We are spending more time than we expected doing one specific kind of work: figuring out where the AI vendor's pitch deck and the AI vendor's actual product stop matching.
Here is a representative example. A client wanted to add an MCP connection between Claude and their accounting system. The vendor's marketing said live data, agentic queries, the whole thing. We bought the story. We pulled it into a real engagement and started building against it.
Then we ran the first real query.
The MCP was not pulling from the production accounting system. It was pulling from a data warehouse downstream of the accounting system. The warehouse refreshed on a multi-day cache. And on certain query types the response payload was so narrow that the agent's answer was functionally useless, even when the cache was fresh. None of this was disclosed in the documentation. All of it was discoverable only by running real queries against real data and watching what came back.
The time burn to find that gap is real. Days, in this case.
The gap is not the vendor's fault, exactly. The marketing department is shipping decks at AI speed. The engineering team is shipping integrations at enterprise software speed. Those two clocks are not synchronized right now. Nobody at the vendor is lying. Everybody at the vendor is selling the version of the product the engineering team is still building.
We are seeing this on multiple projects, with multiple vendors, across multiple stacks. The specific gap is different every time. The pattern is the same. The pitch is six to twelve months ahead of the implementation, and the customer is the one who pays for the gap in calendar time and engineering hours.
Here is the part that matters for the mid-market firm shopping AI tools right now. You are not going to find these gaps by reading the spec sheet. You are going to find them by running real workloads against real data and watching what breaks.
That work has a name. It is called diligence. It used to be standard procurement practice. Somewhere along the way it stopped getting applied to AI vendors. The demos look too good in the room.
Reinstate it.
If you do not have a senior operator on the inside who is willing to call a flashy vendor demo a flashy vendor demo, you are going to keep paying for the gap.