AI for operators4 min read
The question before the AI pilot
An assistant can draft a handoff in seconds. Before it sees a real one, there are three things worth checking.
An AI assistant can turn a shift handoff into a checklist. That's an appealing little job: take a messy note and give the next person something they can actually use.
The useful question is what happens to the handoff after you give it one.
Not the imaginary version in a demo. The actual note with a staff name, a delivery problem and a phone number in the margin.
Before putting that note into a tool, get three answers: which account is handling it, who can reach it, and whether the task works without it. The first two depend on the product and the account. The third is a test the team can run.
The account matters
OpenAI says it does not use inputs and outputs from ChatGPT Business, ChatGPT Enterprise and its API platform to train or improve its models by default. Its business data page describes an explicit opt-in for sharing data to improve services. That is a statement about the named business products and API. It is not a blanket description of every consumer account or every app built on a model.
Anthropic describes a similar default for data generated through its commercial products. It also describes a voluntary Development Partner Program through which an organization can choose to share Claude Code sessions for service improvement, including model training.
Those statements answer the training question. They do not, on their own, answer who can access a conversation, how long it stays in a system or what happens in a connected service. For those, ask about the exact plan, the settings in that account and the agreement attached to it. Write the answers down with the date and a link to the documentation.
That's not paperwork for its own sake. The name on a tool is a poor substitute for knowing which account a person actually opened. If the answer is still unclear, a fictional handoff gives the team something to test while it gets the answer.
The handoff to the tool
The next boundary is access. A connected assistant may need permission to look at files, messages or calendars to answer a request. The question is not just whether the tool can connect. It's what the person using it can already reach.
Google says Gemini in Workspace has the same access to Workspace data as the user, subject to administrator and content-owner controls. Its documentation names Drive files shared with that user and calendar events the user can view. It also says a Drive owner's restrictions on downloading, copying and printing can prevent Gemini from accessing a file.
That makes the check specific. Which staff account is signed in? What can that account open? What has the administrator allowed? The answers can differ from one person to the next, even when the team calls the whole thing “the same drive.”
For a pilot, start with a small test folder containing only sample material. Sign in as the person who will use the tool. Ask an approved question. Compare the response with the permissions the team intended to give that account. If the result does not match the plan, stop and work out why before connecting more material.
There is a second handoff after the answer is generated: the output goes somewhere. Decide who needs it, where it belongs and who removes the test material when the pilot ends. A statement about model training does not answer those questions for you.
A test a team can actually run
Make up a shift note. Give it an imaginary delivery delay, a maintenance item and a deadline. Leave out real names and contact details. Then try this:
Turn this handoff into a short checklist. Keep the deadline, mark anything uncertain as uncertain, and don't add tasks that aren't in the note.
That's a small enough job to see what the tool actually did. It also gives the next person a clear place to check the work.
Put the result next to the original. Check the deadline first. Then look for anything the assistant added, omitted or converted from an uncertain note into a definite instruction. Have the person responsible for the handoff make the comparison; a tidy checklist is not much use if it sends the next shift after the wrong thing.
If the first test is useful, make a second fictional note with a different shape and run it again. Keep both inputs and outputs together. A useful result is worth getting curious about: where did the assistant help, and where did the instruction need to be clearer? The point is to give the team something it can inspect, not to collect a flattering screenshot.
Only then move the conversation to real material. Name the person who can approve that step, decide what stays out of prompts and check the terms and permissions for the actual account. This test does not establish that a tool is secure, and it does not establish that its next answer will be right. It does show whether this particular task has earned a closer look.
Keep the input beside the output, and make the receiving person check the handoff.

