Second Brain AI Agency
Operations ยท 6 minute read

Why AI projects stall halfway

Not at the choice of tool, and not on the technology. They stall where the work meets itself: the handover, the second data entry, the check by the one person who still sees the whole process.

Dudley Haddocks {date}

A company of thirty people rarely has an AI problem. It has a process that grew without ever being redesigned. At eleven people it worked because everyone could see all of it. At thirty it works because steps have been glued on: an extra email to confirm something, a second entry because two systems do not talk, a check by the person who still knows the exceptions.

Nobody decided on those steps. They are written down nowhere either. That is why they stay invisible until someone measures them, and why they survive every tool placed on top of them.

Three ways it stalls

In conversations with companies that have already tried something with AI, the same three patterns keep coming back. All of them are understandable, and none of them is a technical problem.

01
The tool was chosen before the problem was defined A demo convinces, and the use case gets found afterwards. What follows is a pilot nobody can settle, because there was never a number to measure against.
02
The same data sits in three places Any model is as good as its input. Without a single source it produces answers slightly different from what the team already knows, and that is enough to lose their trust.
03
Nobody inside the company owns it The agency leaves and the knowledge leaves with it. Six months later the team is doing it by hand again, and the conclusion is that AI did not work for them.

What measuring gives you

The moment you put hours per process on paper, the conversation shifts. Not to which tool, but to which order. And almost always part of the list turns out to need no AI at all.

Example, a 28-person consultancy
Assembling quotes 6.5 h partly AI
Hours into invoicing 4.0 h no AI
Client onboarding 3.0 h no AI
An illustration of the format, not a result we promise. In a real diagnostic every line names its source.

Two of these three lines are solved by one integration and a fixed order. That is not an exciting answer, but it is the cheapest time back in the list. The line where AI does add something only works once the process underneath it holds.

Where to start

You do not need an agency to take the first step. Take your three heaviest processes, ask the people who run them how much time they spend per week, and ask where work sits waiting. Write the answers down with a number next to them. That is half a diagnostic.

The other half is the order: hours per week divided by the lead time of the fix. Anyone holding that on a single page can judge any proposal from any agency, including ours.

Want that page made for you?

The diagnostic takes two weeks and delivers one document. After that you decide, including deciding to execute it yourself.

Start the intake How it works