The short answer.
You tell whether AI is paying back by measuring one whole piece of work against your own baseline, written down before the workflow’s first run. Pick one workflow, record how that work goes today (how long from start to finished, how often it comes back with a mistake, how long it waits), then record the same things once the workflow is in daily use, including the time a person spends checking. If checking eats what preparation saves, the numbers show it, and you fix the workflow or stop it.
Hours of assistant use, active seats and a survey of how people feel measure activity, not the work. The ambition is to 10X the ROI of your existing AI, and the proof is your own before and after on named work.
Ask three questions, in order.
To know whether an AI rollout is working, ask whether each workflow is installed, then whether it is used, then whether it is useful, and do not skip ahead.
- Is it installed?The workflow is set up in your company’s ChatGPT, Claude, Copilot or Gemini, the people who run it have access, and the owner has run it once on real work.
- Is it used?The people who own the work run the workflow every time that work comes up, without being reminded. Count the runs against the number of times the work happened.
- Is it useful?The work is done faster, comes back with fewer mistakes, or waits less than it did before, with the checking counted. Only this question needs the baseline.
A workflow that is not used cannot be useful. Installed, used or useful? has the full sheet for all three questions.
Write the baseline before the first run.
A baseline is a short written record of how one piece of work goes today, taken before the workflow touches it. Once the workflow is in use, nobody remembers how long the work used to take.
Put four things in the baseline, measured on the last ten to twenty times the work was done.
- Time from start to finished.From the moment the work arrives to the moment it leaves the team, in hours or days.
- Hands-on time.The minutes a person actually spends on the work, including gathering what it starts from.
- How often it comes back.The share of cases returned with a mistake, a missing item or a question from the next team or the client.
- How long it waits.The time the work sits between steps, waiting for a person, a record or an approval.
Write down who measured the baseline and how: system timestamps, a week of people noting start and finish, or the team lead’s count of returns. What is an AI workflow worth? shows how to turn the same baseline into a value for your company.
Count the checking honestly.
The time a person spends checking a workflow’s output is part of the work, and it goes into the after measurement in full. A workflow prepares; a person reviews and decides. If checking adds more time than preparation takes off, the workflow costs time, and the before and after shows that plainly.
Three habits keep the checking count honest.
- Time the review separately.The reviewer notes how long the check took on each run for the first month.
- Log every correction.Each time the reviewer changes the output, they note what changed. A correction that repeats is a fault in the workflow, and the owner fixes it once instead of the reviewer fixing it every week.
- Count the rework downstream.A mistake the next team caught still counts as work that came back.
Decide what the time that comes back is for.
Time a workflow saves pays back only when it goes to named work, decided before the workflow is in daily use. Write one line on the baseline sheet: what the person who runs this work will do with the time it frees, such as calling clients the same day, clearing the backlog or taking on the next workflow. What will your team do with the capacity it gains? sets out how to decide it.
When to expand, adjust or stop.
Decide after the workflow has been in daily use for four to six weeks and run on enough cases to compare with the baseline. The workflow’s owner decides with the team lead.
- Expand.The work is faster, returns less or waits less than the baseline, with checking counted. Roll it out to the rest of the team, or start the next workflow.
- Adjust.The work is better in one measure and worse in another, or checking takes longer than it should. Fix the repeated corrections, retest, and measure another four weeks.
- Stop.After one round of adjustment, checking still eats what preparation saves, or the team does not use the workflow. Stop it, record why, and move the effort to the next candidate. A stopped workflow with a written reason is a result.
A 90-day priority gives this decision a natural date. Make an AI pilot your next quarterly priority shows how to set one up with the baseline at the start and the decision at the end.
A one-page baseline sheet to copy.
Copy these headings into a document, one sheet per workflow. Fill in the first three before the workflow’s first run.
- The work.The one piece of work, what it starts from, what it produces, the team that does it and the person who decides.
- Measured before.Dates and number of cases. Time from start to finished. Hands-on time. How often it comes back. How long it waits.
- What the time is for.The named work the freed time goes to.
- Measured after.Dates and number of cases. The same four measures, plus checking time per case and the corrections logged.
- Who measured.The person, by name, and how: system timestamps, a tally or a log.
- The decision.Expand, adjust or stop. The date, who decided and why, in one sentence.
Illustrative example.
Take a 60-person distributor. Its customer service team prepares a written answer to every order query that involves a delivery problem. The head of customer service owns the workflow and decides every answer before it is sent. Before the first run, she pulls the last fifteen queries from the inbox and records when each arrived, when the answer went out, how many came back with a follow-up question and how long each sat waiting for a stock check from the warehouse. She writes on the sheet that the time freed goes to calling the five largest accounts each week. After five weeks of daily use she measures the next fifteen queries the same way, with each reviewer noting their checking time. The answers go out sooner, but the warehouse wait has not moved and checking is slow on part orders. She decides to adjust: the workflow gets a step that asks the warehouse for stock in the first message, and the part-order instructions are tightened. She measures again in four weeks.
What is different in a company of under 200 people.
A company of under 200 people measures return on the work itself, because it has no analytics team and needs none. Most published advice on AI return is written for large enterprises, with usage dashboards across thousands of seats. You do not need that.
- The person who measures is the person who owns the work.A team lead with a sheet and a week of notes is enough for a baseline.
- Ten to twenty cases is a real sample for you.You are deciding about one workflow in one team, and the people who do the work can tell you why each case went the way it did.
- The decision is quick.The owner, the team lead and whoever owns the plan can decide to expand, adjust or stop in one conversation.
Measure the work, then plan the next one.
Every workflow on your plan should carry its baseline and its decision. FullyPowered Planner is where the plan lives: you set out your teams, their roles and the work between them, and choose what to build first, so the next workflow is ready when the first one earns its place. Measuring is part of step three, deploy them, in How organizations plan their use of AI.