Buy, switch or build
Almost every tool decision in the AI era comes down to the same question: keep paying licence fees for the off-the-shelf product that is now retrofitting AI, switch to an AI-native product, or build it yourself. The answer does not depend on the tool but on seven properties of your own operation.
The three routes
Off-the-shelf product
Grown feature set, deep integrations, contract readiness and administration built for many users. You pay with licence costs that grow with seats and contacts, and with slow development.
AI-native product
Younger, usually much cheaper, faster to operate because AI is built in rather than bolted on. You pay with a thinner certification record, less integration breadth, and features needed for evidence sitting in the most expensive plan.
Build it yourself
Full data sovereignty, exact fit, no licence. You pay with one-off build effort and permanent operating load: on-call, permissions, evidence, maintenance.
Seven criteria
The more criteria point one way, the clearer the route. Where they contradict each other, the criterion with the highest damage on failure usually wins, and that is almost always the duty to provide evidence.
01
Process breadth
Does the tool cover one process or half a company?
Buy broad processes. A narrow, clearly bounded workflow is the only case in which building yourself regularly stays cheaper than a licence.
02
Number of users
Do three people use it or three hundred?
Permissions, training and support grow faster than user count. That work is already paid for in an off-the-shelf product.
03
Burden of evidence
Does the use have to be evidenced to auditors, customers or a regulator?
The higher the burden of evidence, the more expensive building yourself becomes, because logging, permissions and documentation have to be built too. A vendor with certifications sells that work along with the product.
04
Data sovereignty as an edge
Is the data in this process a differentiator or just operating material?
Where your own data carries the edge, the process belongs in your own house. Where it is only operating material, data sovereignty is an argument without a return.
05
In-house capability
Are there two people who can do it, or one?
Building yourself with exactly one capable person is not building, it is a personnel risk. The second person is the actual precondition.
06
Licence cost threshold
Does the annual licence exceed build effort plus two years of operation?
Only when the answer is clearly yes does building add up. A narrow yes means no, because build effort is regularly underestimated and operations regularly forgotten.
07
Maintenance horizon
How long does the result have to live?
For weeks and months the fastest route is the right one. For years, what decides is who still knows in year three how it was built.
Where building yourself usually fails
Not at the building. Four places keep coming up, and they only come up after the euphoria.
Operations
Building is a project, running is a state. On-call, updates and backups appear in no effort estimate and still cost every month.
Permissions and tenants
Who may see what, and how do you prove it. Retrofitting a permission model costs more than building it in, and it is almost always retrofitted.
Evidence
Logging, deletion periods and documentation decide the approval. They are invisible until someone asks and blocking the moment someone does.
Bus factor
What you build survives its builder leaving only if a second person can read it. That is a staffing question, not a technical one.