How Do You Scope a Fixed-Price AI Consulting Project?

A fixed-price AI consulting project works when the outcome, boundaries, assumptions and acceptance tests are explicit before work begins.

The answer: sell a defined outcome, not unlimited effort

To scope a fixed-price AI consulting project, define the business problem, the specific deliverables, what is excluded, what the client must provide, how completion will be accepted, and how changes will be priced. A fixed fee should purchase a bounded piece of work, not an unlimited promise to “make AI work.”

That distinction matters for new consultants. A seemingly small workflow can turn into a systems integration, data cleanup project, security review and staff training program. The fixed price does not become sustainable just because the proposal sounded simple. For the choice between pricing models, start with our hourly versus project pricing guide. This article focuses on constructing the actual scope once you choose a fixed fee.

Start with a discovery note before writing the SOW

Ask the client what is happening today, who owns the workflow, how often it runs, which systems are involved, what information moves between them, and what a successful result would look like. Observe a real example instead of accepting a broad description. “Automate our customer enquiries” is not a scope. “Draft a suggested category and response for incoming non-urgent enquiries, with an employee approving every message” is closer.

Collect five pieces of evidence: a sample input, a sample correct output, the current process steps, a baseline measure and a named decision maker. If these are unavailable, consider charging for a separate assessment before quoting implementation. Our discovery call guide offers useful questions.

Worked example: a fictional $3,000 engagement

Entirely fictional example; the price is illustrative, not a recommended market rate. Cedar Street Repairs, a six-person service business, receives about 60 email enquiries each week. The owner wants a system to suggest a category and a draft response. The consultant proposes a supervised workflow, not autonomous replies. Staff will approve every outgoing message. No customer records are used in an unapproved external tool.

The agreed fixed fee is $3,000 USD for a two-week pilot with one inbox, one approved tool environment, four enquiry categories, up to 40 test examples supplied by the client, one staff training session and a handover guide. The client provides access, an authorized owner and feedback within two business days. This example is intentionally small so the work can be estimated and tested.

The project does not include CRM integration, historical data migration, a new customer portal, 24/7 monitoring, production autonomy, or custom model training. These exclusions are as important as the deliverables.

Copyable sample statement of work

This is educational sample wording, not a ready-to-sign legal contract. A real engagement needs appropriate confidentiality, intellectual property, privacy, liability, payment and dispute terms reviewed for the relevant jurisdiction.

Acceptance criteria that can actually be tested

Avoid “client is satisfied” as the only acceptance test. Satisfaction matters, but it is too subjective to define whether the project is complete. For this fictional pilot, the acceptance checklist might require that the workflow is demonstrated on the agreed 40 cases; every suggested response is clearly labelled as a draft; staff can edit or reject it; the system logs the outcome in the agreed location; and the owner can disable the workflow using documented steps.

The parties might also agree that at least 36 of 40 cases must receive the correct category in the test set, while any unsafe suggestion triggers investigation before acceptance. That 90% example threshold is invented for this scenario, not an industry benchmark or evidence that the workflow is production-ready. For consequential tasks, the acceptable threshold may need to be much stricter or the task may not be appropriate for automation at all.

Define who judges each test, how defects are classified, and what happens if the tool cannot meet the agreed criteria. Our project success measurement guide explains how to distinguish output delivery from business results.

Milestones and payments: a worked schedule

For the $3,000 example, a possible schedule is 40% ($1,200) on signed scope and kickoff, 40% ($1,200) when the supervised workflow and test results are ready for review, and 20% ($600) after agreed acceptance and handover. These are illustrative commercial terms; actual payment timing, deposits, taxes and legal requirements vary.

Each payment should correspond to a defined milestone. “Halfway through” is vague; “configured workflow demonstrated with agreed test inputs” is inspectable. Specify what happens when the client misses a dependency, and do not silently absorb unlimited waiting time or extra revisions.

A simple change-request process

Imagine the client asks, on day six, to connect the workflow to a CRM and send responses automatically. Those requests change both the work and the risk profile. The consultant should not casually include them because the client says they are “small tweaks.”

Use a written change request: Requested change: add CRM write-back and unattended sending. Impact: new integration, permissions review, security testing, additional acceptance criteria and likely a revised timeline. Decision: not included in the original $3,000 scope; quote separately after technical and risk assessment. Authorization: work starts only after both parties approve the amendment.

Even when you choose to include a minor change at no charge, document it. That prevents ambiguity later and protects the relationship. See what an AI consultant should deliver for more about handover obligations.

How scope creep can erase the margin

Suppose the consultant planned 20 hours of work. At $3,000, the gross revenue per planned hour is $150 before expenses, taxes, sales effort and non-billable time. If unpriced CRM work, repeated revisions and extra meetings push the project to 35 hours, gross revenue per actual project hour falls to about $85.71. The fee has not changed; the economics have.

Those numbers are arithmetic from the fictional scenario, not a market-rate recommendation. The lesson is to estimate the full workload: discovery, configuration, testing, documentation, meetings, training, corrections and handover. Add a realistic contingency and be clear about how many review cycles are included.

When a fixed fee is the wrong choice

Do not force fixed pricing when the inputs are unknown, the client cannot define the outcome, the systems are inaccessible, or the main task is exploratory research. A short paid discovery phase, a time-boxed assessment or a capped pilot may be safer for both parties. The client should understand what uncertainty remains and what the first engagement will resolve.

Also avoid fixed-fee commitments to guaranteed AI accuracy, guaranteed revenue or unlimited support. An ethical proposal explains assumptions and dependencies. The NIST AI RMF is a helpful reference for identifying risk that can affect scope, testing and governance.

Pre-signature scoping checklist and sources

• Write the business outcome in one sentence.
• Name the workflow owner and approver.
• Specify exactly which systems and data are in scope.
• List every deliverable and its format.
• List exclusions and the number of revision cycles.
• Define test cases, acceptance criteria and sign-off timing.
• Set client dependencies and delay rules.
• Agree milestone payments and change-control steps.
• Define confidentiality, access, ownership and handover responsibilities.
• Make ongoing support a separate decision, not an assumed free extra.

For small-business planning and contracting context, consult the U.S. Small Business Administration business guide. For AI-specific governance considerations, see NIST AI RMF. These sources support the general business and risk-management principles; the SOW and pricing illustration are original examples.

Once a pilot is complete, the next commercial question is whether the client needs an ongoing AI consulting retainer or simply a clean handover.