Agentic AI value tracking per run: chasing is easy, catching is not

Forrester's mid-year verdict on agentic AI is blunt: companies are chasing, few are catching. Three quarters of enterprise leaders report they are adopting agentic AI, yet only a small minority run it meaningfully in production. The gap is not model quality. The gap is that the value of an agent evaporates the moment a run finishes, because nobody wrote it down. Agentic AI value tracking per run is the missing discipline, and it is embarrassingly cheap to add.

The value evaporates when the run finishes

An agent completes a task and the outcome lands where the work was needed: a ticket closed, an invoice extracted, a report drafted. The result is consumed immediately. What does not survive is the record that the run happened, what it replaced, and what that replacement was worth.

Six months later the budget review arrives and the team that shipped the agent has an anecdote pipeline: "support loves it", "it saved the ops team a bunch of time", "we think it handles most of the intake now". Anecdotes lose to line items every time. Forrester's same analysis points at ROI uncertainty as the reason enterprises stay stuck in pilots: they cannot justify production expansion beyond narrow efficiency wins, because nothing was measured while the wins were happening.

The pattern has a predictable ending. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, with unclear business value listed ahead of technical failure as the cause. The projects rarely die because the agent was too dumb. They die because nobody could show the catch.

What agentic AI value tracking per run looks like

The unit that survives an executive review is a run-level record with three fields: what task the agent completed, how long a human would have taken, and what that time is worth. Everything else (latency, token counts, eval scores) matters for engineering, but finance does not buy tokens, it buys hours.

Instrumenting this is one HTTP call at the end of each successful run:

curl -X POST https://humanhours.dev/api/v1/track \
  -H "Authorization: Bearer hh_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "invoice-triage",
    "task_type": "invoice_data_extraction",
    "outcome": "success",
    "human_baseline_minutes": 12
  }'

The baseline is the number people argue about, so settle it before the first run, not at the review. Take the task your agent replaces, time how long a person actually spends on it, and record that number with a one-line note on where it came from. A defensible 12 minutes beats an optimistic 45 every time, because the first question in any review is "says who".

Note what this is not: it is not workflow observability. Tools in the automation stack track execution health, and n8n's Insights documentation is explicit that its time-saved tracking is limited to parent workflows. Useful, but scoped to one tool. Value tracking has to sit above the stack, because your agents do not live in one tool.

From anecdote pipeline to P&L line

Per-run records compound. A team running 5,000 agent executions a month at a 12-minute baseline logs 1,000 human hours a month. At a loaded cost of $55 an hour that is $55,000 of monthly cost saved, itemized per agent and per task type, with a timestamped trail behind every number.

That sentence is the entire difference between chasing and catching. The chasing team says "the agent handles invoice intake now". The catching team says "invoice intake ran 4,820 times last month, saved 964 hours, and here is the month-over-month trend". Same agent. Same work. One of them survives the budget review.

The volumes involved are not exotic. Most teams running agents in production today sit somewhere between 1,000 and 100,000 executions a month across their fleet. At those volumes a per-run tracking call adds no meaningful latency and no meaningful cost, and the aggregate becomes a defensible P&L line within the first month.

There is a second-order effect: per-run value data ends the internal argument about which agents deserve investment. When every agent reports hours and money saved on the same scale, the portfolio ranks itself. The agent nobody talks about that quietly logs 300 hours a month beats the demo darling that logs 12.

Start before the budget review

Forrester notes that more than half of enterprises report agentic sprawl even after adopting governance frameworks like the NIST AI RMF. Sprawl without measurement is the worst position available: many agents, no ranking, no evidence, and a security organization (49% of security decision-makers already name agentic AI as a concern) asking why any of it exists.

The fix does not require a platform migration or a committee. Pick the one agent with the clearest replaced task. Agree the baseline minutes with the team that used to do the work. Add the tracking call. Let it run for 30 days.

At the next review you will be the only team in the room with a number per run instead of a story per quarter. In a year where 40% of these projects are headed for cancellation, that number is the difference between being the cut and making the case.