Guide
Why an AI project costs more than the licence
Updated
The gap between a vendor quote and a project cost is not a markup or a surprise. It is three categories of work that nobody quoted because nobody sells them.
Data is the project
Almost every business AI project turns out to be a data project with a model attached. The work is finding the data, getting permission to use it, making it consistent, and discovering that a field everybody relies on has been filled in three different ways since 2019.
This scales with the mess rather than with the volume, which is why two companies of identical size can pay wildly different amounts for the same software.
Integration is not included
A tool that cannot reach your systems is a demo. Connectors to CRMs, ticketing systems, document stores and finance systems are work, and the older the system the more work.
List the systems it must touch and ask, for each one, who writes the connector and who maintains it when the other end changes.
Change is the one that decides the return
People have to work differently, and most will not unless the new way is genuinely easier and somebody senior expects it. This is the line most often cut and the one whose absence guarantees no return.
Count the people whose work actually changes and budget for them properly. It is not a training day.
Then year two
Running cost, monitoring, evaluation and the periodic rework as inputs and models change. Budgeting year one and forgetting year two is how a project that worked gets cancelled.
Put a running percentage in the plan from the start so nobody is surprised by it at the first renewal.