Score agent workflows by hours returned per dollar

Usage and a satisfying answer are not evidence that an AI workflow paid off. Name the outcome before you spend, either revenue or time returned, and count human minutes, because agent runtime is not human labor time. Jiaona Zhang's contrast is a deck font you could fix with one click, versus a task that took humans five hours a week and an agent finished in five minutes.

Jiaona Zhang is Chief Product Officer at Laurel. This conversation aired in June 2026.

How do you tell a useful agent run from token maxing?

Treat hours returned per dollar as a partial efficiency metric: human minutes before, human minutes after including review and correction, and the token cost for that workflow. Evaluate revenue outcomes separately against the relevant costs. A rise in usage, or a reply that felt good, does not show that the workflow returned time or revenue.

Zhang calls the failure mode token maxing. It follows an order that everyone must use AI, sometimes with that use written into the performance review. People then spend tokens to show they complied. A font change that a button already handles is her picture of spend with no time returned. A workflow earns a review only when the task, the minutes before, the minutes after, and the cost are filled in. Blank fields mean you are still experimenting.

If you use AI to go like redo the font on this deck when you could have just clicked a button, it is not an efficient use of your tokens. But if you used AI and you took something that took humans five hours per week, an agent did in five minutes, that is Time back.

Jiaona Zhang, Chief Product Officer at Laurel, on Chain of Thought ep 64

Stop Token Maxxing: Find Where… - Chain of Thought | AI Agents, Infrastructure & Engineering - Apple Podcasts

Podcast Episode · Chain of Thought | AI Agents, Infrastructure & Engineering · June 25 · 58m

podcasts.apple.com

What should you write down before the workflow spends tokens?

Write one line that says whether the workflow should move revenue or return time to a named role. Teams often skip that sentence, so a session that felt productive and a counted gain look the same.

Zhang frames the gap as visibility. People take yesterday's task and try to run it through a model, before naming the time they can get back or the revenue they can drive. Few then quantify the result. Many organizations also skip a regular look at token spend, so the bill arrives while the mandate is still use AI everywhere.

Are you trying to drive revenue? Are you trying to drive time allocation efficiency? What is it that you're really trying to drive?

Jiaona Zhang, Chief Product Officer at Laurel, on Chain of Thought ep 64

Keep one row per workflow:

1. Outcome: revenue, or time returned to a named role.
2. Captain: the person accountable for that outcome.
3. Human minutes before, from the people who did the task.
4. Human minutes after, on the same task, including review and correction. Agent runtime does not count.
5. Token cost sliced to that workflow, over the same period as the minutes.

A minimal sketch to adapt, not a drop-in library.

def hours_returned_per_dollar(task_logs):
    # Partial efficiency metric for time returned, per week.
    # Revenue outcomes are evaluated separately against relevant costs.
    report = []
    for task in task_logs:
        # Agent runtime is not human labor time, so it is not counted here.
        human_after = (
            task["hands_on_minutes_after"]
            + task["review_minutes"]
            + task["correction_minutes"]
        )
        minutes_saved = task["human_minutes_before"] - human_after
        hours = minutes_saved / 60
        cost = task["weekly_token_cost_usd"]
        report.append({
            "workflow": task["workflow"],
            "hours_returned": hours,
            "hours_per_dollar": (hours / cost) if cost else None,
        })
    return report


# Hypothetical dollar amount and review minutes, not measurements.
token_cost_for_workflow = 12.0

# Five human hours a week, from Zhang's example. The agent's five minutes
# are runtime, so the human minutes after are only hands-on, review and fixes.
example_logs = [
    {
        "workflow": "named_weekly_task",
        "human_minutes_before": 5 * 60,
        "agent_runtime_minutes": 5,
        "hands_on_minutes_after": 0,
        "review_minutes": 10,
        "correction_minutes": 5,
        "weekly_token_cost_usd": token_cost_for_workflow,
    }
]

hours_returned_per_dollar(example_logs)

How do you keep one person accountable for an agent fleet?

Put one person on the hook for the outcome, and treat the agents as an extension of that person's work. Zhang's picture was a group of 10 humans on a job, then five humans, each with hundreds of agents. The owner is still one person.

At Laurel that person is a captain. The work is orchestration and delegation: which piece an agent runs, how agents hand work to each other, and where a human still decides. People who have managed before often pick this up faster, because they already had to get a result without doing every piece themselves. A strong individual contributor is used to doing the thing. A captain stays accountable and still ships and audits. Delegation is the shift from individual contributor to manager, and she says you cannot be hands off.

You still supply context and clarity. Motivation stays with people, because an agent does not bring a colleague's interpersonal dynamics. If the outcome slips, start with the captain.

The reason why we love the term captains and we use it very often here at Laurel is because the idea of a captain is you know, that accountability, right? Like, there is one throat to choke, one person responsible.

Jiaona Zhang, Chief Product Officer at Laurel, on Chain of Thought ep 64

How do you get a team to adopt the workflow?

Start with one workflow a person already spends a lot of time on. Zhang's examples were timekeeping, entering data in a CRM, moving pixels, and unit testing. Sit with that person and build the replacement with them, on the bottoms-up path she describes.

Pair it with a top-down pass she calls an ontology of the work. Write the buckets that make up what the function should do to drive those outcomes. Mark which chunks you will automate and which still need a person's judgment. She pointed at law, accounting, consulting, and private equity, fields whose steps sat still for a long time, where people are now splitting a mechanical step from a step that still needs a person.

FAQ

What is token maxing?

It is Zhang's name for using AI to prove you used it. The pattern shows up when leaders require AI use and put that use in a performance review. Her contrast is a one-click font fix against a five-hour weekly task an agent finishes in five minutes.

How do I start if agents are already in production and nothing is measured?

Write the outcome for one workflow, revenue or time returned. Add minutes before, minutes after, and token cost, then compute hours returned per dollar. A usage total will not tell you which workflow to keep.

Who should own an agent workflow?

One captain owns the outcome. Zhang treats agents as an extension of that person, so the human you ask when the result is wrong is still that captain.

Takeaway

• Write the outcome, revenue or time returned, before you fund a workflow with tokens.
• Log minutes before, minutes after, and token cost on one row per workflow.
• Report hours returned per dollar, and leave out rows you cannot fill in.
• Name one captain who is accountable for the outcome the agents support.
• Start adoption on a single time-heavy task, and write the buckets of that function.

The full conversation, with the transcript, is on Chain of Thought.

Subscribe to the Chain of Thought newsletter for new episodes and write-ups like this one.

Drafted with AI assistance from the episode transcripts.