AI statement of work: what to include, with a template.
What an AI consulting statement of work must say that a normal software SOW doesn’t: measured acceptance criteria, data terms, model changes, ownership and exit. With a complete template to copy.
An AI statement of work needs everything a software SOW has (scope, deliverables, milestones, ownership) plus what a normal one leaves out. It must define quality as a measured result on an agreed test set, because an AI system’s output varies and “works as specified” means nothing without numbers. It must say what data you provide and who may use it for what. It must name the models and services the system depends on, who holds those accounts, and what happens when a provider changes a model. And it must say who owns the prompts, evaluation sets and any fine-tuned weights, not only the code.
For an enterprise generative AI or automation project, add human handover rules, security testing that covers prompt injection, monitoring after launch and an exit plan. This guide explains each section from the buyer’s side and ends with a template to copy.
Why does AI work need a different statement of work?
A normal SOW assumes that if the supplier builds what is described, the result works. For AI, that assumption breaks in four places.
- The output is probabilistic. A system that is right 95 times in 100 is still wrong five times. NIST’s Generative AI Profile calls one form of this confabulation: “the production of confidently stated but erroneous or false content”. The SOW has to say how often wrong is acceptable, and which kinds of wrong never are.
- Quality is measured, not inspected. You accept an AI system by scoring it on agreed cases, not by clicking through screens.
- The result depends on your data. Out-of-date documents and incomplete records become wrong answers. Nobody can promise a result before someone has looked at the data.
- The model underneath changes. Providers update and retire models on their own schedule. NIST’s AI Risk Management Framework (AI RMF 1.0) notes that AI systems may need more frequent maintenance “due to data, model, or concept drift”.
Public buyers reached the same conclusions. The UK government’s Guidelines for AI procurement (June 2020) say “lifespan testing, not a one-time decision, is required”, and tell buyers not to expect “a fully detailed and definite plan, as AI development is an iterative process”. In the US, OMB memo M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government (3 April 2025), which replaced the earlier memo M-24-18, tells agencies to test proposed solutions before award and to include contract terms that let them monitor performance regularly. The reasoning holds for any company buying AI work.
If acceptance isn’t a number, it’s an opinion.
What sections does an AI statement of work need?
Each section settles a question you would otherwise argue about later. The last column is the red flag when a draft leaves it out.
| Section | The question it answers | Red flag if missing |
|---|---|---|
| Objective and business measure | What problem does this solve, and how will we know it helped? | The goal is “implement an AI assistant”, with no number attached. |
| Scope and out of scope | What exactly is being built, and what isn’t? | No out-of-scope list, so every new idea becomes an argument. |
| Data | What data do we provide, who may use it, and for what? | Silence on whether the supplier or its model provider may train on your data. |
| Evaluation set and acceptance | What test must the system pass, and who owns the test? | Acceptance is a demo, or “to the client’s reasonable satisfaction”. |
| Human oversight and handover | When does a person decide, and how does the system hand over? | No route to a person. |
| Models and third-party services | What does it depend on, and whose accounts are they? | Model keys sit in the supplier’s account with no plan to transfer them. |
| Security | How is it tested against misuse, including prompt injection? | Security means only “industry-standard encryption”. |
| Ownership | Who owns the code, prompts, evaluation sets and weights? | Ownership covers “software” without naming prompts and test sets. |
| Milestones and change control | What must be true at each step, and what if the scores fall short? | Milestones are dates, not results, and there is no decision point. |
| Monitoring after launch | Who watches quality after go-live, and how often? | Support ends at launch, or covers outages only. |
| Warranty | What exactly does the supplier promise about behaviour? | “Accurate” or “error-free” with no test behind it. |
| Exit and handover | Can we run it, or move it, without this supplier? | No documentation, knowledge transfer or data return at the end. |
How do you write the objective and the scope?
Start with the problem and a number. “Answer supplier invoice queries the same day instead of within two days” is an objective. “Build a RAG chatbot” is a method. Name the business measure, today’s baseline and the target.
The World Economic Forum’s AI Procurement in a Box: AI Government Procurement Guidelines (June 2020) puts it plainly: “Focus on developing a clear problem statement, rather than on detailing specifications of a solution.” Fix the outcome and the acceptance test tightly, and leave the model and technique open unless you have a reason to fix them.
The scope lists the workflows, users, channels and languages, the systems the AI reads and writes, and the actions it may take. Then write the out-of-scope list: the obvious next steps you are not buying yet, such as a second channel or write access to the CRM.
What should the SOW say about data?
Many AI projects slip on data. Be specific.
- What you provide, and when. Sources, format and the date each is ready, and what happens to the timeline if data arrives late.
- Access. How the supplier reaches the data, where it is stored, and who on the supplier’s side may see it.
- Quality. Who checks the data before the build. The UK guidelines have a heading that reads “Conduct a data assessment before starting your procurement process”. If you haven’t done one, make it the first deliverable.
- Use and training. Say whether the supplier, or any model provider it uses, may train on your inputs or outputs. M-25-22 says US agencies’ processes should ensure contracts “permanently prohibit” using non-public agency data and outputs to further train publicly or commercially available AI, absent explicit agency consent. You can ask for the same.
- Personal data. What the system handles, where it is processed, how long it is kept and how it is deleted. The WEF guidelines suggest a date for deletion and written confirmation from the vendor that the data is gone.
- Data the system produces. Logs, conversation records and scores, which stay yours like the data you supply. A WEF principle is to “Ensure that you have the required access to data used and produced by the vendor(s) solution”, because you can’t check performance without it.
How do you set acceptance criteria for an AI system?
This is the section that makes an AI SOW work. The system runs against an agreed evaluation set, and it passes if its scores meet numbers written in the SOW.
Build the evaluation set
An evaluation set is a list of realistic inputs, each with the correct outcome: the right answer, the right action, a refusal or a handover. Build it from real cases, including the awkward ones, and have someone who does the work today sign off the expected outcomes. Coverage matters more than size. Evals before features explains how to build and score one.
Keep part of it back. M-25-22 tells US agencies to evaluate on data they define, and says that data “should not be accessible to the vendor, and should be as similar as possible to the data used when the system is deployed”. A held-back portion stops a system being tuned to the exact questions it will be graded on.
Write the numbers
| Measure | How to write it in the SOW |
|---|---|
| Task pass rate | At least [90]% of held-back cases meet the expected outcome. |
| Critical failures | Zero in the critical category, such as a wrong eligibility decision, personal data disclosed or an action taken without approval. |
| Supported answers | At least [95]% of factual answers cite a source passage that supports them. |
| Abstention | When the answer isn’t in the data, it says so or hands over in at least [95]% of cases. |
| Handover | Every case marked “must hand over” is handed over, with a summary. |
| Latency | [95]% of responses within [x] seconds. |
| Security | Zero successful attacks from the agreed prompt-injection cases. |
| Per category | Every category scores at least [x]%, so a good average can’t hide a weak area. |
The bracketed numbers are placeholders. A wrong internal search answer and a wrong eligibility decision deserve different bars.
Agree how it is scored
Say who scores each case (a script, a model grader checked against human labels, or a named reviewer), who settles disagreements, and how many runs count, since results vary between runs. M-25-22 says contracts must detail the vendor’s “examination, testing, and validation procedures”. Ask for the same, plus the right to share the results internally. Our guide to AI acceptance testing covers building the test set and scoring it.
Where does a person stay in charge?
List the decisions the system makes alone, the ones it proposes for a person to approve, and the ones it never makes. Then write the handover triggers as rules: the user asks for a person, there is no supported answer, an action is above a limit, the user is upset, or a connected system is down. Require triggers and limits to be enforced in code, not only in the prompt. A rule in code can be tested, so every trigger becomes a case in the evaluation set.
For when the system is down, NIST’s Generative AI Profile (NIST AI 600-1) suggests testing rollover and fallback for generative AI systems, “acknowledging that rollover and fallback may include manual processing”. Name that manual process and who runs it.
Which models does it depend on, and whose accounts are they?
List every third-party dependency (model provider and version, embedding model, vector database, hosting) and whose account each runs in. The safest arrangement for a buyer is to hold production accounts and keys in your own organisation and grant the supplier access you can remove. A system whose keys live in the supplier’s account is a system you can’t run without them.
Then cover change: model versions are pinned and recorded, no model, prompt or provider change reaches production until the held-back set passes again, and the supplier gives notice before switching models. M-25-22 encourages agencies to require vendors “to meet performance standards before deploying a new version of an AI system or service or to roll-back to a previous version if a new version fails to meet performance standards”. Ask for the same.
How should the SOW handle security and prompt injection?
A system that reads outside text (messages, emails, documents, web pages) can be steered by instructions hidden in it. NIST’s Generative AI Profile describes indirect prompt injection, where attackers inject “prompts into data likely to be retrieved”, and notes researchers have shown it can steal proprietary data or run malicious code remotely. The security section should list every entry point for untrusted text, limit tools and permissions to what each task needs, require a person’s approval before irreversible actions, and put attack cases in the evaluation set with a target of zero successes. Prompt injection is an input-validation problem explains why permissions, not prompts, limit the damage.
Who owns the code, prompts, evaluation sets and weights?
Standard clauses talk about “software”. An AI build produces more, and the valuable parts are easy to leave out: system prompts, tool definitions, the evaluation set and its expected outcomes, scoring scripts, retrieval settings, cleaned or labelled data, and fine-tuned weights or adapters. Name each one, say which become yours, and what licence you get to the supplier’s pre-existing tools.
M-25-22 tells agencies to include terms that “clearly delineate the respective ownership and IP rights of the government and the contractor”, especially where agency information is used to train or fine-tune the system. If a model is fine-tuned on a provider’s platform, check whether the weights can be exported at all; if not, owning them means controlling the account they live in.
How should milestones and change control work?
Tie each milestone to a result, not only a date:
- Discovery and feasibility. Data assessment, evaluation set and acceptance numbers agreed, a baseline, and a prototype scored on the development set. It ends with a written go or no-go recommendation.
- Build. Ends when the system passes the held-back set at the agreed numbers in a test environment.
- Pilot. Live for a limited group, with a person reviewing outputs, until the live measures hold for an agreed period.
- Launch and handover. Full rollout, documentation, training and transfer of accounts.
The feasibility gate is the most important line in the SOW. Sometimes the approach won’t reach the bar, and it is far better to learn that early. Decide now what happens then: the supplier reports scores and causes, and you choose to change the approach, the scope or the criteria in writing, or to stop. The criteria and the held-back set change only by a signed change request, never by quietly editing test cases until the number passes.
What should happen after launch?
A system that passed acceptance can get worse without anyone touching it: content changes, users ask new things, and the model is updated or retired. The support terms should name the live measures and who reviews them, rerun the evaluation set on every change and on a schedule, require model migrations to pass it again, define a serious incident, and set a regular review against the objective.
M-25-22 says contract terms must let the agency “regularly monitor and evaluate (e.g., on a quarterly or biannual basis, based on the needs of the program) performance, risks, and effectiveness”. The UK guidelines say: “establish with the supplier how the efficacy of the model will be monitored once deployed”.
What should the warranty and exit terms say?
Don’t accept a warranty that the system is “accurate” or “error-free”. Nobody can honestly promise it, and it gives you nothing to enforce. Ask for a warranty tied to measured behaviour: at acceptance the system met the agreed numbers on the held-back set, and for a set period the supplier fixes any fall below them on the same set, model version and data.
Exit terms decide whether you can leave. At the end the supplier hands over everything in the ownership list plus documentation, removes its access to your accounts, deletes your data with written confirmation, and gives knowledge-transfer time to your team or a new supplier. M-25-22’s closeout section asks agencies to agree the “format and usability of any data” and a plan for transferring data and derived assets.
The AI statement of work template
Copy it, replace every [square bracket], and delete what doesn’t apply. It sits under your master services agreement, which holds the legal terms.
AI STATEMENT OF WORK
Project: [project name]
Customer: [company name] ("Customer")
Supplier: [company name] ("Supplier")
Governed by: [master services agreement, date]
Version: [1.0], dated [date]
1. OBJECTIVE
1.1 Problem: [the problem in one or two sentences]
1.2 Business measure: [measure]. Baseline [value] on [date]. Target [value] by [date].
1.3 This SOW fixes outcomes and acceptance tests. The Supplier chooses models and methods unless section 6 names them.
2. SCOPE
2.1 In scope: [workflows, users, channels, languages]
2.2 Systems the AI reads: [system, access level]
2.3 Systems the AI writes to: [system, actions]
2.4 Out of scope: [e.g. second channel, new language]. Anything not in 2.1 to 2.3 is out of scope.
3. DATA
3.1 Customer provides: [source, format, date ready]
3.2 Access and storage: [method, system, region]
3.3 Supplier staff with access: [roles or names]
3.4 Data gaps and quality issues reported by [date].
3.5 Customer data, inputs and outputs are used only for this SOW. Neither the Supplier nor its model providers may train models on them without the Customer's written consent.
3.6 Personal data: [categories, legal basis, processing locations, retention period]
3.7 The Customer has full access to system logs, records and scores.
3.8 On exit, Customer data is returned or deleted within [x] days, confirmed in writing.
4. EVALUATION SET AND ACCEPTANCE
4.1 Evaluation set: [number] cases from [source], covering [request types, edge cases, handover and attack cases]. Expected outcomes signed off by [name].
4.2 Development set: [x]% of cases, shared with the Supplier.
4.3 Held-back set: [x]% of cases, kept by the Customer and run only at milestone gates.
4.4 Acceptance criteria, on the held-back set:
a. Task pass rate at least [x]%
b. Critical failures: [0] in [categories]
c. At least [x]% of factual answers cite a supporting source
d. Says so or hands over in at least [x]% of cases where the answer is not in the data
e. Required handovers made: [100]%
f. [x]% of responses within [y] seconds
g. Successful attacks: [0]
h. Every category scores at least [x]%
4.5 Scoring: [script, model grader checked against human labels, or named reviewer]. Disputes are settled by [name]. The set runs [3] times; the [lowest] run counts.
4.6 The test method, data and results belong to the Customer.
5. HUMAN OVERSIGHT AND HANDOVER
5.1 Decisions the system makes alone: [list]
5.2 Decisions it proposes for approval: [list]
5.3 Decisions it never makes: [list]
5.4 Handover triggers: user asks for a person; no supported answer; action above [limit]; distress or complaint; system down; [other].
5.5 A handover carries a summary, checks done and anything promised.
5.6 Triggers and limits are enforced in code, not only in prompts. Each one has a test case.
5.7 Manual fallback: [process, owner].
6. MODELS AND THIRD-PARTY SERVICES
6.1 Dependencies and account holder for each: [model and version, embedding model, vector store, hosting]
6.2 Production keys are held by the Customer, which grants and can remove Supplier access.
6.3 Model versions are pinned. No model, prompt or provider change reaches production until the held-back set passes 4.4.
6.4 [x] days' notice before a model or provider change.
6.5 If a new version fails 4.4, the Supplier rolls back to the last version that passed.
7. SECURITY
7.1 Entry points for untrusted text: [list]
7.2 Tools and permissions: [list], each limited to what the task needs.
7.3 Actions that need a person's approval: [list]
7.4 Prompt-injection cases are part of 4.4 g.
7.5 Logs: [location, retention, who can read them]
7.6 Security incidents reported within [x] hours.
8. OWNERSHIP
8.1 The Customer owns: source code, prompts, tool definitions, evaluation sets and expected outcomes, scoring scripts, configuration, cleaned or labelled data, and fine-tuned weights or adapters.
8.2 The Supplier keeps its pre-existing tools: [list]. The Customer gets a [perpetual, transferable] licence to use them in the system.
8.3 Weights that cannot be exported stay in an account the Customer controls.
8.4 Provider of the system under applicable AI law: [party]. Supplier documentation: [list].
9. MILESTONES
M1 Discovery and feasibility, by [date]: data assessment, agreed evaluation set, baseline, prototype scored on the development set, written go or no-go recommendation.
M2 Build, by [date]: passes 4.4 on the held-back set in a test environment.
M3 Pilot, [dates]: live for [group] with review; live measures hold for [x] weeks.
M4 Launch and handover, by [date]: rollout, documentation, training, account transfer.
Each milestone is accepted in writing by [name].
10. CHANGE CONTROL
10.1 Changes to scope, data, the evaluation set or acceptance criteria need a written change request signed by both parties.
10.2 If 4.4 looks unlikely to be met, the Supplier reports scores and causes within [x] days. Options: change the approach, change the scope, change the criteria in writing, or stop.
11. AFTER LAUNCH
11.1 Live measures: [list], reviewed [weekly] by [names].
11.2 The evaluation set reruns on every change and every [month].
11.3 The Supplier tracks model deprecation notices and migrates before retirement. A migration must pass 4.4.
11.4 Serious incident: [definition]. Notice within [x] hours.
11.5 Review against the objective every [quarter]. Retirement criteria: [criteria].
12. WARRANTY
12.1 At acceptance the system met 4.4 on the held-back set.
12.2 For [x] days after acceptance the Supplier fixes any fall below 4.4 on the same set, model version and data.
12.3 Not covered: Customer changes, use outside the scope, [other].
13. EXIT AND HANDOVER
13.1 On expiry or termination the Supplier delivers everything in 8.1, plus architecture, test and operations documents.
13.2 Supplier access to Customer accounts is transferred or removed within [x] days.
13.3 Knowledge transfer: [x] hours with the Customer's team or a new supplier.
14. FEES AND PAYMENT
Fees and payment: [as agreed in the commercial schedule]
15. SIGN-OFF
Customer owner: [name, role]
Supplier lead: [name, role]
Signed for the Customer: [name, date]
Signed for the Supplier: [name, date]
How do you use the template?
Who writes it. You own the SOW, even if the supplier drafts it. The owner of the business measure fills in sections 1 and 2, the data owner section 3, and someone who does the work today signs off the evaluation set.
When. Write sections 1 to 3 before you talk to suppliers. Agree section 4 during discovery, once someone has looked at the data, because numbers set before that are guesses. Many teams sign a short SOW for discovery first and the build SOW after the go decision. If you are still choosing a supplier, see how to choose an AI development company.
Before the SOW. If you are inviting several suppliers, our AI RFP template sets up the comparison, and how long an AI project takes helps you judge their milestones.
The numbers. Measure how your team does on the same cases today, set critical failures to zero, and set the pass rate where the system is useful with a person handling the rest, not where a demo looked good.
This guide is not legal advice: a lawyer should review the final contract, including how the SOW fits your master agreement and the law that applies to you.
For how we scope AI work, including the discovery phase and the evaluation set we agree before building, see our services and AI consulting.
Questions buyers ask
What should an AI consulting statement of work include for an enterprise GenAI automation project?
The usual sections plus acceptance criteria written as numbers on an agreed evaluation set, data terms covering access, training and personal data, handover rules, model dependencies and who holds each account, prompt-injection testing, monitoring after launch, a warranty tied to measured behaviour, and exit terms.
How do you write acceptance criteria for an AI system?
Agree a set of real cases with the correct outcome for each, keep part of it back from the supplier, and write the pass marks as numbers: overall pass rate, zero critical failures, correct handovers, abstention when the answer isn’t in the data, latency and security cases. Score each category separately.
Is a statement of work the same as a statement of objectives?
No. A statement of work describes the work; a statement of objectives describes outcomes and leaves the method to the supplier. An AI SOW works best as a hybrid: fixed outcomes and acceptance tests, with the method left open.
Who should own the prompts and the evaluation set?
Usually you. You need both to retest after a model update or to change supplier, so name them in the ownership section.
Should the SOW name a specific model?
Only if you have a reason, such as a data residency requirement. Otherwise fix the acceptance criteria, let the supplier choose, and require any model change to pass the evaluation set again.