Proceed
Acceptance criteria are met, material risks are controlled and the organisation can own the next scope.
Public working method
The Simlyst evidence-to-operation method for practical AI
A four-gate method for turning AI interest into an auditable decision. Each gate asks for specific evidence and permits four honest outcomes: proceed, prepare, narrow or stop.
- Problem before model
- Acceptance and stop criteria
- Ownership before operation
In brief
How should an SME take an AI use case from idea to operation?
Start by defining the workflow, owner and baseline rather than selecting a model. Compare AI with simpler alternatives, test the smallest credible intervention against representative cases, and define ownership and monitoring before operational use.
The method is intentionally stage-gated. A good outcome can be a controlled release, more preparation, a narrower intervention or a recorded decision not to proceed.
What you leave with
- A decision record that survives the sales pitch
- Comparable evidence at each investment gate
- An explicit reason to proceed, prepare, narrow or stop
Four evidence gates
One decision question at each stage
Do not carry weak assumptions forward. Record the evidence, its limits and the next decision before investment increases.
01
Problem fit
Is there a defined workflow and a meaningful baseline?
- Evidence required
- Named owner, affected users, current process, baseline cost or quality, constraints and a measurable decision.
- Gate decision
- Proceed only when the problem can be observed. Otherwise gather evidence or stop.
02
Intervention fit
Is AI better suited than a simpler change?
- Evidence required
- Comparison with process repair, rules-based automation and existing software; value, feasibility, risk and full-cost assumptions.
- Gate decision
- Choose the simplest credible intervention, narrow the use case or reject it.
03
Pilot evidence
Does a bounded implementation meet agreed acceptance criteria?
- Evidence required
- Representative evaluation cases, failure tests, user feedback, quality measures, unit cost and human-review requirements.
- Gate decision
- Progress, revise, constrain or stop based on the recorded evidence—not the quality of a demonstration.
04
Operational control
Can the organisation own and monitor the capability?
- Evidence required
- Named owner, permissions, documentation, monitoring, incident path, supplier dependencies, change control and review date.
- Gate decision
- Release at an appropriate scope, retain human approval, or defer operation until control gaps are resolved.
Reusable artefact
The minimum evidence record
Use these ten fields as a one-page decision record. Link to detailed research, evaluation results and controls instead of hiding the decision inside a long presentation.
| Record field | What to capture |
|---|---|
| Decision | The specific choice this evidence needs to support. |
| Workflow and owner | Where the work happens and who remains accountable for the outcome. |
| Baseline | Current time, cost, quality, risk or user experience before intervention. |
| Hypothesis | The expected change, for whom, and over what observation period. |
| Alternatives | Process change, existing software, deterministic automation and AI-assisted options considered. |
| Acceptance and stop criteria | Minimum evidence for progress and the conditions that should narrow or end the work. |
| Evaluation cases | Representative normal, edge and failure examples with expected handling. |
| Full cost | Delivery, licences, model use, integration, review, training, monitoring and maintenance. |
| Controls | Permissions, human decisions, logging, escalation, incident handling and review cadence. |
| Outcome and next review | The recorded decision, evidence limits, owner and date for reassessment. |
Decision rules
Keep evidence and enthusiasm separate
The method becomes useful when the team agrees how evidence will change the decision before seeing the result.
Prepare
The opportunity remains credible, but data, process, skills, controls or ownership need deliberate preparation.
Narrow
Evidence supports a smaller user group, lower-autonomy workflow or more limited outcome than first proposed.
Stop
Value, feasibility, cost, risk or operability does not justify further investment under current conditions.
Consensus before differentiation
The method does not replace law, regulation or specialist assurance. Its control questions are grounded in primary UK guidance, then adapted to the workflow and evidence in front of the team.
- UK introduction to AI assurance
Department for Science, Innovation and Technology
- Guidance on AI and data protection
Information Commissioner’s Office
- Guidelines for secure AI system development
National Cyber Security Centre
Apply the method to a real decision
A self-assessment can identify where evidence is thin. A consultancy engagement can examine the real workflow, stakeholders, systems, risks and operating constraints behind the score.
Start with the decision, not the technology
Use the method on the decision that is blocking progress
Bring the workflow, available evidence and uncertainty. Simlyst can help build the record, test the riskiest assumptions and make the next investment decision explicit.