Buying AI project management software is harder than it looks, because "AI" now covers three different things sold under one name. Gartner calls assistants the precursor to agentic AI, notes they depend on human input and do not operate independently, and labels calling them agents agentwashing. It also estimates that only about 130 of the thousands of agentic AI vendors are real. This guide turns that into a buying process: decide the job, sort the features into kinds, ask eight questions, and watch a fifteen-minute demo that answers them.
1. Start With the Job, Not the Feature List
Write down the three coordination jobs that eat your week: the Monday status report, chasing overdue work, triaging new requests, turning meeting notes into tasks. Then ask of each whether you want it done when you ask, done automatically on a rule, or done on its own toward a goal. That answer decides the category you are buying, and stops you paying agent prices for an assistant, or expecting an assistant to run your Monday without you.
2. Sort Every Feature Into One of Three Kinds
An assistant answers when asked. Automation follows what Anthropic calls predefined code paths. An agent is a system in which the model directs its own process and tool use. Ask the vendor to label each AI feature this way. Most suites sell all three: Asana pairs AI Studio automations with prebuilt AI Teammates; monday.com has an agent builder, an assistant and AI workflows; Notion has Notion Agent and Custom Agents that run on a schedule.
3. The Eight Questions
One: which kind is each feature? Two: does it work on the real tasks, or on a copy in a chat? Three: what can it reach, and whose permissions does a scheduled run use when nobody is signed in? Four: which categories of action wait for a person? Five: is there a record, and can a mistaken change be undone? Six: what runs without being asked, and where does the output land? Seven: what does one scheduled run cost, and can unattended spend be capped? Eight: can you choose the model? Three and four come straight from OWASP's mitigations for Excessive Agency: limit permissions to the minimum necessary, and have a human approve high-impact actions before they are taken.
4. The Fifteen-Minute Demo
Do not watch the vendor's script. Ask them to show four screens. The setting that holds an outbound email or a bulk reassignment until a person approves it; monday.com, for example, describes a simulation mode before an agent goes live. The log of what the AI did yesterday. The undo for one of those actions. And the usage page that shows what yesterday cost. A vendor that can show all four in fifteen minutes has built for the job; one that shows how clever the model is has built a demo.
5. Answers That Should End the Call
"The agent uses an admin account." "There is no way to stop a category of action, only to turn the agent off." "We log prompts, not actions." "Pricing for agents is custom." Each one means you would carry the risk or the bill blind. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls, and every one of those answers is a step toward that.
6. Run a Two-Week Trial on One Job
Pick one job from your list, give the AI the narrowest permissions that do it, hold every external action for approval, and read the log daily. In Kavanah that means one AI employee, scoped to tasks and reports, with action policies holding outbound messages, in safe mode if you want risky tools removed entirely, and a daily token ceiling on unattended runs. Widen only what the log shows was safe. Anthropic's own advice for agents is the same shape: the simplest solution possible, increasing complexity only when needed.
Decide the job, sort the kinds, ask the eight questions and watch the four screens. For how Kavanah, Asana, ClickUp, monday.com, Wrike and Notion answer the questions from their own pages, see our AI project management software page.
