Open any project tool's pricing page and the word AI appears a dozen times. Underneath, there are only three shapes: an assistant you prompt, an automation you configure, and an agent you delegate to. They look alike in a demo. They behave completely differently at 4pm on a Thursday when a client moves a deadline and nobody is watching the board.
1. The Assistant: Fast, and Entirely Dependent on You
An assistant answers. Summarise this thread, draft this update, what is overdue on the mobile project. It is genuinely useful; a good assistant takes the writing out of status work. But nothing happens unless a person is at the keyboard asking. The coordination still belongs to you. Fastio's reviewers call this the reactive model: useful, but you ask and they answer. Quire is blunter about the chat-in-a-sidebar version, calling it a smarter clipboard, not a teammate.
2. The Automation: Reliable, and Blind
An automation fires a fixed action on a trigger. When the due date passes, move the card and notify the owner. When a PR merges, close the ticket. Automations are the backbone of every mature workspace and Kavanah ships dozens. Their failure is that they cannot judge. The rule runs on the exception exactly as it runs on the norm, and the case that hurts is always the one nobody wrote a rule for: the deadline that moved for a good reason, the owner who is on leave, the ticket that merged but is not actually done.
3. The Agent: Judgment, With Limits
An agent watches the project, decides when something needs attention, and acts in a sequence of steps, within limits a person set. It can tell that the overdue task belongs to someone on leave and reassign it, or notice that three tickets in one project slipped this week and raise that in the standup rather than nudging each owner separately. The failure mode of an agent is not blindness, it is reach: an agent with no scopes, no approvals and no ledger is a very fast new hire with every key in the building. The limits are not a constraint on the product. They are the product.
4. How to Tell Which One You Bought
Three checks. Can it act without a prompt, on a schedule or on an event, and did it do so this week? Can you put an approval in front of a category of action, not a single prompt, and does the agent stop and wait rather than route around it? When it does act, can you read a ledger afterwards and reverse the ones that were wrong? A yes to all three is an agent. Two yeses is an automation with a model attached. One is an assistant. In Kavanah those three are the autonomy level, the action policies and the action ledger with undo, and they are set per AI employee, because triage can run free while a client email should wait for a person.
None of the three is wrong. Assistants and automations are how most teams should start, and a workspace with agents still needs both. The mistake is paying for an agent and getting an assistant, or deploying an agent with an automation's lack of limits.
