What AI Consulting Actually Costs — And What Drives the Number

The four pricing models

Almost every engagement uses one of four structures, and each fails in a different way.

• Hourly — flexible and honest for exploratory work, but it prices your consultant's time rather than your outcome, and nobody is incentivised to be quick.
• Fixed scope — a defined deliverable for a defined fee. Best for a consultation or a single workflow, because both can be specified in advance. Falls apart when the scope was never really understood.
• Retainer — ongoing capacity, sensible once systems are live and need maintenance and iteration. Wasteful before that.
• Outcome-based — payment tied to a measured result. Attractive in principle, rare in practice, because it requires a baseline both sides trust.

What actually moves the number

Cost tracks complexity of integration far more than sophistication of AI. A conceptually simple workflow touching four systems with inconsistent data will cost more than a clever one touching a single clean API.

• How many systems the work touches, and whether they have usable APIs.
• Data quality — cleaning and deduplicating records is frequently the largest line item.
• How well the process is documented. Undocumented processes have to be mapped first, and that is billable discovery.
• Number of edge cases the system must handle rather than escalate.
• Compliance requirements, which add review cycles more than engineering.

We scope and price this work in an AI consultation before any build begins, so the number is attached to a specific plan.

How to read a quote

A quote you can evaluate has specifics in it. A quote you cannot is usually hiding the fact that nobody has scoped the work yet.

• Named deliverables, not capability descriptions.
• What is explicitly out of scope — its absence is the most common source of overrun.
• Who owns the finished system, the accounts, and the credentials.
• Ongoing costs stated separately: model usage, platform subscriptions, maintenance.
• What happens when something breaks after handover, and who pays for it.

Cheap and expensive both have failure modes

The cheapest quote usually excludes error handling, monitoring, and documentation. Those are not extras — a workflow that fails silently and is not noticed for three weeks is worse than no workflow, because people stop trusting the system and quietly revert to doing it by hand.

The most expensive quote is often paying for a discovery phase that produces a strategy deck rather than a running system. Ask what exists at the end of each phase. If the answer to phase one is a document, ask what phase two costs before you commit.

A sane way to budget

Budget the first project against the cost of the process it replaces, not against a technology budget. If a process consumes a few hours a week, the annual cost of that time is the number the build has to beat — and back-loaded economics mean a one-off build against a recurring saving usually clears it quickly.

Then fund the second project from the proven result of the first. That sequence keeps the programme honest and makes it far easier to get approval for the larger work later.