Buy off-the-shelf AI when your problem is generic and a tool already does 90% of it. Build custom AI when the work is specific to how your business runs, when errors cost real money, or when the tool would touch data you can't hand to a vendor. Most businesses need a mix of both.
What does "building" custom AI actually mean?
Building custom AI means paying someone to assemble a system around one specific job in your business: your data, your workflow, your rules. It does not mean training a giant model from scratch. Nobody sensible does that for a small business, and anyone who says they will is selling you something.
In practice, a custom build wires proven AI models into the way your team already works. The model is a component, like the engine in a van. The build is everything that makes it useful: the connection to your systems, the checks around it, the rules for when a person takes over, and the boring plumbing that means it runs every day without someone pushing it.
A concrete example from our own work: a payroll team was checking every payslip line by hand, month after month. The custom system now does the checking, flags the handful of lines that genuinely need a person, runs in the client's own cloud, and keeps a full audit trail. No off-the-shelf tool could do that job, because the job is made of that company's specific rules.
When is off-the-shelf AI the right answer?
Buy a tool when your problem is common to millions of businesses and a product team has already spent years polishing the answer. This is the right call more often than a custom builder would like you to believe.
Buying wins when all of these are true:
- The job is generic. Writing first drafts, transcribing meetings, summarising documents, answering common questions. Your business does it the same way as everyone else.
- Being wrong is cheap. If a draft is mediocre, you edit it. Nothing breaks, no customer is harmed, no regulator calls.
- The workflow is standard. The tool's idea of how the work happens roughly matches yours, so your team isn't fighting it.
- The data is not sensitive. You're comfortable with the content living in the vendor's cloud, under the vendor's terms.
A good subscription tool gives you speed for tens of pounds a month. If you've read this far and your problem fits that list, stop here and go buy the tool. That's the honest answer, and it costs you nothing to take it.
When does a custom build earn its keep?
Build when the work is specific to you and the value of getting it right is large. The pattern we see across real builds is consistent: the jobs worth building for are repetitive, rule-heavy, and expensive precisely because they are yours.
Building wins when one or more of these is true:
- The job runs on your rules. Payroll exceptions, compliance checks, report assembly, candidate matching against your criteria. A generic tool doesn't know your rules and can't learn them properly.
- Errors cost real money. When a wrong answer means a mis-paid employee, a compliance breach or a lost client, you need a system with checks designed for your risks, not a chatbot's best guess.
- The data can't leave. Client records, payroll, candidate details. A custom system can run inside your own cloud so nothing is handed to a third party. UK GDPR stops being a worry and becomes a design input.
- The job eats serious time. One team we work with loses eight to ten days, every single month, building one report by hand. At that scale, a system that does the assembly pays for itself quickly, and keeps paying every month after.
- The tool would be a moat. If the system encodes how you win work or serve clients better than rivals, owning it outright matters. You can't build a moat out of a subscription everyone else can buy.
What does each route actually cost?
Buying looks cheap and building looks expensive, until you do the arithmetic over a couple of years. The honest comparison is between a subscription that never ends and a project cost that stops.
Off-the-shelf pricing is per seat, per month, forever. Ten seats on a serious AI tool is a few thousand pounds a year, every year, rising when the vendor's investors want it to. You also pay in workarounds: the hours your team spends forcing a generic tool to fit a specific job never show on an invoice. And much of the spend quietly evaporates: industry analysis of company software portfolios finds nearly half of paid-for applications go underutilised or unused.
On the build side, there is no honest universal price. Published agency pricing guides put UK custom AI work anywhere from roughly £21,000 to £400,000 and beyond depending on complexity, and that spread tells you the real answer: the job defines the number. Treat those figures as a vendor-published anchor, not independent research.
A custom build is a one-off project cost plus optional support. Ours run as a paid mapping phase first, then a build quoted per project and shipped in two-week sprints, then a simple monthly amount for support if you want it, which you can stop any time. The number depends entirely on the job's complexity, which is why we won't print a price here and neither should anyone else who hasn't seen your week. What we will say in advance is the shape, and we'll put a straight number on it after mapping.
| Question | Off-the-shelf tool | Custom build |
|---|---|---|
| Cost shape | Per seat, per month, forever. Rises over time. | Project fee that stops, plus optional monthly support you can cancel. |
| Time to value | Days. Sign up and start. | Weeks. Mapping first, then two-week sprints. |
| Fit to your workflow | You adapt to the tool. | The tool adapts to you. |
| Your data | Lives in the vendor's cloud, on their terms. | Can stay in your own systems, on yours. |
| Ownership | You rent access. Cancel and it's gone. | You co-own the system. It's still yours if we part ways. |
| When it wins | Generic jobs where being wrong is cheap. | Specific, rule-heavy jobs where errors cost real money. |
For a deeper breakdown of build pricing, including what makes a project cheap or expensive, see our guide to what custom AI costs.
Who owns what you pay for?
With a subscription you own nothing. The day you stop paying, the capability walks out the door, along with any workflow your team built around it. Your prompts, your fine-tuning, your integrations: read the terms, because usually the useful parts aren't yours to take.
With a custom build, ownership is whatever the contract says, so read that too. Our standard is co-ownership: the client owns what we build, can see inside it, can host it in their own cloud, and can switch any part off. No lock-in is a design decision, not a slogan, and it's the single biggest difference between the two routes when something goes wrong or the relationship ends.
Ownership also covers the data the system learns from and produces. In a well-built custom system, your data stays in your systems, nothing is sold on, and nothing is used to train anything outside your business. With a bought tool, the answer lives in the vendor's privacy policy, and it can change with thirty days' notice. We've written more on this in who owns the AI you pay for.
Where does each route go wrong?
Both routes fail, in different ways, and anyone selling either one without saying so is not being straight with you.
Bought tools fail quietly. The classic pattern: a subscription gets signed in January, the team pokes at it for a fortnight, it doesn't quite fit how the work happens, and by March nobody has opened it. The invoice keeps arriving. The tool wasn't bad, it was generic, and the gap between generic and your reality landed on your busiest people to bridge. The numbers back the anecdote: in S&P Global's 2025 survey of over a thousand IT and business leaders, the share of companies abandoning most of their AI initiatives jumped from 17% to 42% in a single year.
Custom builds fail loudly. A build goes wrong when the problem was never mapped properly: the wrong job gets automated, scope drifts, and six months later there's a demo and no working system. This is common. By some estimates cited by RAND, more than 80% of AI projects fail, and MIT's 2025 research found roughly 95% of enterprise AI pilots deliver no measurable profit impact. That failure is expensive and visible. It's also avoidable, which is why we insist on a paid mapping phase before any code: a plan first, so the build targets the job that's actually eating the week, and you see working software every fortnight rather than a slide deck at the end.
The common thread in both failure modes is the same: nobody did the honest arithmetic up front about which jobs were worth touching at all. Some jobs aren't worth automating, and hearing that early is worth more than a glossy roadmap.
How do you actually decide?
Don't start with the technology. Start with the week. Write down where your team's hours actually go, then run each time-eater through five questions:
- Is this job generic or ours? If a million businesses do it identically, buy. If it runs on your rules, build is on the table.
- What does a mistake cost? Pennies: buy. Real money, real clients, or a regulator: build, with checks designed for that risk.
- Can the data leave? If handing the data to a vendor makes you uneasy or breaches your obligations, that's a build signal.
- How many hours a month is it eating? Small numbers don't justify a project. Days per month do.
- Would owning this outright be an advantage? If yes, a rented tool can't give you that.
Mostly "buy" answers: buy, and spend the saved money on something else. Two or more "build" signals on one job: that job is a candidate, and the next step is not a contract, it's a mapping conversation where someone does the arithmetic with you and tells you straight if it's worth building. That's what our free Impact Call is for, and "no, don't build this" is a real answer we give.
Sources
Figures and claims in this guide draw on our own delivery work and the sources below. We only publish numbers we can stand behind.
- AI Nativ.es delivery experience, 2026: the payroll checking build and the reporting build described are real client projects, told without names until we have written permission to use them.
- Zylo, SaaS Management Index (2026 edition): 46% of applications underutilised or unused. zylo.com. Published 2 April 2026, accessed 8 July 2026. Vendor research, enterprise-skewed sample.
- Appinventiv, "How much does AI software cost in the UK?": £21,000 to £400,000+ indicative range. appinventiv.com. Published 25 June 2026, accessed 8 July 2026. Vendor pricing guide, not independent research.
- S&P Global Market Intelligence, 451 Research, "Voice of the Enterprise: AI & Machine Learning, Use Cases 2025" (n=1,006): companies abandoning most AI initiatives rose from 17% to 42%. spglobal.com. Published 2025, accessed 8 July 2026.
- RAND Corporation, "The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed": by some estimates more than 80% of AI projects fail. rand.org. Published 13 August 2024, accessed 8 July 2026.
- MIT NANDA initiative, "The GenAI Divide: State of AI in Business 2025", via Fortune: about 95% of enterprise generative AI pilots show no measurable P&L impact. fortune.com. Published 18 August 2025, accessed 8 July 2026. Methodology debated; read alongside the RAND and S&P figures.