Run a one-week AI credit test

Choose one job the AI would do repeatedly, such as summarizing sales calls. One job keeps the number clean. If you test five things at once you cannot tell which one used the credits.

Give the job to one person for five working days. Pick someone who will actually log the numbers. A careful person on a typical week beats an enthusiast on a busy one.

Every day the tester writes down two numbers. Copy this table into a spreadsheet.

• Runs. How many times they ran the job that day.
• Credits used. Read from the tool's usage page at the end of the day. Most tools show credits on a usage page. If yours does not, ask the tool's support where to find it.

Write down what counts as one run before you start, so every day is measured the same way. For example: "one run is one call summary, from pasting the transcript to getting the answer back." If the tester asks a follow-up question, decide now whether that is part of the same run or a new one. Either choice works. Switching halfway does not.

If your usage page shows a running total, read it twice a day if you can, once before the first run and once after the last. The difference is that day's credits. If you only check once, write down the number you saw yesterday, or the subtraction will be wrong.

Common mistakes

• Counting runs from memory on Friday. People forget, and they forget the cheap runs first. Fix: fill in the row at the end of each day, even if it takes one minute.
• Letting the tester skip the odd runs. A failed run, a retry or a long document still uses credits. Fix: log every run, and note the odd ones in the Notes column.
• Testing on a quiet week. A holiday week or a slow week shrinks the number. Fix: start on a normal week, or add a margin to the result.
• Using the plan page as the tester's number. The plan page tells you what you can spend. Only the usage page tells you what you did spend. Fix: log from the usage page, then compare.

After day five, add up the week.

The test is not finished until you do this step. A week of logs with no math gives you no answer.

1. 1

   Credits per run = total credits used ÷ total runs.

2. 2

   Credits per day = credits per run × runs per day you expect for ONE person.

3. 3

   Monthly credits for this job = credits per day × number of people × 21 working days.

That is about 21 working days in an average month. Use your own number if your team works a different schedule.

The result is the monthly number for that one job. If you plan to roll out several jobs, repeat the test for each one.

A worked example

Picture a 6-person bookkeeping firm. The owner wants the AI to summarize client calls. She picks one careful team member, Sam, and defines a run as one call summary.

By Friday, Sam's log shows 30 runs and 90 credits in total. Monday looked cheap and Wednesday looked expensive, because Wednesday's calls were long. The Notes column says so.

1. 1

   Credits per run: 90 ÷ 30 = 3.

2. 2

   Sam did about 6 summaries a day in the test and expects the same once this is normal work. Credits per day: 3 × 6 = 18.

3. 3

   Four people will use it. Monthly credits: 18 × 4 × 21 = 1,512.

She opens her plan page. The monthly allowance is well below 1,512. So the answer is over. She now has three real choices instead of a surprise bill. She decides to roll it out to two people first, which cuts the monthly number to 756, then recheck the usage page after a month.

She pastes the same numbers into the math check below, and Claude shows each step so she can confirm the 1,512. The test gave her the size of the decision before the bill arrived.

Open your plan's page and find how many credits it includes each month. Then compare. You can paste your numbers into Claude to check the math.

• Under your plan. Roll it out. Keep checking the usage page for the first month.
• Over your plan. You have three choices. Buy more credits, run it less often, or give it to fewer people.

Either way, you decide before the bill does.

Before you pick, look at the three levers one at a time. Running less often is the cheapest to try, because it needs no new purchase. Giving the job to fewer people also works, and it lets you start small and widen later. Trimming the job, for example by sending shorter documents, can lower credits per run without changing who uses it. Change one lever, then run a shorter test again, so you know which change made the difference.

If you are under your plan, leave a margin. Real work is messier than a test week, and people find new uses once the tool is in front of them.

If the numbers still do not make sense after a day or two, stop and ask the tool's support how one run is counted. The math only works when you know what a credit pays for. Do not keep logging numbers you cannot explain.