AI automation consulting pricing: scope the cost you can verify

Ask for a scoped estimate

“How much for an AI agent?” is missing the part that sets the price: what will it do, which systems will it touch, and who handles mistakes? A tool that drafts a reply is a different job from one that updates a live CRM. A useful estimate starts with that scope.

This guide does not present an industry rate survey or a price quote for Fazal Shah’s services. The numbers below are explicitly hypothetical planning examples. A real estimate follows review of the scope and acceptance criteria.

Calculate operating cost from assumptions

Suppose a workflow completes 1,000 tasks per month. Assume model usage costs $0.02 per attempt, with an average 1.2 attempts per task; tools cost $0.01 per completed task; hosting is $30 monthly. The estimated monthly machine cost is $24 + $10 + $30 = $64. That is $0.064 per completed task before human review, maintenance and taxes.

If 100 tasks require three minutes of review each, that adds five hours of operator time. Apply your actual internal rate and include the cost of failed tasks and maintenance. The machine budget alone is not the project budget.

Compare proposals on deliverables

A fixed-price milestone can be appropriate when the deliverable is bounded. A discovery phase may be more suitable when the workflow or constraints are unclear. These are scoping options, not a promise of a specific contract structure.

Evaluate value against the same quality bar

Record the manual baseline: time spent, correction rate and who handles exceptions. Estimate recoverable time conservatively and keep quality comparable. A workflow that saves drafting time but creates substantial review work may have a smaller benefit than expected.

Do not turn a hypothetical time-saving assumption into a case-study result. After deployment, measure actual completion rates and operator time over a stated period, then revisit the estimate.

What to bring to a scoping call

Prepare a process outline, sample inputs with sensitive details removed, a list of connected systems and the desired output. State which actions require approval and what a failed run should do. Avoid sending credentials through the public form.

Use the agent cost calculator, read the workflow selection guide, or discuss an automation project.

Make acceptance part of the commercial scope

For a draft-response workflow, an acceptance test can require the correct ticket, supported facts, an editable draft and no automatic send. For a record-updating workflow, include record ownership, approved fields and a safe response to timeouts. These are materially different deliverables even if both are described as an AI assistant.

Ask how changes after acceptance will be handled. A new connected application, higher request volume or expanded write permission can alter both implementation effort and operating risk. Keep those changes visible rather than treating them as an unlimited extension of the original brief.

A useful proposal should also describe the handover: where configuration lives, who receives alerts, how to pause the workflow and how a partial action is reconciled. These details allow the business to assess maintainability alongside initial delivery. They are not a promise that every engagement includes the same support period; the agreement should say which responsibilities are included and which remain with the client.

Method and review notes

All numerical examples are original calculations using the assumptions printed beside them. No market-wide consulting rates, client ROI figures or provider price claims are used. Review edition prepared October 14, 2026.