Home  ›  Guides  ›  Managing AI tokens and consumption spend  ›  Run a monthly usage review: what to check and who approves increases

Run a monthly usage review: what to check and who approves increases

Usage-based software can cost a different amount every month. A short monthly review shows where usage is growing and why. It also gives one named person the decision on any increase.

You need the vendor's usage report, your forecast and the owner of each automated process. All figures in this article are illustrative. Prices are before tax and in no particular currency.

Set up the review

Hold the review in the first week of each month, after the vendor's usage period closes. Invite three roles:

Agree in advance who approves each kind of change, and keep the record with the contract.

ChangeWho approves
Usage stays within the allowanceNo approval needed. The process owner notes it in the review.
Change an alert thresholdPlatform admin, recorded in the review
Raise a hard limit or buy extra unitsNamed budget owner
Turn on a new feature that uses unitsNamed budget owner and the process owner
Move to a larger allowance or planNamed budget owner, through your normal purchasing process

Name a deputy for the budget owner. A hard limit can stop a process while the approver is on leave.

Decide who receives each alert

In the example, the plan includes 20,000 credits a month for $700. Overage costs $0.04 per credit. Alerts fire at 75% (15,000 credits) and 90% (18,000 credits). A hard limit stops usage at 22,000 credits.

For each alert, agree who receives it, how fast someone acknowledges it and who it goes to next. The table shows an example set.

AlertSent toAcknowledge withinIf nobody acknowledges
75% of allowancePlatform adminOne working dayDeputy admin
90% of allowancePlatform admin and budget ownerFour working hoursDeputy budget owner
Hard limit reachedPlatform admin, budget owner and support managerOne hourDeputy budget owner

Send alerts to a shared mailbox or channel as well as to named people. The alert then still reaches someone when a recipient is away.

Collect the numbers

Gather three sets of figures for the previous month from the usage report, which shows more than the invoice total.

  1. Usage. Units used by each process.
  2. Volume. How many times each process ran, such as emails answered or records enriched.
  3. Results. What the process achieved, such as emails closed without a staff reply.

Confirm that the report shows usage by process. Some reports show totals by feature only. If two processes use the same feature, you cannot split its total without a tag or a separate setting. Ask the vendor, or estimate the split from volume and mark it as an estimate.

Compare expected and actual usage

Use the forecast for each process. Compare actual usage with the expected operating usage, before the buffer. October was a normal month in the example. Expected operating usage was 14,800 credits, and the buffered budget was 16,280.

ProcessExpected operating usageActual creditsVarianceVariance %
Support email agent10,00012,600+2,600+26%
Call summaries3,2003,040-160-5%
Contact enrichment1,6002,880+1,280+80%
Total14,80018,520+3,720+25%

Total usage reached 18,520 credits, about 93% of the allowance. That is 2,240 credits, or about 14%, above the buffered budget. Both alerts fired, and usage stayed below the hard limit.

Find the cause of each variance

Divide units used by volume to get units per action. Then compare both numbers with the forecast.

When units per action rise, look at message length, extra steps, retries and test runs. When volume rises, ask the process owner what changed in the business.

Counts show the scale of a change. Investigate causes in the source system, and keep customer content out of the review record. If you need to look at one case, open it in the platform with the right access.

Check results and quality against usage

Record one result for each process, to show what the usage achieved.

In the example, the agent closed 1,890 of 2,520 emails without a staff reply. That is 75%. Each closed email cost about 6.7 credits, because the agent also used credits on emails it passed to staff.

A closed email meets the vendor's definition of a resolution. It does not show whether the customer got a useful answer. Add a quality measure, such as one of these:

Run the sampled review inside the support platform. Record only the scores and counts. If the vendor bills per resolution, ask how it treats cases that reopen.

For enrichment, check how many event contacts became useful leads. If few did, enrich contacts only after someone qualifies them.

Prepare a request before any limit increase

October usage stayed below the 20,000-credit allowance and the 22,000-credit hard limit. A change to the limit needs a forecast for the coming months.

In the example, November and December are the seasonal peak. The support manager expects the launch to add 500 emails a month to the peak forecast. The marketing manager agrees to enrich only qualified contacts.

Process (each month)VolumeCredits eachCredits
Support emails3,500517,500
Call summaries50084,000
Contact enrichment2,00012,000
Expected operating usage23,500
Buffer (10%)2,350
Buffered budget25,850

Both figures are above the 22,000-credit hard limit. At the current limit, work would pause before the end of each month.

At the current limit, overage cannot pass 2,000 credits, which is $80.00 a month. A 26,000-credit limit allows up to 6,000 credits, which is $240.00 a month. The higher limit can add up to $160.00 a month, or $320.00 over the two months.

List the alternatives beside the cost:

Put the forecast, the cost and the alternatives to the budget owner. Change the limit only after they approve it.

Record the decision

Copy this record into a spreadsheet and complete one row for each process every month. The example shows the support agent.

FieldExample (illustrative)
Month reviewedOctober
ProcessAI agent answers support emails
Expected operating usage10,000 credits
Actual usage12,600 credits
Variance+2,600 credits (+26%)
CauseVolume rose to 2,520 emails after a product launch. Credits per email stayed at 5.
Result1,890 emails closed without a staff reply (75%)
Quality95 closed emails reopened within 7 days (about 5%)
Current limitsAlerts at 75% and 90%. Hard limit of 22,000 credits for all processes.
RecommendationSet the hard limit to 26,000 credits for November and December, from the revised buffered budget of 25,850. Extra overage of up to $160.00 a month. Change it only after approval.
ApproverNamed budget owner, such as the head of finance
DecisionApproved for November and December. The limit returns to 22,000 credits on 1 January.
Data checkCounts and totals only. No customer records or prompts included.

Give every increase an end date. Without one, a temporary increase becomes the new normal.

Common questions

How often should we review AI usage?

Monthly suits most usage-based plans, because many allowances reset each month. Add a quick weekly check in the first months after go-live, or when an alert fires.

What should we do if we keep using less than our allowance?

Record the gap each month. Use three to six months of actual usage to ask for a smaller allowance at renewal.

Should test and build usage go in the monthly usage review?

Yes, if it counts against the same allowance as live work. Record it as its own line. A rise in test usage can mean someone is building a new automation.

More on managing AI tokens and consumption spend

All guides in this area

How usage-based software pricing works, and what drives the bill

Learn how tokens, fixed credits, outcome charges, allowances and limits work in usage-based software, and what makes the bill grow.

Read the guide: How usage-based software pricing works, and what drives the bill

How to forecast AI and automation usage before you sign

Forecast AI and automation usage before you sign. Count billable events, test units per action, model peaks and growth, then compare plan costs.

Read the guide: How to forecast AI and automation usage before you sign

Browse guides by area

All guides