A statement of work is supposed to protect the agency and align the client. In practice, most AI agency SOWs do neither.
They are either too vague to enforce, too rigid to accommodate the inherent uncertainty of AI work, or missing critical sections that would prevent the disputes that eat margins and damage relationships.
The cost of a bad SOW is not just legal exposure. It is the operational chaos that unfolds when the engagement has no clear boundaries and both sides are working from different assumptions.
Mistake 1: Vague Scope Definitions
The most common and most expensive SOW mistake is describing scope in terms that are open to interpretation.
Problematic language:
- "Develop an AI solution for customer service"
- "Implement machine learning capabilities"
- "Build an intelligent automation system"
These descriptions mean different things to different people. The agency interprets them as a specific deliverable. The client interprets them as a comprehensive outcome. The gap between those interpretations is where disputes live.
Better approach:
Define scope with specificity:
- list the exact workflows, processes, or systems being addressed
- define the inputs the system will receive and the outputs it will produce
- specify what is included and what is explicitly excluded
- identify the boundaries of the AI system's decision-making authority
- name the specific models, platforms, or tools that will be used
"Build an AI-powered invoice classification system that categorizes incoming invoices into five predefined categories (utilities, supplies, services, equipment, other) with a target accuracy of 90% on the provided test dataset" is a scope that can be delivered, tested, and closed.
Mistake 2: No Performance Criteria
AI systems do not have binary pass/fail outcomes like traditional software. They perform on a spectrum. Without defined performance criteria, the question "Is this done?" has no objective answer.
What to include:
- accuracy, precision, recall, or other relevant metrics with specific thresholds
- acceptable response time or latency
- volume and throughput requirements
- conditions under which performance will be measured (test data, production data, specific scenarios)
- the evaluation methodology
What to avoid:
- "The system should perform well"
- "Results should be satisfactory to the client"
- "Industry-standard accuracy"
Subjective criteria create situations where the agency believes the work is complete and the client believes it is not. Both are right according to their own interpretation. The SOW should make interpretation unnecessary.
Mistake 3: Ignoring Data Responsibilities
AI projects depend on data. SOWs that do not clearly assign data responsibilities create delays, blame, and rework.
Define who is responsible for:
- providing the training or input data
- ensuring data quality and completeness
- cleaning and preprocessing data
- labeling or annotating data if required
- providing ongoing data access for monitoring and maintenance
Define what happens when:
- data quality is insufficient for the proposed solution
- required data is not available on schedule
- data formats or structures change during the engagement
- additional data is needed beyond what was originally identified
Data delays are the most common cause of AI project timeline slippage. The SOW should make it clear that data delays from the client side affect the project timeline and are not the agency's liability.
Mistake 4: Fixed Price Without Fixed Scope
Fixed-price contracts work when the scope is well-defined and the risks are understood. In AI projects, scope and risk are often uncertain at the start.
Signing a fixed-price contract for a discovery-level engagement creates a situation where the agency absorbs all uncertainty. Every unexpected challenge reduces margin.
Better approaches:
- Phased pricing: Break the engagement into phases with separate pricing for each. Discovery and scoping are priced separately from implementation.
- Time and materials with a cap: Charge for actual work with a maximum budget. This shares risk between the agency and client.
- Fixed price with change order provisions: Set a fixed price for the defined scope with a clear process for pricing additional work.
The key is matching the pricing model to the level of certainty about the scope.
Mistake 5: No Change Order Process
AI projects evolve. Requirements change as the team learns more about the data, the model's capabilities, and the client's actual needs.
Without a change order process, changes happen informally. The agency absorbs extra work to maintain the relationship. Resentment builds. Margins erode.
Include in the SOW:
- how change requests are submitted
- who evaluates and approves changes
- how changes affect timeline and budget
- required documentation for approved changes
- what happens if the client requests changes that conflict with existing scope
A change order process is not about being difficult. It is about making sure both sides agree on what is changing before the work is done.
Mistake 6: Missing Acceptance Criteria
Without defined acceptance criteria, project completion becomes a negotiation instead of a verification.
Define acceptance criteria for each deliverable:
- specific tests or evaluations that determine whether the deliverable meets requirements
- who performs the acceptance testing
- the timeline for acceptance review
- what happens if the deliverable does not meet criteria (rework scope and timeline)
- how many rounds of revision are included
Include a acceptance timeline:
- how long the client has to review and accept deliverables
- what happens if the client does not respond within the review period (auto-acceptance is common)
- how disputes about acceptance are resolved
Mistake 7: No Intellectual Property Clarity
AI projects create multiple types of intellectual property: code, models, training data, prompts, workflows, and documentation.
The SOW should clearly define:
- who owns the final deliverables
- who owns underlying tools, frameworks, or libraries the agency used
- whether the agency can reuse approaches, patterns, or anonymized learnings
- how pre-existing IP from either side is treated
- who owns models trained on the client's data
- licensing terms for any third-party components
Ambiguity about IP creates conflicts that are expensive to resolve after the work is done.
Mistake 8: No Warranty or Support Boundary
What happens after the project is delivered? If the SOW does not say, the client assumes ongoing support is included.
Define:
- the warranty period and what it covers
- the boundary between warranty issues and new work
- how bugs are distinguished from feature requests
- response time expectations during the warranty period
- what ongoing support options are available after warranty expires
This section transitions naturally into a maintenance or retainer conversation, which benefits the agency's recurring revenue strategy.
Mistake 9: Missing Assumptions and Dependencies
Every AI project is built on assumptions. Making those assumptions explicit protects the agency when reality does not match expectations.
Common assumptions to document:
- client will provide data access by a specific date
- client will assign a project owner with decision-making authority
- client infrastructure meets specified requirements
- third-party APIs or services will remain available at current pricing
- the volume or complexity of data will remain within estimated ranges
When an assumption proves false, the documented assumption gives the agency standing to discuss timeline or budget adjustments.
Building Better SOWs
The fix for all of these mistakes is the same discipline: specificity.
A well-written SOW takes more time upfront. It requires thorough discovery, honest risk assessment, and sometimes difficult conversations about what is and is not included.
That investment pays for itself many times over. It prevents disputes, protects margins, and creates the kind of client relationship where both sides know exactly what to expect.
Agencies that write clear SOWs do not just avoid problems. They build reputations for professionalism that win more business and better clients.