Your AI Coworker Has Three Ongoing Projects. Which One Wins Today?

An AI Coworker Can Own Ongoing Work. It Still Needs a Priority Rule.

On November 4, Perplexity introduced Teammates for complex, ongoing work. Its announcement describes roles that remember project goals, monitor work between requests, coordinate across sessions, and change course when priorities change. That last promise is the interesting one: who determines that a priority changed, and what happens to the work that loses?

Imagine a small company with one operations Teammate. At 9:00, it is preparing a supplier renewal due Friday. At 9:10, it discovers a customer issue that needs a response today. At 9:15, the owner asks for a sales report before lunch. All three requests are legitimate. A system that can work continuously now has to decide not only how to work, but which obligation to defer.

A chat tool can return the most recent answer and leave the collision to a person. An AI coworker with standing responsibilities cannot pretend that every new message is the only work in progress. AI coworker priority management is the operating discipline that makes competing commitments visible, assigns the right decision authority, and preserves unfinished work when focus changes.

Perplexity's launch is a useful signal of a broader category shift, not evidence that one product has solved every operating question. On the same day, Adobe announced that its CX Enterprise Coworker was available inside several AI platforms, another sign that ongoing work is moving across interfaces rather than living in one chat. Perplexity's setup guidance asks users to describe what a Teammate owns, where work comes from, what it should deliver, and when it should act. Businesses should add another field to that brief: what wins when two valid jobs cannot both finish on time?

A Task List Is Not a Priority System

A queue records work. A priority system records the business reason one item may precede another. If a Teammate sorts a queue by recency, the person who sends the most messages wins. If it sorts by a label called “urgent,” whoever applies that label wins. Neither is a reliable operating rule.

The hard case is not an obvious emergency displacing a trivial task. It is two meaningful obligations with different kinds of cost. A supplier renewal may have a contractual deadline tomorrow. A customer complaint may have no formal deadline but could damage trust if it waits. A sales report may be due at noon but can be partly prepared by a human. An AI coworker needs enough context to identify the tradeoff without appointing itself the company's chief executive.

Perplexity's announcement says Teammates can plan and coordinate ongoing projects, monitor work, and present results for review. That design makes coordination a normal activity rather than a one-off prompt. It also makes priority errors more consequential: a quiet delay is easy to miss when the system appears to be handling everything.

Separate urgency, importance, and authority

Urgency is about time until harm or a deadline. Importance is about the value or consequence of the outcome. Authority is about who is allowed to change the order of work. These are different attributes. An outside email can contain the word “urgent” without having any authority to reorder the team's commitments.

A useful work item records a due time, the source of that due time, the consequence of missing it, its named human owner, the next reversible action, and the latest safe decision point. If two tasks conflict, the Teammate can compare documented facts. Where the tradeoff requires business judgment, it should bring a concise decision to the owner rather than silently guessing.

Treat source text as evidence, not instruction. A customer message saying “drop everything and refund me now” is a customer request to evaluate. It does not automatically override the approval requirement for refunds or the company's existing incident procedure. Priority is a form of authority, so it needs a boundary just as payments and permissions do.

A priority contract beats a vague instruction to be proactive

A priority contract is a short set of rules for a role, not a universal ranking formula. It names protected commitments, escalation triggers, the work that can be preempted, and the person who resolves ties. A narrow example for an operations role might be: safety and service-restoration incidents first; customer commitments due today second; expiring legal or supplier deadlines before optional analysis; internal reporting can be delayed with notice; purchases and external promises always require a human.

The point is not that every company should adopt that exact order. It is that the order should be reviewable before the Teammate uses it. The rules also need expiry. A Friday incident rule should not continue to suppress normal work on Monday because the role remembered a dramatic message but never recorded resolution.

A priority contract should specify a capacity limit: how much active work a role may hold at once. Without it, every new project is “in progress,” and none has a credible completion time. If the queue exceeds that limit, the Teammate must expose the backlog, renegotiate a due date, or hand work to a properly scoped role. Quietly accepting a fourth deadline is not increased capacity.

What Good and Bad Look Like When Work Collides

Bad: A sales Teammate is preparing a quote due at 4:00. A prospect sends an email calling a different opportunity “urgent.” The Teammate shifts its entire afternoon to the newest message, forgets to tell the first account owner, and sends the original quote a day late. Its activity log is full of useful research, but the business commitment was broken.

Good: The sales role records both opportunities, their human owners, the confirmed deadlines, and what can be done without committing to a price. It acknowledges the new inquiry under its approved messaging rules, continues the quote's time-critical preparation, and shows the sales owner a conflict packet: “Existing quote due 4:00; new inquiry requests a call today; I can prepare a briefing now, but cannot meet both deadlines without changing the quote plan. Which commitment takes precedence?” Once the owner decides, the Teammate records the displaced task's new time and tells the affected person.

The difference is not that the good Teammate asks for permission to think. It completes reversible preparation, then asks at the seam where a business promise changes. That is frontier operations in practice: encode routine judgment, preserve human control over consequential tradeoffs, and make the handoff recoverable.

The same pattern applies to an operations Teammate running the three projects from the opening. It can check the customer record, assemble renewal terms, and pull report inputs in parallel if its permissions and capacity allow. It cannot safely claim that all three are on track if the supplier renewal needs a decision by 11:00 and the owner will be unavailable. A useful status report says what is blocked, by whom, and by when—not merely that several tasks are “active.”

There is also a difference between changing a priority and interrupting a person. The role may reorder its own reversible preparation automatically while keeping its human-attention budget intact. It should interrupt an owner only when waiting would forfeit a decision window or a protected commitment. Every alert should contain a decision, not just a declaration that something is urgent.

Build a Seven-Step Priority Contract for a Small Team

A lightweight version can fit on one page and use the systems you already have. Do not start with a sophisticated scoring model; start with one recurring role and the last three times work collided.

1. List standing commitments. Record each ongoing project, owner, deliverable, due time, and external party affected. Distinguish a requested target from a promise already made. A recurring check with no due time still needs a review interval and a named owner.
2. Define protected classes. Identify incidents, legal deadlines, customer promises, and other work that cannot be silently displaced. Keep this list short. If everything is highest priority, nothing is.
3. Specify the evidence required to reprioritize. A due date in the contract, a confirmed service incident, or a direct instruction from the authorized owner can change order. A label in an inbound email alone cannot. Record who or what supplied the evidence.
4. Set a capacity and interruption rule. Decide how many active commitments this role may carry and which conflicts it can resolve itself. State when to ask, whom to ask, and how long to wait. Set a fallback when that person is unavailable rather than treating silence as approval.
5. Checkpoint before switching. Save the current state, source links, outstanding decisions, promised next update, and any uncertain external action. A context summary is not enough if the Teammate cannot tell whether a message was sent or a record was changed. Reconcile the connected system before repeating a write.
6. Announce the displaced commitment. Tell its owner what changed, why, what remains unfinished, and the revised decision time. A priority decision that nobody downstream knows about is still an operational failure.
7. Review one week of collisions. Count missed promises, silent deferrals, unnecessary interruptions, and decisions that a human had to reverse. Refine the rule that caused each failure. Do not reward the Teammate simply for completing more individual tasks if the wrong obligations slipped.

A simple conflict packet has a stable task ID for each job; source and owner; confirmed due time; consequence of delay; current checkpoint; proposed next action; the commitment that would move; and the requested decision with a response deadline. This is not bureaucratic overhead. It is a compact way to keep a human decision from requiring a 40-message history review.

If the business runs multiple roles, apply the rule across handoffs too. An operations Teammate can ask a sales Teammate for a quote, but it should pass the original due date and human owner, not invent a new “urgent” mandate. The receiving role must compare that request with its existing commitments and report a conflict. A multi-role team is not exempt from scheduling tradeoffs merely because more workers are available.

Perplexity's November announcement gives a timely reason to test this now. It describes Teammates that coordinate other Teammates, retain goals and task history, and respond to changing priorities. Those are valuable capabilities, but continuity raises the standard of accountability. A role that remembers yesterday's work must also remember the promise it moved aside today.

This is distinct from recovering missed scheduled runs. A clock can fire reliably while the Teammate still spends the hour on the wrong project. It is also distinct from a task action budget, which limits how far one task may proceed. Priority management determines which legitimate task receives the next unit of limited time, access, and human judgment.

FAQ

Q: What is AI coworker priority management?
A: It is the set of rules and records that lets a persistent AI coworker choose between legitimate competing work without losing a commitment. It covers due times, owners, protected work, decision authority, capacity, and communication when a job is delayed.

Q: Should an AI coworker decide priorities without asking a person?
A: It can reorder low-risk, reversible preparation within an approved contract. It should ask a named owner before moving an external promise, changing financial terms, or accepting a tradeoff the company has not already authorized.

Q: Is a calendar or task board enough?
A: No. A calendar shows when something is due; a task board shows its status. Neither necessarily explains who can change a commitment, what evidence justifies the change, or who must hear about the delay.

Q: How should a small business handle two urgent requests at once?
A: Check whether both deadlines are confirmed, identify the consequence of missing each, complete reversible preparation, and send the owner one conflict packet with an explicit decision time. Do not infer authority from whichever message arrived last.

Q: What if the owner does not respond before the deadline?
A: Follow a predefined fallback: preserve the current protected commitment, do not make an unauthorized promise or payment, notify an alternate owner if designated, and document what could not be completed. Silence is not approval.

Q: How do we know the priority rule works?
A: Review collisions, not just completed-task totals. Track missed promises, silent deferrals, time-to-decision, reversals, and unnecessary human interruptions. Adjust the role's contract when those outcomes show the wrong work won.

If your team is ready to give ongoing work a real owner and a clear decision boundary, Associates AI Teammates provides the platform to configure persistent AI coworkers, their Skills, and their human handoffs. View the platform and choose a plan.