Design for Shared Control
How to design AI interactions that make progress, authority and recovery understandable.

On this page
A coordinator asks an assistant to rearrange an appointment. It finds a free time, drafts a message and proposes a change to the booking. The coordinator needs to know what is still a suggestion, which details were checked and whether the appointment has already moved.
A smooth conversation can hide those distinctions. The interface needs to show enough of the work for the coordinator to check it and intervene.
Choose a surface for the task
An inline suggestion works when the user is already completing a bounded task. A workspace can support open-ended analysis with files, sources and editable artefacts. A delegated job may need a progress view, an exception queue and a way to return later. A generated table or form may help when the structure of the task varies.
These patterns can coexist. Choose them from the user's workflow, available context and desired control. An empty prompt is not always the best starting point; examples, suggested actions or known fields can help a person see what the product can do.
Show the state that matters

For delegated work, users may need to see what is planned, active, complete or blocked; the sources or context used; and the action that needs their approval. A plan, change preview or concise activity history can provide control without streaming every internal operation.
Distinguish a suggestion from an external action. If a system proposes sending a message or changing a record, make the consequence clear before the user approves it. The action boundary is also a technical control: the interface should not imply a permission the system does not have.
Make uncertainty actionable
A fluent answer can conceal missing evidence. Show relevant sources, distinguish an observation from an inference and identify when required information is missing. Offer a way to verify or correct the result, or to ask for help.
A numeric confidence indicator is useful only if it has been calibrated for the task and people understand what to do with it. Otherwise, a direct explanation of the missing information or required review may be more useful than a number.
Design recovery into the flow
Preserve the user's input and completed work when a tool fails or times out. Explain whether any external action occurred, what remains unfinished and what the user can do next. Offer retry only when it is safe; a timed-out action may already have completed. Provide undo, compensation or a human handoff where appropriate, and carry the relevant context into that handoff.
For voice, make it easy to correct a misunderstood detail and confirm critical information. For images or other input, distinguish what the system observed from what it inferred. Limit collection and retention to what the task needs. The voice playbook covers the operational detail.
Test what people understand
Watch people complete representative tasks and ask them to explain what the system did, what is still a draft, which sources were used and what will happen after the next action. Test a failure and recovery path as well as the successful one.
Acceptance, edits, overrides and repeat use can reveal useful patterns, but they do not establish quality on their own. A person may accept a correct result, accept an error or edit a good draft to suit their situation. Pair behaviour with independent assessment and direct observation. Product metrics covers how to connect those signals to decisions.
v3.1 · Updated Oct 2026