We build custom AI systems, so treat what follows with appropriate suspicion. It is also the advice we give at intake, including when it costs us the engagement: if a tool fits your workflow off the shelf, buy the tool. The interesting question is not build versus buy in general. It is what to do about the very common case in the middle, where the tool fits about seventy per cent.
Seventy per cent is where the money gets lost, because seventy per cent looks like success in a demo and behaves like failure in production.
The rule
Ask one question about the missing thirty per cent: is it the part that makes you different?
If the gap is incidental — a field you would name differently, a report format, a step you could adapt to — buy the tool and adapt. You will be productive next week and the difference was never worth defending.
If the gap is your actual edge — the judgement your best people apply, the sequence that makes your margins work, the rule that exists because you learned something expensive in 2019 — then renting a generic version of it is the worst available option, because you are paying to have your differentiator smoothed away.
A 70% fit is a bargain when the missing 30% is cosmetic, and a slow disaster when the missing 30% is the reason you win.
Why the middle case is so expensive
The failure is not that the tool does seventy per cent. It is what fills the remaining thirty.
Usually it is people. A spreadsheet beside the tool. A Slack channel where the exceptions get handled. A senior person who reviews everything because the tool cannot encode the one rule that matters. That labour is invisible on the invoice and often exceeds the licence cost several times over.
Worse, the workaround is where all the interesting information now lives — and it lives outside any system, so nothing accumulates. You are paying a subscription to have your most valuable decisions happen in a spreadsheet.
Three cases where buying is clearly right
Being specific about this matters, because "it depends" is not a decision rule.
And one where it clearly isn't
When the workflow is where your margin comes from, fires often enough to produce real volume, and already has a person applying judgement to every case — that judgement is a training signal you are currently throwing away.
That is the profile worth owning. Not because custom is better in the abstract, but because in that specific shape the thing you would be renting is the thing you would be trying to protect.
How to run this on your own list
For each tool you are evaluating or already paying for, write down the thirty per cent it does not do. Then mark each item cosmetic or differentiating.
An all-cosmetic list is a buy. A list with two or more differentiating items in a high-volume workflow is worth pricing properly — and a Scan will tell you which, along with what the workaround is costing you today, before anyone commits to building anything.
If it comes back cosmetic, we will say so, and the cheapest engagement we ever run is the one that ends with "keep the tool."
