When Not To Build A Bespoke AI System
Bespoke is the wrong answer more often than it is the right one, and the tests are simple enough to apply in a meeting.
Most organisations asking for a bespoke AI system should buy a product instead. That is an unusual thing for anyone who builds bespoke systems to write down, but the alternative is a project that costs more than the problem and takes longer than the patience available.
Test one: is the process actually unusual
Ask whether a competent competitor would recognise the process immediately. Invoice processing, expense approval, meeting notes, customer records, first-line support: these feel specific from the inside and are near-identical across an industry from the outside. If a product models it adequately, the residual difference is rarely worth a build.
The genuinely unusual is narrower than it feels: a process shaped by a regulator specific to that organisation, an asset nobody else operates, or a judgement encoded over years that exists nowhere else.
Test two: is the volume there
A bespoke system has to earn its cost against the time it saves. A workflow run fifty times a week by two people rarely does. The same build against a workflow run five hundred times a day usually does, comfortably. This is arithmetic, and it should be done before the design conversation rather than after.
Test three: is the constraint real
Many organisations believe their data cannot go to a cloud when in fact it can, with controls. DORA, the FCA and PRA regimes, EBA guidance and GDPR all permit it under conditions. The genuinely narrow cases are material classified above a threshold, ITAR-controlled work, isolated operational technology, and cases where an impact assessment lands negative.
If the constraint turns out to be a preference rather than a prohibition, the cloud options widen considerably and a product may serve. That is worth establishing honestly, because building an on-premise system to satisfy a constraint that does not exist is an expensive way to feel safe.
Test four: is anyone going to own it
A bespoke system needs somebody inside the organisation who understands it, wants it, and will keep it aligned to the process as the process changes. Without that person it decays into a system nobody quite trusts, which is worse than not having built it.
“If a studio already does what you need, take the studio. It is faster, it costs less, and any supplier worth commissioning will tell you so before taking the work.”
What passing all four looks like
A process specific to the organisation, run at volume, under a constraint that genuinely rules out the alternatives, with a named owner who wants it. That combination is not common, and where it holds the case is usually obvious to everyone in the room within twenty minutes.
Where it does not hold, the useful answer is a product recommendation and a shorter invoice.