Introduction
An AI automation statement of work (SOW) closes the gap between a clear proposal and a messy delivery. The SOW records what an AI build delivers and how the client accepts the result.
Data Prism builds RAG assistants, document processing, and workflow automation. Before launch, Data Prism validates each workflow against real business scenarios, test data, and expected exception cases. This article covers an annotated template, a responsibility matrix, and a review checklist for one invoice workflow.
What an AI Automation Statement of Work Covers
An AI automation statement of work covers the agreed delivery scope for one AI build. The SOW names each deliverable and its acceptance check. Every task also gets an owner and a date. Anything the SOW leaves out is not owed.
The US Census Bureau’s business trends survey reported a national AI use rate among businesses of 19.8% as of May 3, 2026. More AI builds mean more buyers reviewing SOWs for probabilistic output. This guide is not legal advice, so have contract terms reviewed by a qualified professional.
SOW vs Proposal vs Service Agreement
A proposal persuades and prices, while an SOW fixes delivery scope. A service agreement sets the commercial and legal terms the SOW sits under. An SOW is not the same as a proposal, since the proposal feeds the SOW but never becomes the delivery record.
|
Document |
Main Job |
Who Drafts |
What the Document Fixes |
When Signed |
|---|---|---|---|---|
|
Proposal |
Pitches approach and price |
Provider |
Proposed approach and price, usually nonbinding |
Accepted before the SOW |
|
Statement of Work (SOW) |
Records delivery scope |
Provider, with client review |
Deliverables, acceptance criteria, roles, and dates |
Per project, under the agreement |
|
Service Agreement |
Sets legal and commercial terms |
Either party, often with legal counsel |
Payment, liability, IP rights, and termination |
Once, before the first SOW |
When the proposal and the SOW disagree, the signed SOW governs delivery.

Define Objectives and Scope Boundaries
A signed SOW governs delivery once objectives and scope boundaries are clear. A 2024 RAND report cites estimates of more than 80% failure for AI projects. RAND’s interviews name a misunderstood or miscommunicated problem as the most common root cause. An AI automation scope of work fixes the misunderstanding in the objective.
Business Objective and Workflow
The business objective names one workflow and the team owning the workflow. The objective also records the current manual steps and the target outcome. Each objective ties to a baseline, such as handling time or error rate. The client supplies the baseline, because a provider-set baseline is easy to beat.
Illustrative clause: The solution will extract supplier, invoice number, date, and total from inbound PDF invoices and create a draft entry in the client’s accounting system for review by the finance team.
Workflow mapping precedes drafting and shows which manual steps the build replaces.
Included and Excluded Tasks
The in-scope list names each task, system, and document type the build covers, plus expected volumes. Disputes start where the SOW is silent. For the invoice workflow, the out-of-scope list names five exclusions.
- Other document types
- Other languages
- Historical data cleanup
- Hardware
- End-user devices
Written exclusions stop a casual offer on a call from becoming a commitment.
Illustrative exclusion: Handwritten invoices and scanned images below the agreed quality threshold are excluded from automated extraction and will route to manual review.
Assumptions and Dependencies
Assumptions are conditions the plan relies on, and dependencies are inputs the client supplies. The invoice workflow assumes working APIs and representative sample data. The plan also counts on a client contact who replies on time. The dependencies are four client inputs.
- System access
- Test accounts
- Data samples
- Security approvals
Even so, undocumented steps and messy records often surface during delivery. A data quality check on sample data before scheduling catches many problems early. A failed assumption then triggers a timeline review or a change request.
Specify Automation Project Deliverables and Responsibilities
Automation project deliverables are nouns, each with one line describing done. A vague deliverable such as “an AI solution” leaves nothing to check at handover. The table lists six invoice workflow deliverables.
|
Deliverable |
Done Looks Like |
|---|---|
|
Design Document |
Client approves fields, routing rules, and exclusions in writing |
|
Configured Workflow |
Email intake, extraction, and routing run in the client’s accounts |
|
Integration with the Accounting System |
Draft entries appear in the accounting system for finance review |
|
Test Evidence Pack |
Per-field scores on the client’s test set, every failure listed |
|
Training Session |
Accounts payable clerk clears the exception queue without help |
|
Handover Documentation |
Versioned configuration, runbook, and account owners on file |
The integration row needs an AI integration clause naming the target system and fields written. The clause also names the failed-write route.
Responsibility Matrix
A responsibility matrix gives every task one owner, since unowned tasks get missed or disputed. Buyers often forget the client’s share, such as data, access, output reviews, and sign-off. The matrix below splits the tasks behind the automation project deliverables between provider and client.
|
Task |
Provider |
Client |
Notes |
|---|---|---|---|
|
Process Walkthrough and Current-State Map |
Owns |
Supports |
Finance team explains today’s invoice steps |
|
Sample Documents and Test Set Selection |
Supports |
Owns |
Client picks the invoices, including poor scans |
|
System Access and Test Accounts |
Supports |
Owns |
Mailbox and accounting access granted before the build starts |
|
Workflow Design |
Owns |
Approves |
Fields, routing, and exclusions fixed in the design document |
|
Build and Configuration |
Owns |
None |
Built in accounts the client owns |
|
Review of Extraction Output |
Supports |
Owns |
Finance team checks sample draft entries during the build |
|
Acceptance Testing and Evidence |
Owns |
Approves |
Provider runs the test set; client checks the evidence |
|
Sign-Off |
None |
Owns |
Named client approver signs against the acceptance criteria |
|
Training and Handover |
Owns |
Supports |
Accounts payable clerk attends; client team receives the documents |
The matrix belongs in the signed SOW. A slipped client task then shows whose delay moved the date.
Write Measurable Acceptance Criteria for AI Output
Acceptance criteria in an AI automation statement of work score output on a fixed sample. AI output is probabilistic, so promising perfect accuracy starts a dispute at the first error. The criteria also route exceptions and name who signs off.
Output Quality on an Agreed Test Set
A test set is the fixed sample behind every acceptance result. The SOW names the set’s source, size, and selector. The set is fixed before testing begins. Scores run per field, because one blended number hides a weak field.
Illustrative criterion: At least 95% of invoices in the agreed test set of 200 documents have invoice number, date, and total extracted correctly.

Under OMB memo M-25-22, dated April 3, 2025, federal agencies evaluate AI on test data they define. The federal guidance says the data “should not be accessible” to the provider. The memo binds only federal agencies, but any buyer can hold back part of the test set.
Exceptions and Approval Behavior
Every AI system receives some inputs too unclear to process with confidence. Exception handling sends each one to a named person with a reason. In the invoice workflow, an unclear scan goes to the accounts payable clerk and never passes through silently. Exceptions are expected, so a queued scan isn’t a defect by itself.
Approval steps need a named person too, plus a response time. The SOW also names a backup approver. Even AI agents need a person on the approval step.
Performance and Sign-Off
Accuracy alone doesn’t show a system is ready for daily use. Performance terms add processing time per item and concurrent load. Availability during acceptance counts too. Setting figures before testing prevents acceptance debates.
The provider presents evidence, and the client decides. Sign-off names a client role, such as the finance operations manager, and a review period. A deemed acceptance clause accepts the deliverable once the period ends without a written rejection.
Record Milestones and Change Control
Delivery disputes often start with an unwritten date or request. Milestones give each stage a review date, and change control sets how requests get priced and approved. The SOW records both, plus the rule for a slipped schedule.
The six milestones are kickoff, design approval, build complete, test complete, go-live, and handover. Each milestone carries a review date, and payment follows acceptance. If client access or sample invoices arrive late, the SOW states how far each date shifts and who gets told.

Each change request gets written down and numbered the same day, even a verbal “can you also.” Each request then carries an estimate of effort, cost, and timeline. No work starts before the client approves the estimate. The log below shows one illustrative change request.
|
CR Number |
Description and Reason |
Effort |
Cost |
Timeline Impact |
Approver and Date |
|---|---|---|---|---|---|
|
CR-001 |
Add credit notes as a document type for supplier refunds |
Estimated build and test days |
Estimated fee at the agreed rate |
Estimated shift to the go-live date |
Client finance lead, with sign-off date |
Clarify Ownership and Post-Launch Support
Many SOWs give the client the deliverables, but an AI build runs on more than code. Prompts, settings, and test data keep the system working. The SOW should assign intellectual property and access for each item below.
- Custom code
- Configuration and workflow definitions
- Prompts and retrieval settings
- The test set and test results
- Cloud and platform accounts
- Documentation
- Third-party licenses
- API and platform usage costs

One prompt edit can change extracted invoice totals. Without the test set, the client can’t tell whether the edit helped. A RAG assistant also depends on retrieval settings, so the handover includes them.
The SOW also names the account behind each usage bill. LLM APIs bill per token, while workflow platforms bill per task, credit, or execution. Ownership stops at third-party components, which keep their own license terms.
After handover, support and maintenance go in a separate support schedule.
AI Automation SOW Template and Review Checklist
A blank AI automation SOW template helps only when each field forces a specific answer. Annotated fields come first, then a filled example. Fill the fields for one workflow, then check the draft before anyone signs.

Annotated SOW Template
Each field in the statement of work template answers a question a buyer would otherwise ask mid-delivery. The note beside each field says what belongs there. A blank field leaves the question for a dispute to settle later.
|
Field |
What to Write |
|---|---|
|
Objective and Baseline |
The business result, plus today’s measured starting point |
|
Workflow and Owner |
One named workflow and the accountable client person |
|
In Scope |
Inputs, steps, and outputs the solution handles |
|
Out of Scope |
Inputs and tasks the provider will not handle |
|
Assumptions |
Conditions the plan depends on, such as data access dates |
|
Dependencies |
Client inputs and approvals, each with a due date |
|
Deliverables |
Nouns the client receives, such as code, prompts, and runbooks |
|
Responsibility Matrix |
Who builds, reviews, approves, and supports each part |
|
Acceptance Tests |
The test set, the pass threshold, and who runs the tests |
|
Exception Handling |
Where failed or low-confidence items go, and who resolves them |
|
Sign-Off Role and Review Window |
The approving job title and the days allowed to respond |
|
Milestones and Payment |
Each payment tied to a milestone’s pass or fail check |
|
Change Requests |
Written, numbered, estimated, and approved before work starts |
|
Ownership and Running Costs |
Who owns code and prompts, and who pays usage bills |
Completed Example
The statement of work example below fills the template for the invoice workflow. Every number is illustrative and sets no benchmark. A real SOW sets each figure from the client’s own invoices.
|
Field |
Illustrative Entry |
|---|---|
|
Objective |
Draft accounting entries from inbound PDF invoices for finance review |
|
In Scope |
Supplier, invoice number, date, and total from emailed PDF invoices |
|
Out of Scope |
Handwritten invoices and scans below the agreed quality threshold |
|
Test Set |
An illustrative 200 client-chosen invoices, fixed before testing |
|
Threshold |
An illustrative 95% of invoices with all three fields correct |
|
Exceptions |
Low-confidence invoices route to a named accounts payable clerk with a reason |
|
Sign-Off Role |
Finance operations manager |
|
Review Window |
An illustrative 5 business days after test results arrive |
|
Ownership |
Client owns code, prompts, configuration, and usage accounts |
Review Checklist
The checklist turns the template into eight yes-or-no checks. Each no answer flags a gap to close before signing. Afterward, only the written text counts.
- Does the objective state a measured baseline?
- Is one workflow tied to one client owner?
- Does the out-of-scope list name inputs and tasks?
- Does the client pick a fixed test set?
- Is the pass threshold a number tied to named fields?
- Does every exception route to a named role?
- Are the sign-off role and review window named?
- Are ownership and running costs assigned?
Phrases to Challenge
Some proposal phrases still fail the checklist. The provider reads each phrase narrowly, and the client reads broadly. Each row below gives a checkable term to ask for.
|
Phrase in the SOW |
Why the Phrase Fails |
Ask for Instead |
|---|---|---|
|
“High Accuracy” |
No number, no test set, and no pass rule |
A per-field threshold on a fixed test set |
|
“As Needed” |
Neither side agreed on when or how much |
A named trigger, a frequency, and an hour cap |
|
“Full Ownership of Deliverables” |
Third-party components keep their own license terms |
A list of owned items and licensed items |
|
“Ongoing Support Included” |
No end date, response time, or covered scope |
A support period, response times, and covered issues |
|
“Integration With Existing Systems” |
Systems, data direction, and access stay unnamed |
Each system, data direction, and access owner |
|
“Minor Changes Included” |
The size of a minor change goes undefined |
A change request process with an effort cap |
Final Checks Before Signing an AI Automation SOW
The checks above help only before signing. Afterward, each gap in the AI automation statement of work becomes a negotiation. So, review the scope of your AI automation engagement against the template before anyone signs.
Book a Free 30-Minute Meeting
Discover how our services can support your goals — no strings attached. Schedule your free 30-minute consultation today and let's explore the possibilities.
Book a Free Call

