A proposal appears line by line in an internal AI tool. The account manager starts reading, then the connection drops halfway through a pricing condition. The interface has already made the work feel nearly finished, even though the manager cannot yet rely on the terms.
Treat streamed AI output as provisional until the relevant response or tool-call boundary is complete and the application has checked it. Displaying a preview and committing a business change should be separate decisions. A completed model response still requires business validation; a successful write requires confirmation from the destination system.
Keep approval disabled while the proposal is provisional. The account manager can read ahead, but an interruption cannot leave an incomplete pricing condition available for sign-off.
Choose what the user can safely see early
Streaming delivers output incrementally. It gives the reader something before the entire response is available, but it does not establish that the result is complete or accepted.
OpenAI’s streaming guide describes separate text-delta and completion events. It also warns that partial completions are harder to moderate. If the product must check an entire answer before showing it, exposing raw deltas defeats that requirement. Buffer the answer or release independently checked sections instead.
For an internal drafting interface, provisional text can be useful. Label it as still generating and prevent export or approval until the required checks finish. A user should be able to distinguish an interrupted draft from a ready proposal without inspecting a developer console.
Where the content itself is unsuitable for early exposure, show progress recorded by the application, such as a completed source fetch. Do not invent progress percentages or present model-generated claims about work as execution evidence.
Match completion to the object being used
Different consumers require different completion boundaries. A text preview can append a fragment; a tool invocation requires usable arguments. Neither operation proves that the entire business task has succeeded.
Anthropic’s streaming documentation explains that tool-input deltas are partial JSON strings. It documents accumulating those strings and parsing them at the content-block boundary, with SDK helpers available for incremental handling. The same page describes errors that arrive inside an established stream.
For consequential actions, use an explicit application rule: assemble the complete tool call, validate its arguments, check authority, and only then dispatch it. A partial parser is useful for display; it does not authorise execution. Follow the provider’s event contract rather than assuming that closing a browser connection means the server completed its work.
A tool call can legitimately finish before the model’s whole turn ends. Waiting for the final conversational answer before executing every tool would break that interaction. The correct boundary is the completed, validated and authorised call, followed by the destination’s execution result.
Structured output helps define the data contract at that boundary. It leaves a separate question about whether the values are supported by evidence, covered in Do Structured Outputs Prevent Hallucinations?.
Make the proposal screen reflect the actual state
Consider a hypothetical consultancy drafting a renewal proposal. This is a product-design example, not a Model Operator deployment result.
While the model generates, the screen displays a provisional draft. Commercial fields remain read-only. Once the required response completes, the application checks the account, permitted pricing and source-backed terms. Only an accepted version becomes available for approval.
If the user approves a proposal, bind that approval to the exact version. An automatic regeneration must not inherit consent given to different wording. The application then submits the approved version and records the destination’s response. The screen can show “saved” only after confirming that save.
| Observed condition | What the interface should permit | What remains unproven |
|---|---|---|
| Text is arriving | Read a labelled preview, if early exposure is allowed | Completion and acceptance |
| Stream ends without the expected completion boundary | Retain an interrupted draft and offer a recovery route | Whether generation finished successfully |
| Complete output fails business checks | Explain the rejected condition and allow correction | Suitability for approval |
| Approved write times out | Show an unresolved submission while checking the destination | Whether the write happened |
| Destination confirms the intended record and version | Show the saved result with its reference | Any later workflow steps |
The timeout row is commercially important. A second click can create duplicate work when the first submission succeeded but its acknowledgement was lost. Keep the submission identity stable and reconcile before retrying a consequential write. The AI agent rollback guide develops that recovery design.
Test the interrupted experience alongside the happy path
Use the following as a proposed acceptance worksheet. These are tests to run against the implementation, not benchmark findings.
Interrupt generation halfway through a commercial condition. Check that the visible content remains provisional and cannot be approved or exported as final. Then interrupt a tool’s argument stream and verify that no destination action occurs from the partial input.
Next, return a complete response with a disallowed price. The interface should leave generation state successfully while withholding commercial acceptance. This catches products that confuse a provider completion event with permission to use the result.
Simulate a save that succeeds at the destination but loses its acknowledgement. Record whether recovery finds the original record or creates another. Finally, run an ordinary valid proposal through the same route and confirm that it reaches an approved, saved version. A design that blocks every submission has not passed.
Capture the rendered state and destination record for each case. Logs showing that a stream handler returned successfully cannot establish what a user was able to approve or what the business system stored.
Approve the latency experience, not just the first token
Before commissioning streaming, agree what useful content can appear early and which actions must wait. Measure time to the first useful preview separately from time to an accepted result. A fast opening fragment does not compensate for a proposal that the user must repeatedly restart.
For a product owner, the next decision is concrete: select one consequential screen and define its interrupted state before signing off its fast demonstration. If that ownership is unclear, the implementation scope is unfinished.
Model Operator is a Dubai-based AI product studio founded by Alexander Chrisostomou. Through AI Initiative Consulting, Alexander works with teams on workflow priorities and implementation decisions. Bring one AI-assisted screen and its destination action to a Model Operator build conversation to agree the boundary between a useful preview and work the business can accept.