What you spent and what you got

Every piece of AI work in Harmony follows the same chain:

1. Harmony quotes the work before it starts.
2. It charges the work when it finishes.
3. People give the result a verdict.
4. Those verdicts add up to a useful rate.
5. The useful rate turns into value.
6. You read value against your budgets and goals.

Each step is recorded where it happens, so at the end of the month the answer comes from the work itself, not from a survey or a hunch.

This page walks the chain from start to finish and shows who does what along the way. The articles linked from each step go deeper.

The chain at a glance

1. Before work starts: the quote

When you ask Harmony Agent for something, start a Loop, start a background agent, ask for a report or send an agent to work on a task, Harmony quotes it first.

• The quote is the middle value (the median) of the last 20 charges for the same kind of work: your own recent chat turns, or that Loop's recent runs.
• With fewer than 3 earlier charges, a chat turn is quoted at the least that kind of work can cost. A Loop shows no quote beside Run now until it has run 3 times.
• Work on Harmony's models is quoted as a share of your limits, for example "About 3% of your 5-hour limit". Work on your own key is quoted in money.

If the quote won't fit what is left of your 5-hour or weekly limit, Harmony doesn't start the work and tells you so: "Nothing has been charged." Work on your own keys isn't refused for limits.

Loop runs and reports also get a ceiling of four times the quote. A background agent's ceiling is its own spending limit, set when it starts. If work reaches its ceiling, Harmony stops it and says so, rather than letting it run on.

More in Quotes and charges.

2. When work finishes: the charge

Harmony charges work when it finishes. Work that fails or that you cancel doesn't count against your limits.

Under each finished piece of work, the person who asked sees the charge next to the quote, for example "Used 2% of your five-hour limit (quoted 3%)". On your own key it reads like "$0.41 on your OpenAI key".

Work on your own keys carries a 10% Harmony fee. There's no fee in a month where it would come to less than £5 ($6), on local or self-hosted models, or on failed work. The fee shows on the Spend and Models pages, not under each piece of work.

3. The output and its verdict

Five kinds of work produce an output you can judge:

• an answer from Harmony Agent;
• a background agent's result;
• a Loop run;
• a report;
• an agent's run on a task.

When a piece of work finishes, Useful and Not useful buttons appear under it. Tap one and the verdict is saved straight away. After Not useful, Harmony asks what was wrong: Wrong or made up, Missed part of the ask, Didn't need it, Too slow or too costly or Something else. You can skip the question.

You can judge work wherever it shows:

• in a room or a thread;
• in the background agent tray;
• on a Loop run or a run page;
• on an agent's run card in a task;
• in a report preview;
• in the activity panel;
• in Profile → Usage.

Anyone who can see the work can judge it, except a background agent's result, which only the person who started it can judge. Each person has one verdict per piece of work. You can change yours or tap the same button again to clear it, and nobody can change yours.

Some actions count as a verdict without anyone tapping a button:

More in Judging work and the useful rate.

4. When people disagree

Several people may judge the same piece of work. Harmony settles on one outcome per piece of work, in this order:

1. The person who asked for the work decides. Their verdict is the outcome, whatever others say. For a Loop run, that is the person who created the Loop. For a background agent, it is the person who started it.
2. If they haven't judged it, the majority of everyone else decides. A tie leaves the work without an outcome.
3. A verdict from a person always beats one read from an action, such as a report being generated again.

Harmony never averages verdicts and never splits a piece of work into part useful, part not. The other verdicts stay on the work, so you can see who disagreed.

In practice, this means the person who asked for the work is the one to judge it. Everyone else's verdicts matter most for Loops and shared work, where the person who asked may not be watching every run.

5. Harmony's estimate is not a verdict

When a piece of work finishes, Harmony reads the request against the result and makes its own estimate: Estimate: likely useful, Estimate: likely not useful or Estimate: unsure. The estimate shows only until somebody judges the work, and its tooltip says it is not a verdict.

Estimates never count toward the useful rate, cost per useful result or hours returned. When Harmony is unsure about work you asked for, it lists that work in Profile → Usage as "waiting for your verdict", so the work that most needs a person's eye gets one.

On Spend → Value, Global Admins see how often Harmony's estimates matched what people said over the last six months. That figure appears once there are at least 5 results to compare.

6. The useful rate

The useful rate is useful results divided by all judged results. Work nobody has judged isn't in it, and neither are Harmony's estimates.

Because a rate from a handful of verdicts can mislead, Harmony shows it as a range, such as 69–82%: with the verdicts so far, the true rate is very likely inside it. The range narrows as more work is judged.

• A headline rate needs at least 20 judged results, covering at least a quarter of the work. Until then, Harmony says how many more verdicts it needs.
• A single figure in a table, such as one team's Useful column, needs at least 5 judged results.

7. What it was worth

Value takes the useful rate and turns it into two figures.

• Cost per accepted output: what the AI work cost, divided by the useful results.

  • On Spend → Value it is a range, shown in US dollars.
  • In Outcomes it is a single figure, from 5 judged results.
  • The cost is the cost of the model work; the Harmony fee isn't part of it.

• Hours returned (estimate): for each useful result, the hours that work would have taken a person. Harmony uses, in order:

  • the hours an admin set for that Loop;
  • the estimate published with a Marketplace App;
  • the hours set for that kind of work.

  With a Value of an hour set, Harmony also shows roughly what those hours were worth.

Global Admins on Enterprise set the Hours per kind of work, Hours per Loop, Value of an hour and What counts at the bottom of Spend → Value. Every change is written to the audit log and shown as "Last changed … by". Harmony works hours out from the current settings each time it shows them, so a change also updates earlier months.

8. Against budgets and goals

Spend and value mean most next to what a team set out to do.

• Budgets cap what a team or a person spends in a month, on the Scale plan and above. A team that has gone over its budget, or is on course to reach 85% of it, appears in Needs attention in Outcomes. See Budgets and spend.
• Goals are optional targets, such as "Close 40 renewal tasks this quarter". A goal is on track while its progress is no more than 10 points behind the share of the period gone. Work linked to a goal shows on the goal page as Ran alongside: Harmony shows the two side by side and never claims the work caused the result. See Setting goals.

Who does what through a month

Reaching one answer at month end

Harmony doesn't send a month-end report on its own. Teams that want a shared answer usually run the same short review:

1. Clear the backlog. Everyone opens Profile → Usage and judges the work waiting for their verdict. This is the step that moves the useful rate from a wide range to a narrow one.
2. Team leads check their team. On the team's Value tab, look at the useful rate and which Loops and people it comes from. A Loop with a low rate is a candidate to fix or stop.
3. Admins check spend. On Spend, compare each team with its budget and look at Needs attention in Outcomes. Export the CSV if finance needs the figures.
4. The Global Admin reads value. On Spend → Value, read the useful range, the cost per accepted output and the hours returned, and check Where the value came from. If the hours look wrong for a kind of work, change the setting. Every change is logged and applies to earlier months too, so everyone keeps reading the same figures.
5. Share one summary. From Outcomes, use Copy summary as Markdown for each team and paste it into a room. People can discuss it there with the figures in front of them.

Disagreement about a single piece of work is settled by the rules in When people disagree. Disagreement about what work is worth belongs in the value settings, where a change is visible to everyone and recorded.

Example scenario

Northwind's finance team runs a weekly "Revenue recognition check" Loop, created by Lena, a team lead.

• Before each run, the Loop page quotes about 4% of the 5-hour limit.
• After each run, the charge appears under the result, and Lena or her analyst Sam marks it Useful or Not useful. One week Sam taps Useful but Lena taps Not useful with Missed part of the ask, because the run skipped two contracts. Lena created the Loop, so the outcome is Not useful. Sam's verdict stays on the run for anyone who looks.
• At month end, Lena clears the two runs waiting for her verdict in Profile → Usage.
• The CFO, a Global Admin on Enterprise, opens Spend → Value. The Loop is at 85% useful with a narrow range, and its cost per accepted output is well below the team's other work. Hours for the Loop are set to 3, and three of the month's four runs were useful, so it shows about 9 hours returned for the month.
• To share it, the CFO copies the finance team's summary from Outcomes into the leadership room.

FAQs

Does Harmony's estimate change our useful rate? No. Estimates show only until someone judges the work, and they never count toward the useful rate, cost per accepted output or hours.

Two people judged the same work differently. Which counts? The person who asked for the work. If they haven't judged it, the majority of everyone else; a tie leaves it without an outcome.

Why does our useful rate show a range instead of a number? A rate from a few verdicts can mislead. The range shows how sure Harmony can be with the verdicts so far, and it narrows as more work is judged.

We changed the value of an hour. Why did last month change too? Harmony works hours and their value out from the current settings each time it shows them, so every month is read the same way. The change is in the audit log.

Is the Harmony fee part of cost per accepted output? No. Cost per accepted output uses the cost of the model work. The fee shows on the Spend and Models pages.

Does Harmony email us a monthly summary? Not today. Use Copy summary as Markdown in Outcomes and the Spend CSV export.