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.
This is why two quotes for the same AI employee can differ enormously without either being wrong. One has priced the integration; the other has priced the integration plus the weeks of extracting how your business actually works. The second is the honest number, and it is the one that predicts whether the thing ships.
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.
Expect this to surface disagreements. Two people doing the same job frequently describe different rules, and the project cannot proceed until someone with authority decides which is correct. That is a management cost, not a technical one, and it lands early.
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.
It is also expensive time specifically. The people who hold the undocumented rules are your most experienced staff, so the work competes directly with whatever else they were doing — which is why it slips, and why projects stall in a state where everything technical is finished.
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.
5. The integration surface
Reading from your systems is usually straightforward; writing to them is where estimates move. Each system needs its own account, scoped permissions and an audit trail — the plumbing that makes actions attributable rather than a shared login.
The cost driver is how many systems are involved and who controls them, which makes it the same question that dominates any software estimate: every system you forget to mention is a change request.
What makes it cheaper
Three things reliably reduce the number. A workflow that is already documented. One channel rather than three. And a genuinely narrow scope, because exclusions remove not just build work but the rules, tests and escalation paths attached to each excluded case.
What does not reduce it is choosing a cheaper model or a simpler tool, since almost none of the cost is there.
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, and it is worth naming as a deliverable in its own right rather than treating as a side effect.