How to scope an AI automation project in one workshop
A practical one-workshop method for turning an AI automation idea into a bounded, testable pilot with owners, controls and success measures.
Written by the Problem Solver Team with AI assistance.

The short answer
Run a focused workshop with the people who own the work. Map one current process, choose the narrow decision or handoff to improve, document the inputs and systems involved, set human-approval boundaries, and agree how a pilot will be judged. The output is a testable scope, not a tool shopping list.
Key takeaways
- Scope a single decision or handoff, not an entire department’s workflow.
- Map the current process before choosing an AI tool or model.
- Write explicit boundaries, escalation rules and system access assumptions.
- End with testable acceptance criteria and a named decision owner.
01The short answer
How to scope an AI automation project in one workshop: map one real workflow, select one decision or handoff, name the inputs and systems, set the human-approval boundary, and agree how a pilot will be judged. Leave with a testable decision record, not a vague request to “add AI.”
02What should the workshop produce?
A good workshop produces a bounded problem statement and a decision about the next smallest safe step. It does not need to select a vendor, model or architecture.
The output should fit on a few pages and answer:
- What work is painful, slow, inconsistent or hard to see today?
- Which single decision, classification, draft or handoff could change first?
- Who receives the result and who remains accountable?
- Which systems are the sources of record, and what access is actually available?
- What must the system never do without approval?
- What evidence would make the team call a pilot useful?
This extends the selection work in which business process to automate first. If the workflow is unclear or unstable, resolve that before adding AI; start with the process.
03Who should be in the room?
Invite the process owner, one or two people who perform or supervise the work, the owner of the relevant system or data, and the person who can authorize a small next step. Bring security, privacy, legal or compliance input when the workflow contains sensitive data, regulated decisions or meaningful customer impact.
The facilitator keeps the discussion anchored in a recent real example. “Automate customer follow-up” is too broad. “Prepare a draft follow-up when a support case has waited two business days, then route it to the team lead for approval” is a scope the group can examine.
NIST’s AI Risk Management Framework encourages teams to document intended purpose, context of use, assumptions, limitations, human roles and impacts. That is why the workshop needs operators and system owners, not only sponsors or tool enthusiasts.
04What is the one-workshop agenda?
Use the following sequence. It turns a solution idea into observable work before any build commitment.
| Segment | Question to answer | Tangible output |
|---|---|---|
| Frame the outcome | What business or user problem are we trying to reduce? | One-sentence problem statement |
| Walk a real case | What happened from trigger to completion? | Current-state process map |
| Choose the slice | Which decision or handoff is narrow enough for a pilot? | In-scope and out-of-scope list |
| Inspect inputs | What data, documents, systems and exceptions are involved? | Input and dependency list |
| Set controls | What requires human approval, logging or escalation? | Control and ownership rules |
| Define proof | What would show the pilot is useful and safe enough to continue? | Acceptance criteria and next decision |
1. Frame the problem before the solution
Start with the user or operator outcome: “When [trigger] happens, [role] needs to [make a decision or complete a handoff], but [specific friction] causes [observable consequence].”
Do not start with “we need an agent.” Discovery guidance from GOV.UK makes the same distinction: reframe a preselected solution as a problem, then examine users, context and constraints. A workshop can legitimately conclude that clearer templates, better routing or a conventional automation is the better next move. AI vs RPA vs AI agent helps make that tool choice after the problem is understood.
2. Map a representative case, including exceptions
Ask an operator to narrate a recent case from first trigger through completion. Capture the systems touched, manual judgments, waiting points and common exceptions. Separate observed facts from assumptions.
For example, an operations team may receive an inbound request, check an account record, consult a policy document, draft a response and ask a lead to approve exceptions. That is not yet an AI project. The workshop must identify the smallest unit that can be tested—perhaps drafting a response only when the account record and policy version are present.
Process and task mining can help teams inspect actual process traces where event logs or recorded tasks are available. It is useful evidence, not a substitute for speaking with the people who operate the process.
3. Draw a hard boundary around the pilot
Write two lists: “the pilot will do” and “the pilot will not do.” Boundaries can include eligible request types, channels, language, systems, customer segments and maximum action authority.
A narrow first scope makes testing and control more realistic. It also makes a no-go decision possible. For a drafting pilot, a sensible boundary might be: generate a proposed reply from approved sources; never send it; route uncertain cases to a named reviewer. This connects to AI agent acceptance criteria: define what “good enough” means before any go-live decision.
4. Make dependencies and controls explicit
For each input, document its source of record, owner, freshness expectation and access method. For each output, identify the recipient, record of action, and whether it can change a customer, financial or operational state.
Then decide the human role. Who can approve? What must cause escalation? What is logged? What happens when a source is missing or the result is uncertain? If the team cannot answer these questions, the scope is still discovery—not implementation.
5. Agree acceptance criteria and the next decision
Avoid promises such as “save time” without defining an observation method. Write a small set of criteria a reviewer can inspect:
- The pilot handles only the agreed request type and leaves other cases untouched.
- A reviewer can see the inputs used, draft output and final disposition.
- Cases missing an approved source are routed for human handling.
- The team has an agreed sample, review method and owner for the go/no-go decision.
Do not invent a target percentage in the workshop. Establish a baseline if needed, then decide what evidence would support continuing, revising or stopping. This makes a later conversation with an AI implementation partner more concrete: they can respond to an evidence-based scope rather than a broad aspiration.
05A worked example: draft, do not send
Consider a fictional internal operations queue where staff prepare replies to routine supplier documentation requests. In the workshop, the team maps the current steps and learns that requests vary widely, but a subset has a standard policy answer.
The initial scope is not “automate supplier communications.” It is: “For requests tagged as standard, create a response draft using the current approved policy; show the policy version and source fields; require a queue lead to approve before anything is sent.” The out-of-scope list excludes exceptions, missing records and changes to supplier status.
The acceptance criteria ask reviewers to inspect whether the correct policy was used, whether the draft remained within the stated boundary, whether a missing source triggered escalation, and whether the audit trail is understandable. The next decision is whether to expand the eligible set, revise the workflow or stop. This example is illustrative; it is not a performance claim.
06What should happen after the workshop?
Within a day or two, turn the board into a short scope record and circulate it to named owners. Resolve open questions about access, policy, risk and measurement before build work begins. If the record shows the problem is too broad, split it. If the process is inconsistent, stabilize it. If a conventional workflow rule solves it, choose that.
The most useful workshop outcome can be a disciplined “not yet.” It protects the team from buying a solution before they know what should change. When the scope is clear, the project has a better foundation for a small, reviewable pilot.
07Ready to turn a process into a bounded pilot?
ProblemSolver can help your team map the work, identify the smallest useful scope and define reviewable acceptance criteria. Book an AI Opportunity Audit.
08Sources
- NIST AI RMF Playbook: Map
- Microsoft Learn: Overview of process mining and task mining in Power Automate
- GOV.UK Service Manual: How the discovery phase works
09Frequently asked questions
Who needs to be in an AI automation scoping workshop?+
Include the process owner, people who do or supervise the work, the system or data owner, and someone who can make the next investment decision. Add risk, security or compliance input when the workflow handles sensitive information or can affect customers. The purpose is to expose real constraints early, not to gather a large audience.
How long should the workshop take?+
For one bounded workflow, allow roughly two to three focused hours plus preparation. The preparation should provide a real example of the work, the current steps and the systems involved. If the team cannot describe one representative case, use the session to investigate rather than forcing a technical solution decision.
What is the main output of an AI automation scoping workshop?+
The useful output is a short decision record: the problem, process slice, inputs, proposed AI role, human controls, systems and data assumptions, risks, acceptance criteria, owner and next decision. It should make it possible to approve a small discovery or pilot, defer it, or decide that a non-AI change is better.
Sources and further reading
- NIST AI RMF Playbook: Map. NIST, 2026.
- Overview of process mining and task mining in Power Automate. Microsoft Learn, 2026.
- How the discovery phase works. GOV.UK, 2025.