AI or automation: what does your workflow need?

Ask what each step needs to do

A business workflow can need several kinds of help. Copying an approved record between systems is different from understanding an unstructured email. Checking a total is different from drafting a reply. Treating all four as one AI problem makes it harder to see what the system should actually do.

Rules-based automation follows conditions you define. An AI model can help interpret variable inputs or prepare an output when the wording and structure change. A person still needs to own decisions whose consequences require their judgement.

The useful question is where each approach belongs in your workflow. You may need a straightforward integration, a small AI step inside an otherwise predictable process, or a system that can choose among several actions. Start by separating those needs.

Use rules where the answer is explicit

If your team can write the condition clearly, test whether ordinary automation can handle it. For example: when a job is approved and its customer ID is present, create a task for the assigned team. The condition, required fields and destination can all be checked directly.

Microsoft’s Power Automate documentation describes this pattern through trigger conditions: a flow runs when the specified condition is met. The broader point is useful regardless of the tool you choose. A predictable decision does not automatically need a language model.

Rules still need care. Define what happens when a record is duplicated, a connection fails or an approval is withdrawn. Confirm which system owns each field, so two integrations do not keep overwriting each other’s updates.

• Use explicit checks for required fields, dates, approved statuses and allowed values.
• Use ordinary calculations for quantities and totals, then check the inputs.
• Keep a record of what ran, what changed and what needs attention.

Consider AI where the input keeps changing

An email may describe the same request in dozens of ways. A document may contain relevant information across several pages. These are candidates for AI-assisted interpretation: classify the request, extract proposed fields, summarise the context or prepare a draft response.

Make the output inspectable. If the system proposes a delivery date, show the source sentence beside it. If the information is missing or contradictory, route the item for review rather than filling the gap with a plausible answer. Test this behaviour using examples from your actual work.

Combine the parts around a clear handoff

Here is a hypothetical enquiry workflow. An integration collects the message. AI proposes the request type and extracts relevant details. Rules check whether the customer and job reference exist. A person reviews the prepared response. The integration sends only the approved version and records the action.

The design makes it possible to identify the point of failure. Was the source record out of date, was the request interpreted incorrectly, or did an approval sit untouched? Each problem has a different fix. A single button labelled “automate everything” would hide those distinctions.

Anthropic’s engineering guidance distinguishes predefined workflows from agents that choose their own sequence of actions. It recommends starting simply and adding complexity where it improves the result. For a business owner, that means asking a builder to explain why each flexible AI decision is necessary.

Test the awkward cases before expanding

Ask for a demonstration using a mix of ordinary and difficult examples. Include incomplete requests, contradictory information, repeat submissions and a source system that is unavailable. Agree which outcomes count as correct, which should reach a person and which must stop the workflow.

Review the whole result, including the effort needed to check it. A draft that takes longer to correct than to write has not removed work. A faster process that creates a new exception queue has simply moved the burden to someone else.

• Can the team inspect the source information and correct the proposed result?
• Is there a clear owner for unresolved items and failed connections?
• Can the workflow avoid repeating an action after a retry?
• Can a person pause the system and continue the work manually?
• Does the measured improvement remain after review and rework are included?