Off-the-shelf AI tools are excellent at generic work: drafting, transcribing, summarising. They stall when the job runs on your rules, lives in your systems, or carries your risk. If your team is building workarounds around a tool, or paying for licences nobody opens, that's the signal to consider building something that fits.
What are off-the-shelf AI tools genuinely good at?
Generic work, and they are genuinely good at it: drafting documents and emails, transcribing meetings, summarising long material, answering common questions. These are jobs millions of businesses do in roughly the same way, which is exactly the problem a product team can spend years polishing.
The economics are hard to argue with. For tens of pounds per person per month you get capability that would have been science fiction a few years ago, live the day you sign up, improving without you lifting a finger. If a draft comes out mediocre, you edit it and nothing breaks.
So let's be plain: if your problem is generic, being wrong is cheap, and the data isn't sensitive, buy the tool. We say this as people who make money building the alternative. The rest of this guide is about the jobs that don't fit that sentence.
Where do they stall?
At the point where the job stops being generic, which in most businesses is exactly where the real time goes. Three walls come up again and again:
- Your rules. The report has to reconcile this figure against that one; the payslip check depends on your rates and exceptions; the candidate match runs on your criteria. A generic tool doesn't know your rules, and prompting harder doesn't teach it them reliably.
- Your systems. The work lives across your accounts package, CRM and spreadsheets. A tool that sits outside those systems leaves a person to ferry data in and out, which quietly becomes a job in itself.
- Your risk profile. When a wrong answer means a mis-paid employee or a compliance question, a chatbot's best guess is not a control. And when the data is payroll, client records or candidate details, "paste it into the vendor's cloud" may not be an option at all.
The market data shows how often the wall gets hit. In a 2025 S&P Global Market Intelligence survey, 42% of companies had abandoned most of their AI initiatives, up from 17% a year earlier, and the average organisation scrapped 46% of AI proof-of-concepts before they reached production. Tools that were easy to start with turned out to be hard to finish with.
What are the hidden costs of forcing a fit?
Workarounds and waste, and neither shows up on the invoice. When a generic tool almost fits a specific job, the gap doesn't close itself: your busiest people close it, by hand, every time the job runs. They reformat what the tool produces, re-check what it got wrong, and maintain the folklore of prompts and fixes that makes it usable. The subscription looks cheap because the expensive part is paid in their hours.
Then there's the software nobody uses. Zylo's 2026 SaaS Management Index, a study of large, mostly US organisations, found 46% of software applications go underutilised or unused. Licence counts drift up, usage drifts down, and the finance team notices at renewal, if at all. AI subscriptions are following the same well-worn path: signed with enthusiasm in January, unopened by March, renewed by default in December.
None of this means the tools are bad. It means a generic tool applied to a specific job transfers the cost of the mismatch onto your team, invisibly, indefinitely.
When is it time to build?
When the job is specific to you and the mismatch costs are large. Condensed from our full build-or-buy guide, five signals point to building:
- The job runs on your rules, and a generic tool can't learn them properly.
- Errors cost real money: a mis-payment, a compliance breach, a lost client.
- The data can't leave: client records, payroll, candidate details that shouldn't live in a vendor's cloud on a vendor's terms.
- The job eats days per month, not minutes, so a project cost has something real to pay back against.
- Owning it would be an advantage: you can't build a moat out of a subscription everyone else can buy.
One signal is a conversation. Two or more on the same job is a candidate for a build, and worth an hour of arithmetic before you renew anything.
What does a graceful mix look like?
Buy generic, build specific, and let each do what it's good at. The businesses that get this right don't rip out their subscriptions; they keep the tools for drafting, transcription and summaries, and build systems for the handful of jobs where the week actually goes: the monthly report, the line-by-line check, the matching and chasing that runs on their rules.
A well-built custom system complements the tools you already pay for rather than competing with them. It lives inside your systems and your rules, keeps sensitive data in your own cloud, and leaves a full audit trail; the bought tools carry on doing the generic work around it. You end up with fewer subscriptions doing more, and one owned system doing the job no subscription could.
How do you decide?
Job by job, with arithmetic, not tool by tool with a demo. For each time-eater in the week, ask whether it's generic or yours, what a mistake costs, whether the data can leave, and how many hours a month it burns. The full method, including the cost comparison over a couple of years, is in our build-or-buy guide and its companion on what custom AI costs.
Or bring us one real job and we'll do the sums with you. A free Impact Call ends with a one-page opportunity map, and "keep the subscription, don't build" is an answer we give regularly, because it's often true.
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.
- S&P Global Market Intelligence, 451 Research "Voice of the Enterprise: AI & Machine Learning, Use Cases 2025" (n=1,006 IT and business leaders): 42% abandoned most of their AI initiatives, up from 17%; 46% of proof-of-concepts scrapped before production. spglobal.com. Published 2025, accessed 8 July 2026.
- Zylo, SaaS Management Index 2026 (study of large, mostly US organisations): 46% of software applications underutilised or unused. zylo.com. Published 2 April 2026, accessed 8 July 2026.
- AI Nativ.es delivery experience, 2026: client work is described without names until we have written permission to use them.