The planning call ends at 10:00. By 10:07, the meeting assistant has created three tasks, drafted a customer update, found two scheduling conflicts, and asked four people to review its work.
The follow-up is impressively fast. It has also restarted the meeting in everybody's notifications.
This is the next productivity problem for workplace AI. Once an assistant can continue working after people leave the room, capture is no longer the hard part. The hard part is deciding when that work should return, where it should appear, and what deserves an interruption.
A deadline does not answer those questions. "Finish by Thursday" tells an agent when the work must exist. It does not tell the system whether to send partial results on Tuesday morning, post every source conflict in chat, wait for the next project review, or interrupt the owner as soon as a risk appears.
Teams need a return contract: a small agreement that defines how delegated work re-enters human attention.
Background work still has a foreground cost
On September 9, Google announced new agentic capabilities across Workspace that can complete cross-application tasks, create files for review, manage scheduling, log reminders, and build task lists. The stated goal is to reduce tool switching and let work continue without breaking a person's flow. Google Workspace
That is a meaningful shift. A meeting can now produce work that continues in the background instead of waiting for someone to open five applications. Yet every background process eventually asks for foreground attention. A draft needs approval. A conflict needs judgment. A finished task needs to be placed where the team can use it.
If the return path is undefined, the agent will tend to use whatever surface is easiest: a direct message, an email, a task notification, a comment, or several of them. Each individual alert may be reasonable. Together they create a second layer of coordination that people must continuously monitor.
Recent research gives a useful reason to treat this as a work-design question rather than a personal focus failure. A September 2026 study of 516 employees in India's IT sector found that AI adoption and digital work intensity were associated with both technostress and psychological resilience. The study did not find that technostress itself significantly reduced wellbeing in its sample, and it cannot establish a universal causal effect. Its more interesting conclusion is the dual one: AI-enabled work can create resources and demands at the same time. Frontiers in Artificial Intelligence
That is exactly what unplanned follow-up feels like. The agent saves effort, while the human absorbs a new obligation to notice, interpret, and route its output.
Add five fields to every delegated follow-up
A return contract can sit beside an action item in the recap. It needs five fields.
Return event: What should bring the work back? This might be completion, a material conflict, a missing approval, a confidence threshold, or a fixed review time.
Return time: When is the earliest useful moment to surface it? Distinguish "complete by Thursday" from "hold all non-urgent results for Thursday's 2:00 review."
Return place: Which channel becomes the single home for the output? Choose the project tracker, a named document, the next meeting record, or one specific conversation. Do not make people reconstruct the result across several surfaces.
Return audience: Who actually needs to see it? The task owner may need the full working record. The broader team may need only the final decision or a short exception note.
Interruption rule: What is important enough to override the planned return? Name the condition narrowly, such as a customer commitment at risk, a source conflict that changes the recommendation, or a deadline that cannot be met without changing scope.
Consider a common action item: "Compare the two launch plans and recommend one by Thursday."
With a return contract, it becomes:
Compare the two launch plans using the approved budget and staffing files. Add the recommendation and source links to the launch decision document by Thursday at noon. Priya reviews it at 2:00. Do not send progress messages. Interrupt Priya earlier only if the plans use conflicting customer dates or neither option fits the approved budget.
The agent's assignment is almost unchanged. The team's attention burden is completely different.
Separate progress from exceptions
People often ask for status updates because they do not trust the process to reveal trouble. Agents can make that habit worse by producing a steady stream of reassuring activity: sources checked, records updated, draft created, analysis 60 percent complete.
Most of those events are useful for an audit trail. They are poor candidates for human attention.
Keep a quiet progress log that a person can inspect when needed. Surface an exception only when the next step requires a decision, a promised outcome is threatened, or the task has crossed a boundary named in the meeting. This preserves visibility without turning visibility into interruption.
The distinction also improves escalation. "The agent hit a problem" is too broad. A useful exception says what changed and what choice is needed:
The staffing plan lists two different support owners for launch week. The recommendation changes depending on which record is current. Confirm the authoritative owner, or ask me to continue with the newer file and mark the assumption.
That message earns attention because it reduces uncertainty that the agent cannot responsibly resolve alone.
Match the return to the rhythm of the work
Not every output belongs in real time. Use three return modes.
Immediate: Reserve this for consequences that become harder to repair with delay. Examples include an external deadline at risk, a permission problem, or evidence that an approved action could harm a customer.
Batched: Collect ordinary completions, drafts, and low-risk questions for one or two planned review windows. This is the default for most project follow-up.
Next conversation: Carry work into the next scheduled meeting when its value depends on collective interpretation. A research summary, scenario comparison, or unresolved tradeoff often benefits from arriving as part of the agenda rather than as another thread everyone is expected to debate asynchronously.
The right mode depends on consequence, not speed. A task finished in three minutes may still wait until the afternoon review. A task due next week may deserve an immediate alert if it uncovers a customer promise nobody had recorded.
Close the loop in the next meeting
A return contract should not disappear once the result arrives. At the next relevant meeting, spend one minute checking the loop:
- Did the work return at a useful time?
- Did it arrive in one reliable place?
- Were any interruptions necessary?
- Did the right person make the final judgment?
This is not a performance review for the agent. It is a way to tune the team's operating rules. If every task triggers an exception, the boundary is vague. If important conflicts sit quietly until the deadline, the interruption rule is too narrow. If people keep asking where results live, the return place is wrong.
Over time, recurring meetings can develop stable defaults. A sales review may batch drafts for 4:00 and interrupt only for customer risk. A hiring debrief may return research to a restricted document and wait for the next panel discussion. An incident call may treat nearly every material change as immediate.
The useful pattern is not "let AI work autonomously." It is "let AI work quietly until human attention can change the outcome."
Caspi is designed to support the full meeting loop with live recap, suggested questions, contextual chat, proactive flags, post-call action items, and persistent meeting memory. That makes the meeting a natural place to record not only what happens next, but how the result should come back. A clear return contract helps Caspi support follow-through without making the follow-through compete for attention all day.