Choose the approach that fits the task. Some work needs a simple rule, some may benefit from AI, and some needs human judgement.
Use rules for predictable steps
An automatic reminder or task creation can work well when the conditions are clear. Even simple connections can fail, so agree how to spot and fix problems.
Use AI where it adds value
AI can help interpret varied information, such as summarising an email. An AI agent may also take actions using connected tools within its assigned limits. A full agent is not needed for every AI task.
Keep people involved
Decide who reviews sensitive or consequential actions and who can correct mistakes. Compare the benefit of each approach with its cost and risk.
Match the approach to the work
For an order update, a fixed rule can alert a colleague when a delivery date changes. AI might summarise the carrier’s free-text message. An agent could look up the order and prepare a draft update using connected tools.
The colleague still checks the source and decides whether to send the update. If the message is ambiguous or the order cannot be matched, the system should leave it for review. This example does not require an agent if a rule and a draft meet the need.
Set permissions around the action
Separate reading information, preparing a recommendation and changing a record. A system that can draft a customer update does not necessarily need permission to send it. An agent that can look up an order does not automatically need access to every customer record.
For each proposed action, name the owner, the information required and the point at which a person approves it. Record what happened and give the team a way to stop, correct or retry the work. Test unusual and incomplete cases before expanding access.
Read what AI automation means for a small business for a straightforward example.
For the privacy risks of giving agents access to information, read the ICO’s guidance on agentic AI.