Appearance
Prove AI's return with business metrics
When you're done, you can answer "how is AI improving our bottom line?" with your own figures. Flowstate counts an outcome your business cares about, such as level 1 support contacts resolved or insurance claims verified. It shows how much of that work AI did without a person, and what one unit costs when AI does it and when a person does.
Flowstate already puts employees, contractors and AI agents on one bill. This checklist adds the other half: what that bill produced.
An operations lead for the work being measured owns this list. They work with a finance lead, a Flowstate administrator and, if readings come from another system, a developer.
Before you start
- Load your people and teams, with salaries and rates for the people who do the work.
- Set up business metrics: the module, access, and your first metric and source.
- Make sure Flowstate records AI use against the people or agents doing the work, so each metric has an AI cost. Either:
- people's AI sessions, from Track AI in your workflows, or
- AI agents that report their own spend, added under Agents.
Connecting an AI provider's bill isn't enough on its own. A bill that arrives as a single total can't be put against the people doing the work, so it doesn't reach a metric's cost.
1. Choose the outcomes to count
What: Pick one to three outcomes where AI already does part of the work. For each, you need to count both the total and the part done with no person involved. Pick a quality measure to watch beside each, such as a satisfaction score. Who: Operations lead, with the finance lead. Where: Your own reports and systems, such as your contact-centre platform or claims system. Done when: For each outcome you've written down its name, its unit, the team that does the work, the system that counts it, and whether that system can tell which units AI handled alone.
| Outcome | Unit | Handled without a person when |
|---|---|---|
| Level 1 support contacts resolved | contact | The AI assistant closed it with no hand-off to an advisor |
| Insurance claims verified | claim | An AI agent verified it and no claims handler reviewed it |
Tip
Count finished outcomes, not activity. "Contacts resolved" says what the business got. "AI conversations started" doesn't.
2. Define each metric and how AI's share is recorded
What: Create a metric for each outcome, and one for its quality measure. Write the rule for "without a person" in the outcome metric's Description, so everyone who records a reading applies it the same way. Then choose the quality metric as its Satisfaction guardrail. Who: The metric's owner, with manage access. Where: Insights → Business metrics → Metrics → New metric. See Create a metric. Done when: Each metric is listed under the people you chose in Whose cost this is. Its Description says what counts as handled without a person, and its page shows the guardrail under Satisfaction.
- Create the quality metric first. Satisfaction guardrail only lists metrics that already exist.
- Under Whose cost this is, choose the team that does the work, including the people who handle what AI can't. Leave them out and a person's unit looks cheaper than it is.
3. Choose a source
What: Decide how readings arrive for each metric: Entered by hand, or Pushed in through the API from the system that counts the work. Every reading carries the total and, where you measure it, how many were handled without a person. Who: The metric's owner. A developer, if readings are pushed in. Where: Event sourcesDone when: Insights → Business metrics → Event sources shows a source for each metric, with status Active.
- Leave Of which, without a person empty only when you don't know. Empty means not measured. Zero means AI handled none.
- Readings from different sources for the same period are added together, so never send the same work from two sources.
4. Load history
What: Load past periods, ideally going back to before AI took on the work, so you can see the change. The Metrics tab opens on the last 12 months. Who: The metric's owner, or the developer sending readings. Where: The metric's page → Record a reading, or the Business metrics API, which takes many periods in one request. Done when: Readings on each metric's page has a row for every past period you loaded, and Volume over time shows them.
- For periods before AI took on the work, enter zero under Of which, without a person. AI handled none, and that's a real measurement.
- Use the same length of period as the metric's Reported setting.
- Recording a period again corrects it, so you can fix history at any time.
5. Link metrics to the work
What: Make sure each metric's cost reaches the right work:
- Teams, cost centres or geographies link through Whose cost this is.
- AI agents count towards a metric when their work lands on a project the metric's team works on. Put each agent on the project it serves. See Put an agent on a project.
- Projects link on their own. When a project's team, or the team of anyone working on it, is covered by Whose cost this is, the project's Value tab shows the metric under Business metric.
Who: The metric's owner, with the delivery lead. Where: The metric's Edit button, and the project's Resourcing plan and Value tabs. Done when: Each AI agent doing the work is on one of the team's projects, and those projects show the metric under Business metric on their Value tab.
Not available yet
You can't link a business metric to an initiative. Under Whose cost this is, choose the teams that deliver the initiative instead.
6. Read cost per unit and AI share
What: Check that each metric gives figures you can explain. Who: Finance lead, with the operations lead. Where: Insights → Business metrics → Metrics, then the metric's page. See Show what AI returns with business metrics. Done when:
- Each metric shows a Cost per unit and an AI share, not Not measured.
- The figures across the top of its page show a cost per unit by AI and by a person.
- Satisfaction shows the guardrail's latest reading.
Tip
More AI doesn't lower cost on its own. If AI takes more of the work but the team stays the same size, you've released capacity, not saved money. Use Model a change on the metric's page to see the difference.
7. Share it
What: Give the people who make decisions access, and show them where to look. Who: Flowstate administrator, with the finance lead. Where: Settings → Users & Access → Roles, and Eddy. Done when: Your leadership can open Insights → Business metrics, see costs, and get an answer from Eddy to "What does a resolved contact cost us, with and without a person?".
- People who can't see cost figures elsewhere in Flowstate see volumes and AI share, not costs.
- To point someone at a metric, send them the link to its page. It opens for anyone with access.
- Eddy also answers "Which teams produce the most without a person?".
Not available yet
You can't export business metrics from their pages. Ask Eddy for the figures, or read them into Claude or ChatGPT through the MCP server.
You're set up when
- Each outcome you chose has a metric with past readings, including how many were handled without a person.
- The Metrics tab shows a Cost per unit and an AI share for each.
- Each metric has a quality guardrail.
- Your finance lead can say what one unit costs when AI does the work and when a person does, from the metric's page.