Adobe Workfront can assign tasks to AI. Watch who gets assigned first.

Adobe calls the feature Task Collaborators. A Workfront administrator connects an agent created in Claude, Microsoft Copilot Studio or Writer, gives the collaborator an access level and a description, then makes it available for assignment to project tasks. The agent can receive task context, do the requested work and post its result back to the Workfront update stream.

The timing needs a correction because launch coverage has described AI Collaborators as generally available. Adobe’s release notes are narrower for Task Collaborators: preview began July 31, production fast release began August 13, and production for everyone is scheduled for October 15, 2026. Adobe’s user documentation still labels the feature as not yet generally available.

That means a team may read about the feature today and still not find it in its Workfront environment. Check the organization’s release track before opening a support ticket or telling employees the new assignee is ready.

The feature also has practical boundaries. Adobe says Task Collaborators require a Standard Workfront license on Select, Prime or Ultimate packages, and setup requires system-administrator access. They can currently be assigned to tasks, not issues, and they cannot review or approve a document. Adobe has a separate Reviewer-type AI Collaborator for brand and asset review.

A person scanning a project board sees an assignee and reasonably assumes the work has an owner. Task Collaborators break that assumption in several specific ways.

Adobe says the agent starts when it is assigned to a task that is ready to begin, when it is the only or primary assignee as a blocked task becomes ready, or when it was assigned before a human teammate. If a person is already assigned and the agent is added afterward, the agent does not start. Mentioning it with an @mention does not start it either.

Two AI Collaborators on one task do not form a team. The first one may already be working; the second one does nothing. Predecessor tasks matter too. If they are incomplete, the collaborator waits until the task can start—and even then its primary status and assignment order still matter.

None of this is necessarily bad. Automatic starts need predictable rules. The problem is presentation: ‘assigned’ and ‘working’ are different states, while a normal project board trains people to treat them as the same. A manager can look at a staffed plan, assume the task is moving and discover the no-op at the deadline.

Workfront is not where the whole agent is built. The model, instructions, skills and tools live in Claude, Copilot Studio or Writer. Workfront supplies the project task, its context, an access level and the allowed after-work actions.

That split gives teams choice, but it also splits troubleshooting. A silent task might be blocked by an unfinished predecessor, a primary-assignee mistake, insufficient Workfront access, an unpublished external agent, exhausted AI credits or provider configuration. Adobe’s own troubleshooting page sends admins back to both Workfront and the agent provider.

Write that split into the support plan before launch. The project manager should know which task state to check. The Workfront admin should own collaborator access and triggers. Whoever owns the external agent should own its instructions, publication state, credentials and usage balance. ‘Ask the AI person’ is not a support path.

There is also a cost boundary. Adobe says AI Collaborators are included for Workfront customers, while the connected third-party agent may create separate charges. A task can therefore be free on the project-management side and still consume paid model or platform usage elsewhere.

Adobe’s product page emphasizes that actions are logged and auditable. Computerworld reports that generated work returns through Workfront’s update stream so people can review, adapt and approve it. That is better than losing the result in somebody’s private chat.

The review still takes time. If an AI Collaborator drafts campaign copy, creates image variations or asks for missing event details, someone has to decide whether the result is usable. A task marked complete by the agent can leave fact checking, brand judgment, source checking and revision sitting in a person’s day without a due date or workload allocation.

For the first pilot, add the review as real project work. Name the reviewer. Give the check its own estimate. Record whether the result was accepted, lightly edited, rebuilt or abandoned. If a human and AI share the same task, make it obvious which one owns production and which one owns the decision.

This protects the team from a common accounting trick: the board says the AI finished ten tasks while one person spends Friday repairing six of them. The project did not gain ten completed tasks. It gained four finished items and six review queues.

Choose one narrow job with a visible output: draft ten event descriptions from complete briefs, create image variations from approved assets, or identify missing fields in ten project requests. Do not begin with customer publishing, budget changes or a broad campaign role.

Set up the same assignment pattern every time. Make the Task Collaborator primary, keep the person’s review step separate and use tasks whose predecessors are already complete. This removes the start-rule ambiguity while the team learns whether the work itself is useful.

For each task, record five things: whether the agent started without help, whether it returned an output, reviewer minutes, correction type and whether the result moved the project forward. Also record the no-start cases and the minutes spent finding out why. Silent debugging belongs in the cost of the pilot.

Compare those ten tasks with ten recent examples done the old way. The rollout earns another round when total human time falls, usable work reaches the next person sooner and nobody has to monitor the assignee list to make sure the machine woke up. If the board looks busier but the reviewer’s week does too, stop expanding and fix the job definition.

Jun Vega would put the start rule directly beside the AI assignee: waiting on predecessor, not started because human is primary, running, or output ready. The person looking at the ordinary task should not have to know Adobe’s trigger table to tell whether work is moving.

Priya Rao would grade the pilot per ten eligible tasks: starts, outputs returned, usable results, review minutes, correction minutes and no-start debugging. Counting AI completions alone gives the machine credit while hiding the person who made the result safe to use.

Jun is fixing the false signal on the screen. Priya is fixing the false result in the rollout report. Teams need both before an AI name in the assignee field can be treated like capacity.

Workfront’s approach solves a real problem. Teams already have project briefs, dependencies, due dates, permissions and review history in the project system. Sending that context to an approved agent is cleaner than asking every employee to rebuild it in a private chat.

The familiar assignee metaphor also creates the sharpest risk. A name on a task carries expectations: it started, it knows what done means, and somebody will notice if it gets stuck. Adobe’s current trigger rules make those expectations conditional.

Treat the AI Collaborator as a new kind of project resource with explicit start, output and review states. Keep the first job narrow. Put the human judgment on the schedule. Then the board can show less busywork instead of one more name everyone learns to babysit.