Gartner's June 2025 forecast that more than forty percent of agentic AI projects will be cancelled by the end of 2027 names escalating costs, unclear business value and inadequate risk controls as the reasons, and adds that many vendors are 'agent washing': rebranding assistants, RPA and chatbots as agents. Unclear is the operative word. Most teams that deploy an agent never decide what it was supposed to change, so six months later nobody can say whether it did. These six metrics are what we watch in Kavanah, and five of them come straight from the workspace's own data.
1. 1. Status Meetings Cancelled
The crudest and the most honest. Count the recurring meetings whose only purpose was to find out what was happening, and count how many you have removed since the agent started writing the standup and the weekly report. If the number is zero after a month, the reports are not good enough to replace the meeting yet, and that is the thing to fix first.
2. 2. Time From Request to Owner
How long a new request sits before it is in the right project with a named owner and a due date. Before an agent, this is hours to days and depends on who is watching the intake. With triage on a schedule it should be minutes. Kavanah records when a request arrived and when it was assigned, so this is a report, not a survey.
3. 3. Actions Taken Versus Actions Undone
The action ledger records every agent action and every undo. The ratio is your trust metric. A rising undo rate means a scope widened too fast or a Negation is missing; a falling one means you can move a category from approve to allow. A ledger with many actions and few undos over a month is what an agent that has earned level four looks like.
4. 4. Approvals Rubber-Stamped
Of the actions that parked for approval, how many went out unedited? If nearly all of them did, the approval is theatre and the category is ready to run on its own. If many were edited or rejected, the gate is earning its keep and the draft quality is the thing to work on. Either answer is useful; not looking is not.
5. 5. Estimate Accuracy
An agent that assigns and plans is only as good as the estimates it plans with. Kavanah calibrates estimates against actuals per person and per kind of work, and the planning layer uplifts forecasts without ever rewriting the stored estimate. Watch whether the calibration ratio is moving toward one. If plans keep slipping by the same margin, the agent is faithfully planning on numbers that are wrong.
6. 6. Credits Per Outcome
AI usage in Kavanah is metered in credits, which makes the cost side measurable in a way a flat add-on never is. Divide the month's credits by the outcomes that matter: reports delivered, requests triaged, sprints planned. A number that rises while outcomes stay flat is usually a scheduled run reading more of the workspace than it needs, and the fix is a narrower scope, not a cheaper model.
Pick three of the six before the agent starts, write down today's number, and look again in thirty days. The teams that get cancelled are the ones that could not answer the question; the ones that keep the budget are the ones that measured.
