Teams often try to buy AI and then ask what work it should take over. OpenAI’s research points in a different direction: workers are already using AI for tasks outside their usual role, then folding those tasks back into the way the work gets done.
That sequence matters. If AI first shows up as a chat window, people usually use it for planning, research, troubleshooting, and drafting. In other words, it becomes a helper at the front of the workflow before it becomes an automated step inside it.
The design question is not “where can we remove a person?” It is “which part of the workflow should change first so the next step becomes safe to automate?” A claims team, for example, might use AI to compare documents, surface missing fields, or draft a summary for review before any decision is automated. That is a narrower and more supportable change than handing the entire case to an agent.
OpenAI’s recent work on agents also reinforces this pattern. Agents are moving from short prompts to longer-horizon tasks, including technical execution by non-technical staff. That expands what software can do, but it also raises the cost of a bad workflow design. If the handoff is messy, automation does not simplify the process; it just moves the confusion faster.
A useful rule: redesign the workflow around the decision, then automate the step that becomes repeatable. That gives operations teams a cleaner path on governance, support, and adoption, because people can still see where judgment stays human and where software takes over.
AI work usually fails when the tool is the first decision. Workflow design should be.