs Buy AI: Should Your Small Business Use Off-the-Shelf Tools or Build Something Custom?

Sooner or later, every business owner using AI hits the same wall. The off-the-shelf tool does 80% of what you need, and the remaining 20% is exactly the part that matters — the bit that touches your pricing rules, your client onboarding sequence, your peculiar way of scheduling jobs. At that point someone suggests building something custom, a developer quotes a number, and the decision suddenly feels much larger than it did last week.

This guide gives you a way to make that decision without guessing. It covers what "building" actually involves in 2026, what both routes really cost over three years, and a short set of questions that settles the matter for most small businesses in under an afternoon.

What "building" actually means in 2026

The phrase covers three very different things, and conflating them is the single most common reason these decisions go wrong.

Configuring a tool you already pay for. Writing a custom GPT, setting up a Claude Project with your documents loaded, building a Zapier or n8n automation that moves data between systems. No code, no developer, typically a few hours of someone's time. Most people who say they want to "build AI" actually want this.

Assembling on a platform. Using a low-code agent builder or a workflow platform to create something that runs on a schedule, calls a language model, and writes results back into your CRM or spreadsheet. This needs a technically confident person — not necessarily a developer — and typically one to four weeks.

Commissioning custom software. A developer or agency writes an application that calls a model's API, stores data in your own database, and presents a purpose-built interface. For a small business this realistically starts around £8,000–£15,000 for a narrow tool and climbs fast from there.

When a vendor, a developer or an enthusiastic team member says "we could just build that," your first job is to ask which of these three they mean. The cost difference between the first and third is roughly a hundredfold.

The real cost of each route

Buying looks expensive because the price is visible every month. Building looks cheap because most of its cost is invisible until it arrives. Here is a fairer comparison over three years for a typical eight-person business automating one workflow.

Buying: a specialist SaaS tool at £80–£250 per month, so £2,900–£9,000 over three years. Add perhaps 10 hours of setup and a few hours a year of admin. The vendor absorbs model upgrades, security patches, compliance paperwork and support.

Building custom: £8,000–£15,000 to build, plus API usage of perhaps £30–£150 a month, plus hosting, plus — and this is the line everyone forgets — maintenance. Budget 15–20% of the original build cost annually just to keep it working as models, APIs and dependencies change underneath it. Realistically £16,000–£28,000 over three years, and you own every problem.

Custom work is not automatically the wrong call. But it has to clear a much higher bar than "the subscription feels pricey." If you have not yet put numbers against the workflow, our guide on how to calculate the ROI of an AI implementation gives you a model you can fill in for either route.

Five questions that settle the decision

Answer these honestly and the choice usually makes itself.

1. Is this workflow genuinely unusual, or does it just feel unusual? Most small businesses believe their processes are unique. Most are not — they are a standard process with local vocabulary attached. Before commissioning anything, spend two hours searching for vendors who serve your exact niche. Search the trade press and your industry association's supplier list, not just Google. The tool often already exists and simply uses different words for the same thing.

2. Does the workflow touch your competitive advantage? If customers choose you because of how you do this thing, it may be worth owning. If it is invoicing, scheduling, note-taking or email triage — work every competitor also does — buy it. Nobody has ever won a customer with a proprietary expense-approval process.

3. How stable is the process? If you have changed how it works twice in the last year, custom software will be obsolete before it is paid for. Buy something flexible instead and revisit in a year.

4. Who maintains it when it breaks at 7am on a Monday? Name the actual person. If the name is "the developer who built it," confirm they offer a support retainer and what it costs. If the name is you, and you do not write code, you do not have a maintenance plan — you have a future crisis.

5. What happens if the underlying model changes? Models are deprecated, repriced and updated constantly. A bought tool absorbs this for you. A custom tool passes it to you as unplanned work. This is worth reading about in depth in our piece on AI model dependency risk for small businesses.

A rough scoring rule: if you answer "unusual, core to our advantage, stable, we have a named maintainer, and we have budgeted for model churn" — building is defensible. Anything less than four of the five and you should buy.

The middle path most businesses should take

There is a third option that gets overlooked because it sounds unambitious: buy the platform, configure the last mile yourself.

In practice this means paying for a capable general tool — a team AI subscription, an automation platform, a CRM with an AI layer — and then investing in your own people's ability to configure it. A well-written custom assistant loaded with your templates, tone guidelines and pricing rules closes a surprising amount of that frustrating 20% gap, at the cost of a few hours rather than a few thousand pounds.

The gap it cannot close is data. If your problem is that your information lives in four systems that do not talk to each other, no amount of clever prompting fixes it. That is an integration problem wearing an AI costume, and it needs an automation platform or a proper integration, not a custom model.

A useful first step here is taking stock of what you already pay for. Many owners discover they are one configuration away from a capability they were about to commission. Our walkthrough on how to audit your AI tool stack takes about 90 minutes and frequently pays for itself immediately.

When building genuinely pays off

Custom work earns its keep in a narrow set of circumstances, and it is worth naming them plainly so you recognise them when they arrive.

Volume makes per-seat pricing absurd. When a tool charges per user or per record and you have thousands of either, custom economics flip. A business processing 40,000 documents a year will often find an API-based tool costs a tenth of the SaaS equivalent.

The data cannot leave your control. Certain regulated contexts — some health, legal and financial work — make it simpler to run within your own infrastructure than to negotiate data processing terms with a vendor. Check whether the vendor already offers a compliant tier before assuming this, because most now do.

The output is the product. If you intend to sell the resulting capability to your own clients, you are no longer buying a tool, you are building a product line. Different decision, different budget, different conversation.

You have genuine in-house capability. If a competent developer is already on the payroll and has spare capacity, the marginal cost of building drops substantially — though the maintenance question in the previous section still applies, and it applies harder if that person leaves.

A thirty-day decision process

If you are currently stuck, run this instead of debating it in meetings.

Days 1–7: define the job precisely. Write one paragraph describing the workflow, the inputs, the outputs and the hours it currently consumes each week. Multiply those hours by a loaded hourly rate to get the annual cost of doing nothing. This number anchors everything that follows.

Days 8–14: search properly for an existing tool. Find at least three candidates. Run free trials on the two best. Assess them against the job you wrote down, not against a feature list. Our guide on how to choose an AI vendor covers the due-diligence questions worth asking at this stage.

Days 15–21: try the configure-it-yourself route. Give one capable person a week to see how close they can get with the tools you already pay for. Even a partial success changes the economics of the decision considerably.

Days 22–30: decide, and write down why. One page: what you chose, the numbers behind it, and the conditions that would make you revisit. That page is worth more than the decision itself, because in eighteen months you will genuinely not remember your reasoning, and the market will have moved.

The bottom line

For the overwhelming majority of businesses under 100 people, the right answer is buy — and then invest the money you did not spend on development into training people to use what you bought properly. Building is justified when volume makes subscriptions irrational, when the capability is genuinely part of what customers pay you for, or when you already have the engineering capacity and a maintenance plan with a real name attached to it.

The failure mode worth avoiding is not choosing wrongly between the two. It is spending six months deciding, and doing neither. A subscription you can cancel next month is a far cheaper way to learn what you actually need than a specification document written before you understood the problem.