Problem-Definition Method
Purpose
Define the problem, compare viable solutions and identify the evidence needed before committing to an AI system. The outputs support AI-013 Phase 1.
Scope
Applies to any proposed AI or machine learning system governed by the AI Management System. Use it for new systems and for systems where a solution has already been selected and must be validated.
Lifecycle Position
This method produces the Problem Definition and inputs to the feasibility study, design specification, impact assessment (AI-009) and risk assessment (AI-008).
- Keep scope and outcomes here; binding functional detail belongs in the design specification.
- Steps 5–6 screen harms and error costs. Record formal findings in the impact and risk assessments and link them; the risk register is the single source of truth for risks.
- Phase 1 is complete when AI-013’s Phase 1 Definition of Done is met.
Principle
Define the problem before committing to a solution. Work the steps in order using available evidence, because each answer informs the next. Do the inexpensive checks first, and test whether simpler options suffice before making expensive commitments; a finding may reopen an earlier step.
Step Summary
| Step | Question | Output |
|---|---|---|
| 0 | Is the solution already decided? | Inherited decisions |
| 1 | Why does the problem exist, and why solve it? | Problem statement, causes, business value, AIMS objective mapping |
| 2 | What happens today, and who is affected? | Current situation, stakeholders, open questions |
| 3 | Is machine learning required? | Capability-by-solution comparison; machine-learning boundary |
| 4 | What must the system do? | Provisional task framing, human baseline, accuracy basis |
| 5 | What limits and risks apply? | Inputs to AI-008 and AI-009 |
| 6 | What does an error cost? | Reliability stance, cost screening |
| 7 | Approved to proceed? | Go/No-Go decision record |
Procedure
Step 0: Determine the starting point
Identify any platform, tool, supplier or constraint already selected. Distinguish inherited decisions from conclusions reached here, with their source. Even an inherited solution must still be tested against simpler alternatives in Step 3.
Step 1: Define the problem, its value, and its AIMS alignment
Identify the actionable causes, why solving them matters to the sponsor and how success would be recognized. State the problem without technology terms and map the benefit to relevant AI-002 objectives. Do not steer the reasoning toward a predetermined solution.
Step 2: Describe the current situation and identify stakeholders
Describe how the work happens, using strategic and day-to-day perspectives. Consult each stakeholder type, or record the consultation gap and whether it changes the scope or a risk; do not infer their views. Where parts of the problem differ in difficulty, risk or success measures, decide deliberately whether they are one problem or several. Record open problem and boundary questions and who can answer them; route rule and scenario questions to the scenario inventory.
Step 3: Identify the simplest viable solution
For each distinct capability, consider solutions in this order of preference: remove the cause; use fixed rules; use or improve an existing tool or third-party product; build an AI or machine learning system. Record, as a table or a list, each capability’s selected solution and the rejected alternatives with reasons, including relevant parallel initiatives. Assign AI only where the simpler options do not meet the need, and keep the boundary at outcome level.
Step 4: Frame the system’s task
State the general input and required output for each AI capability, how a skilled person performs the task and how success will be judged. Use human-approved outcomes as the comparison basis. Record missing baselines or unresolved measures. Binding contracts, accuracy units and thresholds belong in the design specification.
Step 5: Screen constraints and risks
Identify concrete financial, operational, customer, regulatory and data harms. Check available evaluation data and its coverage, integration limits, and speed and volume constraints. Establish which safeguards are effective and which remain proposed. Record formal assessment content in AI-009 and AI-008 records; link them and expose any assessment or constraint still unresolved.
Step 6: Screen the cost of errors
Determine the consequences of errors and the required reliability. Explain whether errors are tolerable and why, including how they are caught. Weigh recurring benefits against rare serious losses, and consider indirect effects such as corrupted feedback data and reputation. Record cost screening in the risk assessment and reflect it in success criteria; unresolved management judgments remain open.
Step 7: Record the decision (Go/No-Go)
Check the objective mapping, linked impact and risk assessments, feasibility and design evidence against AI-013 Phase 1 and the governance review checks. Prepare the decision record with genuine conditions or unresolved matters. Record the authorized management decision, including its date and approver, through GitLab’s native approval workflow; link that record rather than duplicating signatures in the problem definition.
Assembling the Document
The problem definition must include every required template section. State each fact once and link supporting evidence.
| Step | Template section |
|---|---|
| 0 | 1.2 Business Context; 4.2 Constraints |
| 1 | 1.1 Problem Statement; 1.4 Impact of Not Solving; 2.3 Alignment with AIMS Objectives |
| 2 | 1.3 Current State Analysis; 3.1 Primary Stakeholders |
| 3 | 1.5 Alternative Solutions Considered; scope boundary in 4.1 Functional Requirements |
| 4 | 1.3 Current State Analysis (human baseline); 2.1 Primary Objectives; 2.2 Success Criteria; 4.1 Functional Requirements |
| 5–6 | 4.2 Constraints; 5. Preliminary Impact and Risk Identification |
| 7 | Separate decision record and GitLab approval |
Use concise prose, lists or tables. Document decision-bearing evidence, constraints, targets and unanswered questions; a template section without its supporting step output is marked open, with who can resolve it. An open item blocks the Go decision when it prevents setting a required success criterion or accepting the associated risk; otherwise the decision record carries it as a condition or a follow-up with an owner.
Relationship to Requirements Engineering
Problem definition sets the scope and AI boundary. Example Mapping then elicits rules and scenarios within it. Keep one question list per level: problem and boundary questions here, rule and scenario questions in the linked scenario inventory.
Using an AI Assistant
An assistant may gather existing evidence, draft and compare options, identify gaps and check consistency. Stakeholder contact, judgment and approval remain with the responsible people. Require source-grounded claims, willingness to reject AI when unnecessary, and outcome-level wording; never invent stakeholder views, assessment findings or approvals.
Summary
Define the need, the simplest viable response, success and limits. Link formal assessments and supporting evidence, then obtain the AI-013 management decision.
Revision History
| Version | Date | Author | Summary of Change |
|---|---|---|---|
| 1.0 | 2026-06-03 | Field Bradley | Initial version. |
| 1.1 | 2026-11-07 | Field Bradley | Simplified the method within its steps, accepting the capability-by-solution comparison as a table or list, recording consultation gaps by their effect, and blocking the Go only on open items that prevent a required criterion or risk acceptance (AI-2167, AI-2131; Field, 2026-11-07). |