Develop the AI Project Scope Statement
Key Takeaways
- An AI project scope statement scopes a deployed decision capability — the decision, the population it applies to, the workflow it enters, and the human oversight around it — not a model artifact or an accuracy number.
- Explicit, sponsor-signed out-of-scope statements are the primary defense against AI scope creep, which arrives as adjacent use cases ("can it also flag fraud?") and as accuracy chasing that has no natural stopping point.
- You cannot commit to model accuracy before CPMAI Phase 2 (Data Understanding); you can commit to phase-gate dates, the effort and cost of the data work, and a go/no-go decision point with named criteria.
- Every assumption about data access, labeling capacity, and infrastructure availability carries an owner, a verification date, and a stated consequence if it proves false.
- CPMAI scoping is rolling wave: scope the next iteration tightly, keep later iterations coarse, and re-baseline the scope statement at every phase gate — scope most often narrows at the Phase 2 to Phase 3 boundary.
What an AI Scope Statement Actually Scopes
The AI project scope statement is the artifact that closes CPMAI Phase 1 (Business Understanding) and authorizes the organization to spend real money on data. Domain II Task 4 of the Examination Content Outline asks you to develop it, and the trap the exam sets is scoping a model instead of a deployed decision capability. "Build a churn model that reaches 85% accuracy" is not a scope statement — it names a technical artifact and leaves everything that creates value undefined. "Score every residential broadband account monthly, push the 5,000 highest-risk accounts into the retention queue in the customer relationship management (CRM) system, and let a retention agent approve, modify, or reject each recommended save offer" is a scope statement, because it names a decision, a population, a workflow, and the human oversight around it.
The Four Things Scope Must Name
- The decision. Which recurring decision changes? Approve or deny, prioritize, route, price, schedule, or flag for review.
- The population. Which claims, accounts, machines, documents, or transactions sit in the addressable set, across which product lines, geographies, and time window.
- The workflow. Which system receives the output, on which screen or queue, at what frequency and latency, and what the receiving human or downstream system does with it.
- The oversight. Who reviews, who may override, what happens when the model abstains, and what the workflow falls back to when scoring is unavailable.
Those four statements are the project's boundaries, and the deliverables follow from them. A defensible AI deliverable list names the data inventory, the trained model and its model card, the integration into the target system, the monitoring dashboard, the operations runbook, and trained users — not "a model."
Success Criteria That Belong in the Scope Statement
Record two thresholds. The minimum acceptable model performance, expressed against the real operating constraint ("at a review capacity of 16,000 investigations a year, the model must place at least 30% true positives inside that band"), and the business success criterion in the sponsor's own words ("reduce indemnity leakage on personal auto by 5% within two quarters of go-live"). Both are stated as conditions for operationalizing, not as promises; the full performance-metric framework, including leading indicators and monitoring measures, is built in the dedicated section on defining success criteria and key performance indicators.
Out-of-Scope Statements Are the Main Defense
AI invites the sentence "and while you're at it, can it also…", because a model that reads every claim looks like it could also detect fraud, predict severity, and draft the adjuster's letter. Each of those is a different label, a different training set, a different regulatory conversation, and often a different data owner. Write the exclusions down and have the sponsor sign them:
- Out of scope: commercial-lines claims; this iteration covers personal auto only.
- Out of scope: automated denial. Every declined claim routes to a human adjuster.
- Out of scope: real-time scoring. This iteration runs as an overnight batch.
- Out of scope: non-English claim documents.
- Out of scope: automatic retraining on production data; deferred to a later iteration.
The second family of AI scope creep is subtler: accuracy chasing. Marginal model improvement has no natural stopping point, so the scope statement must state that model work stops once the minimum acceptable performance is met and remaining effort moves to deployment.
Assumptions and Constraints
Three assumption families break AI projects. Each register entry gets an owner, a verification date, and a stated consequence if false:
- Data access — "Legal and privacy approve use of the 2019 to 2025 claims archive for model training by week 4."
- Labeling capacity — "Two senior adjusters each give six hours a week for eight weeks to label 4,000 historical claims."
- Infrastructure — "The AI workspace with training capacity is provisioned and accessible by week 3."
Constraints are the fixed conditions you design inside: a latency budget (a score returned in under 400 milliseconds at the point of quote), a cost-per-inference ceiling at production volume, regulatory and approval constraints such as model risk validation sign-off or notice and consent duties under the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA), and data seasonality — needing 24 months of history to observe two full annual cycles sets an earliest possible start date, not a preference.
How AI Scope Differs from Conventional Scope
| Scope element | Conventional software project | AI project | Why it differs |
|---|---|---|---|
| Acceptance criteria | Passes a functional test suite | Meets a performance threshold on held-out data and moves a business metric | Output is probabilistic, so "correct" is statistical, not binary |
| Core deliverable | A working feature | A deployed decision capability with model card, monitoring, and retraining plan | Performance decays after release |
| Requirements | Specified up front | Hypothesis confirmed against real data in Phase 2 | Feasibility is unknown until the data is inspected |
| Dominant effort | Development and testing | Data acquisition, labeling, and preparation | Data work routinely exceeds modeling effort |
| Scope creep source | Feature requests | Adjacent use cases plus accuracy chasing | Marginal accuracy has no natural stopping point |
| Completion | Ship once | The iteration closes at a phase gate; later iterations are re-baselined | CPMAI is iterative by design |
| Change trigger | A stakeholder change request | Also the data itself — fields missing, biased, or unusable | Scope answers to what the data supports |
Scope Iteratively and Re-Baseline at Gates
Use rolling-wave scoping: the next CPMAI iteration is scoped in detail, later iterations deliberately coarse. Re-baseline the scope statement at every phase gate, and expect the Phase 2 to Phase 3 boundary to be where scope narrows most often, as the population shrinks to the segments where usable, labeled data actually exists. Aligning scope to resource availability is the same discipline: map every in-scope deliverable to a named business objective, then test that list against the subject-matter-expert hours, data engineering capacity, budget, and infrastructure you have actually been granted. When resources fall short, cut scope — never the data preparation or evaluation work.
Estimating Honestly with a Sponsor
You cannot commit to model accuracy before Phase 2, because accuracy is an empirical property of data you have not yet inspected. What you can commit to is a calendar date for each phase gate, the effort and cost of the data work, a named decision point with explicit go/no-go criteria, and the business metric the solution must move to be worth operationalizing. Say it plainly to the sponsor: "I can commit to a data-sufficiency answer by 14 October and to the effort it takes to reach it. I cannot promise 92% accuracy today. What you will get on 14 October is an evidence-based estimate and a recommendation to fund the next iteration or to stop." That framing keeps an unverifiable number out of the baseline while still giving the sponsor a dated decision.
The sponsor of a claims-triage AI initiative reads the draft scope statement and says, "Since the model already reads every claim, have it flag likely fraud as well." The team is two weeks from the Phase 2 data-sufficiency gate. What should the project manager do first?
The approved scope statement for a quoting AI names the decision, the population, the workflow, and the oversight, but records no latency budget and no cost-per-inference ceiling. In Phase 4 the best-performing ensemble returns a score in about 2.5 seconds and costs several times what the pricing team assumed per quote. What should the project manager do?
The scope statement for a clinical coding AI assumes two coding specialists will each contribute six hours a week for eight weeks to label training records. In week 2 the operations manager confirms neither specialist is available until the next quarter. What should the project manager do next?