AI Project Management Is a Buzzword. Here's What Actually Works.

Elena spent a week evaluating "AI-powered project management tools" for our ops team last quarter. She started with a list of nine. Asana, Monday.com, ClickUp, Notion, Wrike, Smartsheet, Hive, Teamwork, and Height. Every single one of them had AI in their marketing copy. Every single one claimed it would transform how her team worked.

After testing all nine, she sent a one-line Slack message: "They all do the same three things and none of them do the thing I actually need."

She was exaggerating, but not by much. The state of AI in project management tools right now is a mile wide and an inch deep. Every vendor shipped something they could call AI. Almost none of them shipped something that changes how a project manager actually spends their day.

What "AI" Means in Most PM Tools

Here's what Elena found when she dug past the marketing pages and actually used the features.

Asana launched "AI teammates" and smart fields. The AI teammates can generate status updates from task data, draft task descriptions from a brief, and suggest custom fields based on your project structure. Smart fields auto-populate field values based on task content. Elena tested the status update generation and said it was decent. "It reads the tasks and writes a paragraph. It's better than what I'd write at 5 PM on a Friday. But it's a summary of data I can already see. It doesn't tell me anything I don't know."

Monday.com's AI assistant does similar work. It can summarize board activity, generate task descriptions, compose emails from task context, and build formulas from natural language prompts. The formula generation is actually useful -- Elena got it to write a formula that calculated weighted priority scores, which would have taken her time to figure out manually. The summaries, again, were readable but not revelatory.

ClickUp's AI can write task descriptions, summarize comment threads, generate subtask lists from a parent task description, and translate content. The subtask generation was the feature Elena liked most. She gave it a task called "Launch Q2 campaign" with a paragraph of context, and it generated eight subtasks that were about 70% right. She edited them and saved maybe ten minutes.

Notion's AI writes and edits text in docs and database descriptions. It can fill in database properties from page content. Wrike has work intelligence that predicts project risks. Smartsheet has formula suggestions.

The pattern across all nine tools: text generation, summarization, and light prediction. These are LLM features bolted onto existing products. They're not bad. They're just not what Elena was looking for.

What a PM Actually Needs From AI

Elena's actual pain points had nothing to do with writing task descriptions or generating summaries. She can write a task description in 30 seconds. What eats her time is the operational work that happens between tasks.

She needs to know which of her 23 active projects has a dependency that's about to slip. Not a summary of each project. An actual analysis: Project X depends on the API integration finishing by March 15, but that task hasn't been updated in nine days and the assignee is on PTO next week. The built-in AI features can't do this because they operate within a single task or a single project. They don't reason across projects, across assignees' calendars, and across dependency chains.

She needs incoming requests triaged before she sees them. Her team gets about 40 new task requests per week through a form. Each one needs to be categorized (bug, feature request, operational, content), assigned a priority based on what's already in flight, and routed to the right person. None of the AI features she tested can read a task description, compare it to the current workload distribution, and make an assignment decision. They can suggest a category based on keywords. That's not the same thing.

She needs stale work surfaced automatically. Tasks that haven't been touched in two weeks. Projects where the completion rate has flatlined. Assignees who have twelve open tasks and a pattern of missing due dates. This requires reading data across the entire workspace, computing patterns, and making judgment calls about what matters. Summarization doesn't help here. You need analysis.

The Three Things Native AI Does Well

I don't want to dismiss the built-in features entirely. There are three categories where they genuinely save time.

First, drafting. When you need a task description, a project brief, or an update email and you're staring at a blank text box, the AI drafts are a decent starting point. They don't replace your thinking, but they replace the blank page problem. Elena uses Asana's draft feature for weekly status emails to stakeholders. She edits every one, but the first draft saves her about five minutes per email.

Second, formula and field assistance. Monday's natural-language-to-formula feature and ClickUp's formula suggestions help non-technical users build calculations they couldn't write from scratch. This is a real productivity gain for people who know what they want to calculate but not how to express it.

Third, meeting prep. Summarizing task activity and comment threads before a meeting is tedious. AI summaries give you a 30-second overview of what happened on a task since you last looked at it. Not transformative, but useful when you're walking into a meeting about a project you haven't touched in a week.

What Native AI Cannot Do

The gap between these features and what a PM actually needs is the gap between content generation and operational reasoning. Here's the list of things Elena tested that none of the nine tools could handle.

Cross-project dependency tracking. Task A in Project 1 blocks Task B in Project 2. Task A is late. Nobody's been notified. No tool's AI feature caught this because they don't model cross-project dependencies with any intelligence. They can see that Task A is late within its own project. They can't trace the downstream impact.

Proactive triage. A task triage agent reads an incoming request, understands what it's about, checks the current workload of each team member, evaluates priority against existing work, and assigns it to the right person with a reasonable due date. The built-in AI features can categorize the request (sometimes). They can't check workloads, evaluate relative priority, or make an assignment.

Backlog analysis. Which tasks in the backlog are duplicates of existing work? Which ones are obsolete because the project they were created for was cancelled three months ago? Which ones should be escalated because they've been sitting untouched for 60 days? This requires reading every task, comparing it to every other task, and applying judgment. That's an agent-level problem.

Standup automation. Not "summarize what happened yesterday" but "here's what each person is working on, here's what's blocked, here's what's at risk, and here's what changed since the last standup." This requires pulling data across projects, grouping by person, evaluating status and recency, and formatting for a specific audience. Summaries are one ingredient. The whole recipe requires reasoning.

Sprint health analysis. Are we on track to finish this sprint? Not based on a count of done versus not-done, but based on the velocity pattern, the size of remaining tasks, who's assigned to what, and whether anyone is about to go on PTO. Wrike's "risk prediction" gets closest to this, but it's a general risk score, not a specific analysis that a PM can act on.

Where Agents Fill the Gap

The difference between a feature and an agent is that a feature reacts to a single input and produces a single output. An agent pursues a goal across multiple steps, reading data, making decisions, and taking actions along the way.

When Elena deployed a task triage agent, it didn't just classify incoming requests. It searched the existing backlog for similar or duplicate tasks. It checked each potential assignee's current task count and upcoming deadlines. It evaluated the request against the team's stated priorities for the quarter. Then it assigned the task, set a due date, added a comment explaining the routing decision, and flagged anything it wasn't confident about for Elena to review manually.

That's not a feature you can ship as a button in a PM tool. It's a workflow that requires reading from multiple data sources, applying context-dependent logic, and taking a sequence of actions where each step depends on what the previous step found.

The same pattern applies to every one of Elena's pain points. Cross-project dependency tracking requires reading all projects, building a dependency graph, evaluating timelines, and alerting on risks. Standup automation requires aggregating per-person data, evaluating task freshness, and formatting output. Backlog cleanup requires comparing every task to every other task and making keep/archive/escalate decisions.

These are operational workflows. They require operational reasoning. A PM tool with a "summarize" button is answering a different question than what PMs are actually asking.

The Honest State of Things

Here's where AI in project management actually stands right now, without the hype.

Built-in AI features are convenience tools. They save small amounts of time on content generation and formula building. They don't change workflows or eliminate manual operational work. If you're evaluating PM tools and AI features are a factor, weight them at about 5% of your decision. The underlying project management model (how the tool handles tasks, projects, dependencies, custom fields, and views) matters far more.

AI agents that operate on top of PM tools are operational tools. They handle the work that falls between tasks: triage, analysis, reporting, cleanup. They're not built into any PM tool natively because they require cross-system reasoning and goal-directed behavior that feature-level AI doesn't provide.

The AI in your PM tool will help you write faster. An AI agent will help you manage faster. Those are different problems, and conflating them is how the industry got into the buzzword situation Elena found when she evaluated nine tools and came back disappointed.

If you're a PM evaluating tools right now, pick the tool with the best project management model for your team. Ignore the AI marketing. Then add agents for the operational work that no tool handles natively. That's the combination that actually changes how you spend your day.

Try These Agents

• Asana Task Triage Agent -- Automatically categorize, prioritize, and route incoming tasks based on workload and team priorities
• Asana Cross-Project Dependency Tracker -- Surface dependency risks across multiple projects before they cause delays
• Asana Backlog Cleanup Agent -- Find duplicates, stale tasks, and obsolete items across your backlog