Choose the operating pattern before choosing the AI technology.
The Workflow Suitability Framework helps teams decide whether a workflow calls for deterministic automation, enterprise search, retrieval-augmented generation, an agentic system, predictive machine learning, or no AI at all.
Start with the job, uncertainty, source material, and cost of failure.
Is the work a fixed transformation, information lookup, grounded synthesis, multi-step execution, or prediction?
Are the rules and inputs structured and predictable, or variable and dependent on interpretation?
Can an error be reversed cheaply, or could it create financial, safety, compliance, or customer harm?
Must every answer be traceable to approved sources, records, or model features?
Should the system only recommend, or may it write data, trigger tools, and advance a process?
Can quality be measured before launch and monitored under real operating conditions?
Match the workflow to the least complex pattern that can do the job reliably.
| Pattern | Best fit | Primary control | Warning sign |
|---|---|---|---|
| Deterministic automation | Stable rules, structured inputs, repeatable transformations | Tests, validation rules, exception queues | The task depends on open-ended interpretation |
| Enterprise search | Users need to find exact approved documents or records | Permissions, indexing, ranking, source freshness | The user actually needs a synthesized answer |
| RAG | Answers must synthesize changing, approved source material | Citations, retrieval evaluation, abstention | The corpus is incomplete, stale, or poorly owned |
| Agentic workflow | Bounded multi-step work across tools with recoverable actions | Tool permissions, checkpoints, logs, rollback | Errors are irreversible or cannot be detected |
| Predictive ML | Historical patterns can improve forecasts, scores, or anomaly detection | Backtesting, drift monitoring, decision thresholds | There is no stable outcome label or usable history |
| No AI yet | The workflow, ownership, value, or source quality is unresolved | Process redesign and data remediation | A tool is being used to avoid fixing the process |
Use a six-step routing test.
- Define the exact input, task, output, user, and business consequence.
- Remove steps that can be simplified or eliminated without AI.
- Identify whether the core need is retrieval, synthesis, execution, or prediction.
- Set the maximum acceptable action authority and failure exposure.
- Choose the simplest viable pattern and define a measurable baseline.
- Pilot against real exceptions before expanding scope or autonomy.
Do not use an agent where a rules engine works, do not use generation where search works, and do not use AI where a process fix removes the problem.
Real workflows often combine patterns, but each boundary should remain explicit.
Search exposes exact records while RAG synthesizes an answer with citations. Evaluate retrieval and answer quality separately.
Retrieval grounds a recommendation before an agent takes a bounded action. Require approval for consequential steps.
A score informs a deterministic threshold or human decision. Monitor both model drift and downstream outcomes.
Rules complete the routine path while exceptions route to a person with enough context to decide.
“No AI yet” is a valid and often valuable result.
- No accountable workflow owner or adoption path
- No baseline against which improvement can be measured
- Unusable source data or unclear access rights
- Failure consequences exceed the available controls
- Low process volume or value cannot justify operating cost
Resolve the process, ownership, data, or control gap first. Then repeat the suitability review with a narrower workflow target.
Turn technology selection into an accountable implementation decision.
Use this framework after the AI Readiness Framework establishes that a workflow is worth pursuing. Then connect the selected pattern to the appropriate implementation and evaluation plan.