At 9:10, a growth manager asks an AI assistant to analyze why customers are leaving. By noon, she has a polished cohort table and a recommendation for the next campaign. At 3:00, a data scientist spots the problem: the analysis mixes trial users with paying customers, so the apparent churn pattern is mostly an artifact of how the population was defined.

The model did not invent a customer segment. The manager did not act carelessly. She crossed into a neighboring specialty with enough technical ability to produce an answer, but without the local knowledge that makes the answer trustworthy.

This is becoming an ordinary workplace tension. AI can help people draft code, interpret contracts, analyze data, plan research, and create assets that previously required a longer handoff. The opportunity is real. So is the temptation to treat a plausible deliverable as evidence that the surrounding expertise is no longer necessary.

The right response is not to force every task back through a departmental queue. It is to add a small piece of coordination before the new work begins: a boundary briefing.

Capability is spreading faster than context

Recent research suggests that AI is not simply shrinking existing tasks. It is widening the work people attempt.

In July 2026, Atlassian's Teamwork Lab reported that heavy AI users were nearly twice as likely to take on work from other teams and twice as likely to handle specialized tasks without involving an expert. The same research found a telling mismatch: heavy users reported more cross-functional work, while AI users ranked time connecting with colleagues last among the ways they spent their time. The study was a survey of 1,000 U.S. knowledge workers, so it should not be read as a universal law. Still, its central warning is useful: role expansion without stronger relationships is fragile.

An eight-month ethnographic study at a technology company, described by UC Berkeley Haas researchers in February, found a related pattern. Employees used generative AI to work faster, attempt a broader range of tasks, and extend work into more hours of the day. That study followed one company, but its interviews and direct observation reveal something surveys can miss: people often expand their own workload because AI makes another task feel newly possible.

Possibility is not the same as permission, and production is not the same as judgment. AI may help a marketer write a query. It does not automatically reveal which metric the finance team trusts, which exception the data team learned about last quarter, or who is authorized to change the campaign on the result.

Hold a 15-minute boundary briefing

A boundary briefing is a short conversation between the person stretching into new territory and someone who knows that territory well. It happens before the first consequential deliverable, not after a polished mistake has gathered momentum.

The meeting should answer five questions:

What outcome are we trying to produce? Name the decision or work product, not a broad activity such as "analyze churn."
What local context changes the answer? Identify the source of truth, important definitions, known exceptions, and previous decisions.
What can the newcomer decide? Distinguish permission to explore, recommend, draft, publish, spend, or commit the organization.
Where is expert review required? Choose checkpoints based on consequence, uncertainty, or reversibility.
What should stop the work? Define the signal that calls for a question instead of another prompt.

For the churn analysis, the briefing might establish that the growth manager can explore hypotheses and build a draft analysis. The billing table is the source of truth for paid status. A data scientist must review any new metric definition, and the work stops if customer identity cannot be joined consistently across systems. The growth lead still moves faster, but no one pretends that speed erased the data team's accumulated judgment.

Name the level of involvement

"Loop in data" is too vague to guide anyone. Does the expert need to be informed, consulted, asked to review, or asked to own the work?

Use a simple four-level ladder:

Explore: You may investigate independently. Label the output as provisional.
Consult: Get the expert's constraints and examples before producing the first draft.
Review: The expert checks defined aspects before the work is used.
Own: The domain team keeps decision authority because the risk, regulation, or organizational commitment belongs with it.

The level can change as the work develops. A prototype may begin at Explore, move to Consult when real data enters, and require Review before a customer sees it. Legal advice, safety decisions, hiring judgments, and security approvals will often remain at Own even when AI makes drafting easier.

This ladder avoids two bad extremes. One is gatekeeping, where an expert must perform every step simply because they always have. The other is domain tourism, where someone visits a specialty with an AI assistant, produces a confident answer, and leaves the expert to clean up the consequences.

Make consultation small enough to happen

Experts resist cross-functional requests when "quick review" routinely means reconstructing the entire problem. The boundary briefing should protect their time too.

Bring a one-page working frame:

the outcome and intended audience;
the inputs you plan to use;
the assumptions you are making;
the decisions you believe you can make;
the two or three questions that need domain judgment.

Ask the expert for a reference example of acceptable work and the most common failure mode. Those two pieces of context are often more valuable than a long tutorial. A strong example makes quality visible. A failure mode tells the newcomer where fluent AI output is most likely to conceal a bad assumption.

Then agree on a channel for narrow questions. This can be a project thread, office hour, or named checkpoint. The expert should not become a silent co-owner of every draft, and the newcomer should not have to schedule a formal meeting whenever one definition becomes ambiguous.

Record the reason, not just the task

Cross-functional work often breaks later because the final artifact survives while the boundary conversation disappears.

Capture a compact record beside the work:

Purpose: Identify preventable churn among paying customers.
Authority: Growth may recommend a campaign; the VP of Growth approves spend.
Source of truth: Billing status from the finance-owned table.
Review: Data validates cohort logic before the recommendation is presented.
Stop condition: Conflicting customer identifiers or an unexplained variance above five percent.

This record is more useful than a transcript. It gives the next reviewer, collaborator, or AI tool the operational meaning of the conversation. It also prevents a provisional analysis from returning three weeks later as an apparently settled company fact.

Microsoft's 2026 New Future of Work report argues that human expertise matters more as people shift from producing every step themselves to guiding, critiquing, and improving AI-assisted work. Effective oversight requires situational awareness and visibility into activity, decisions, and outputs. A boundary record makes that principle concrete: the person reviewing the work can see what was delegated and why.

Watch for wider work becoming endless work

AI-assisted role expansion can feel like growth while functioning like overload. If every newly possible task becomes your task, the reward for using AI well is an unbounded job.

Managers should review the role, not just the output. Ask what new responsibility has appeared, what old responsibility has been removed, which expertise is being developed, and whether the person now carries accountability without authority. A boundary briefing can clarify one project, but repeated crossings may signal that the team needs a redesigned role, formal training, or a different ownership model.

The goal is not to keep people in narrow lanes. It is to let capability expand with support underneath it. New reach should create learning and better collaboration, not hidden risk or a longer evening.

A good meeting lends judgment, not just information

The most valuable cross-functional meeting may be the one that teaches someone how to notice the edge of their knowledge. It does not transfer an entire profession in 15 minutes. It reveals the definitions, decision rights, exceptions, and stop signals that a general AI assistant cannot infer reliably from a prompt.

That is a natural role for meeting support. Caspi provides live recap, suggested questions, contextual chat, proactive flags from connected tools, post-call action items, and persistent meeting memory. Used for a boundary briefing, those capabilities can help a team preserve not only what work was assigned, but also the conditions and judgment that make the assignment responsible.