AI development, from the first use case to the system that runs it

We pick one workload before we write any of it. The AI Readiness Assessment takes two weeks at a fixed fee and says which use case is worth building first, what data and permissions it needs, and what it would cost to run for a month at real volume.

Then comes a written rubric of what a good answer looks like, agreed with the people who own the question, and an environment that scores it. The build goes to production against that rubric, and if the scores don't hold on a change, the change goes back. Once the first workload is trusted, the next one reuses what it built.

Three costs come with an AI build, and pilots hide the third. The first is engineering, which we quote against a written scope after the assessment rather than against a brief. The second is the platform it runs on, which is usually one you already pay for.

The third is inference, what the models cost at real volume. A pilot with a dozen users never shows it, which is why the Readiness Assessment prices a production month. A promising use case often needs a different design at that point rather than a bigger budget. We don't publish rates, and these are the things that move the number.

What moves the price of an AI build

• How many systems the agent reads from and writes back to
• Which steps need a person's approval, and how many people approve
• How long a single run lasts, and whether it waits on people
• Whether evaluation data exists or has to be written with your team
• Which models the workload needs at its real volume
• Whether it has to run inside your own cloud account

A shortlist of AI development companies usually starts from a list of firms, and a list can't tell you whether anyone on it has kept an AI system running past its first month. These checks can, and we'd expect any firm on a shortlist to answer them on the first call.

Questions with answers you can check

• A system they built that's still in production, and someone there you can ask about it
• Results with a source, such as a figure the customer confirmed, rather than a percentage on a slide
• How their systems keep an agent from reading data its user can't see
• What happens when the model times out halfway through a run
• How they'll show that next month's version is better than this month's
• Whether the framework underneath is code you keep, as TechFabric Harness is