Integrating AI into workflows while keeping human control

Choose an observable task

Start with a bounded task: categorising enquiries, preparing a summary or suggesting a reply. Describe the input, output and person who will use the result. “Adding AI to the department” defines neither a goal nor a test of success.

If a simple rule is enough, consider that option before adding a model. AI earns its place through its usefulness in the work, not merely by being part of the pipeline.

Prepare the input data

Define the usable sources, required fields and handling of missing or outdated information. A precise instruction cannot compensate for a source that mixes several versions of the same record.

Limit the information sent to what the task actually needs. Check the service's terms and authorised access before entrusting it with internal data. Keep working instructions separate from the content being analysed.

Define an acceptable result

Prepare a small set of representative cases with a reference reviewed by the team. For a summary, check facts retained and omitted; for categorisation, the correct categories; for a reply, the commitments and information that need checking.

Also assess review and correction time. A fast output that needs extensive rework may offer less value than the current method. Fluent wording is not evidence of accuracy.

Put human decisions in the right place

The reviewer should see the relevant sources, the proposed result and anything that commits the business. Run this check before publishing, sending anything outside the company or making a change that is hard to undo.

For tasks with limited consequences, sample-based checking can be considered after a satisfactory pilot. Define which errors trigger a full review or a return to manual processing.

Keep a fallback process

Plan what happens when the service does not respond, the output is unusable or a source is missing. A request should be able to remain pending or move to the team without disappearing from tracking.

Keep the working rules and access needed to continue without AI. Test the handover with the people involved; a written procedure nobody can carry out is not an operational fallback.

Expand after observing the results

Compare the pilot with the previous process over a comparable period and set of cases. Look at total time, significant errors and corrections. Document limitations before increasing volume or the actions entrusted to AI.

• A specific task and an identified source.
• Quality criteria that can be checked.
• Review matched to the consequences.
• A tested manual fallback.