Six questions to answer before automating a workflow with AI

1. Is the workflow defined well enough to improve?

If different people complete the same task in completely different ways, automation may simply accelerate inconsistency. First map the triggers, inputs, decisions, exceptions and handoffs in the current process.

This does not require a large documentation exercise. A simple view of what happens today, where it breaks down and who owns the final decision is enough to identify whether automation is appropriate.

2. What decision should remain with a person?

AI can classify, summarise, draft, extract and suggest. It should not quietly make high-impact decisions without a clear owner, review path and escalation rule. The more significant the consequence for a customer, employee or organisation, the more important human accountability becomes.

A useful design states what the system can do automatically, what must be reviewed and what should be escalated. That clarity protects both users and the organisation.

3. What data is necessary and appropriate?

Automation should use the minimum data needed to perform its job. Before connecting a model or external service, identify data sensitivity, retention requirements, access controls and whether personal or confidential information can be removed or masked.

This is especially important when a workflow involves client records, employee information, commercial contracts or regulated data. Privacy and security should be part of the design brief, not an afterthought.

• What data enters the workflow?
• Who is permitted to access it?
• Can sensitive information be reduced, redacted or separated?
• How will prompts, outputs and decisions be retained or audited?

4. How will quality be checked?

A model can produce a plausible response that is incomplete or wrong. Quality control should be designed around the task: sampling, mandatory review, validation against trusted data, confidence thresholds or escalation when information is missing.

For operational work, the best pattern is often a constrained workflow. The system extracts or drafts, a deterministic rule checks critical fields and a person approves where required.

5. Can the organisation explain what happened?

Teams need to be able to explain which input led to an output, what reference material was used and who approved the final action. This does not mean exposing every technical implementation detail. It means keeping the workflow understandable enough to investigate a concern and improve it.

Versioned prompts, source references and straightforward logs are often more useful than a complex autonomous system with no practical audit trail.

6. What is the smallest safe first use case?

Start with a bounded task that has clear value, manageable data and a visible quality check. Examples include preparing a first draft from approved source material, classifying a controlled set of requests or extracting structured fields from a standard document.

A small implementation gives the organisation evidence about cost, quality, user adoption and risk. That evidence should inform whether and how the capability expands.