The build is mostly not building
Connecting a model to a workflow is days of work. Everything around it — deciding scope, writing down rules, resolving contradictions, and agreeing what good looks like — is where the effort actually goes, and it is staff time rather than supplier cost.
1. Deciding the scope properly
Narrowing to one workflow with explicit exclusions takes real conversations with the people doing the job. Skipping it is the most common cause of a deployment that never stabilises — hence the job description being the first artefact.
2. Writing down what nobody wrote down
This is usually the largest line and the one nobody budgets. Answering thirty questions in writing and resolving conflicting documents is the knowledge base work, and it takes weeks rather than days.
You are not buying software. You are paying to write down how your business actually decides things.
3. Building the evaluation
Twenty real inputs, scored by someone who knows the job, plus the never-events list. It is a small amount of work with an outsized effect, and skipping it is why a demo never becomes a product.
4. The first month of corrections
Reviewing everything while the gaps surface is a real cost with a real end date. Budget it explicitly — the first month is mostly documentation, not configuration.
What it buys beyond the system
The written knowledge outlives the deployment and makes every human hire faster. That is a genuine return even in the case where the AI employee is later switched off.