For FP&A teams

Stop plugging the AI line.

AI is the number in your model with the widest error and the least support behind it. Get the drivers, the scenarios, and variance you can decompose without a week of investigation.

How should FP&A model AI spend?

As a driver-based cost, not a trend line. AI spend is the product of volume and unit cost, and those two move independently: a launch multiplies volume while a model change halves unit cost. Camaze forecasts each workload separately on its own drivers, applies the pricing actually in force including committed tiers, produces a range rather than a point, and decomposes variance into volume, unit cost and scope so a miss can be explained rather than absorbed.

The problem

The AI line is the one you cannot defend.

Every other significant cost in the model has a driver behind it. Headcount has a hiring plan.

More detail

Cloud has committed spend and a growth assumption. AI has last quarter's actual and a percentage, because nothing else is available.

Then it misses, and the miss cannot be explained. Was it volume or price. Was it a launch, a model change, or something new that shipped between planning cycles. Reconstructing it means a week of asking engineering teams what they deployed, and the answer arrives after the pack has gone out.

  • No usable driver, so the line is an extrapolation
  • Step changes at launch that a trend line cannot anticipate
  • Volume and unit cost moving in opposite directions in the same period
  • New workloads appearing between cycles with no owner in the plan
  • Prepaid credits masking the real run rate until they are exhausted
Total AI spend All providers
This month to date$151,400+12.8% vs last month
Forecast, month end$168,200Range $161,000 to $176,000
Identified savings$61,90041% of run rate
Monthly AI spend over twelve months with a three month forecast Spend rises from about 41 thousand dollars to about 151 thousand dollars over twelve months. The forecast tail projects roughly 168, 187 and 208 thousand dollars over the next three months, inside a shaded confidence range. 0 $65k $130k $195k $260k FORECAST AugOctDecFebAprJunAugOct
Actual Forecast Confidence range

Illustrative product view. Figures are examples.

What we do about it

Give the line a driver.

Each workload is modeled on the things that actually determine its cost, then aggregated. That is what makes it possible to answer a question about one launch without rebuilding the model, and to explain a variance without an investigation.

  • Volume and unit cost forecast separately, then combined
  • Per-workload growth patterns, including weekly shape and seasonality
  • Contracted pricing applied, including committed tiers and volume discounts
  • Confidence ranges, with the sources of uncertainty named
  • Prior forecast accuracy tracked, so the range carries weight
Scenarios

Answer the conditional question in the meeting.

Planning conversations about AI are almost entirely conditional. What does this cost if the pilot goes to every user.

More detail

What if we move that workload to the cheaper model. What if usage doubles and we do nothing.

Scenarios are built on any workload and compared side by side, so the answer arrives with a range attached rather than a promise to follow up.

  • Rollout scenarios, from a pilot cohort to general availability
  • Model change scenarios, with unit cost isolated from volume
  • Growth scenarios at several rates, each with a range
  • An optimization scenario, showing the savings plan executed
Scenario planning Monthly run rate
Five quarter AI spend forecast under three scenarios On the current trajectory monthly spend reaches about 447 thousand dollars. Rolling the feature out to everyone reaches about 662 thousand. Following the optimization plan holds it at about 238 thousand. 0 $175k $350k $525k $700k $447k $662k $238k NowQ1Q2Q3Q4Q5
Current trajectoryFeature rolled out to everyoneWith the optimization plan

Illustrative product view. Figures are examples.

Variance

Explain the miss in one line.

Variance against plan is decomposed rather than reported as a single number.

More detail

Volume against plan, unit cost against plan, and scope that was never in the plan are separated, with the largest movements listed first.

The difference is between reporting that AI came in 340 thousand over and reporting that volume was on plan, unit cost fell after a model change, and the entire overrun is one workflow that shipped in month two and has an owner.

  • Volume, unit cost and scope reported separately
  • Largest movements ranked, with the workload and owner named
  • Rolling reforecast on your cycle
  • Drill from a variance figure to the underlying usage
Cost by project and model July
ProjectModel ShareCostChange
Support automation Claude Sonnet $34,800 +22%
Search and ranking GPT-4o mini $21,400 -6%
Sales copilot GPT-4o $18,900 +31%
Docs assistant Llama 3.1 70B, self-hosted $15,200 +9%
Data enrichment Gemini 1.5 Flash $11,600 -14%
Internal agents Mixed $9,800 +64%
Engineering seats Cursor, Copilot $8,300 0%

Illustrative product view. Figures are examples.

Repricing

The market changes your assumptions for you.

AI is the only major line in the plan where the unit price can fall 40% between forecast cycles without anyone doing anything.

More detail

It is also a line where a new model can make an assumption obsolete two weeks after the pack goes out.

New models, new prices and new discounts land every few weeks. You hear about them the week they land, priced against your own usage, in time to plan the change instead of reading about it afterward. Forecasts reprice automatically when a provider changes pricing on a model you run, and a new cheaper option arrives as a scenario you can put in front of the budget owner rather than as news.

Scenario planning Monthly run rate
Five quarter AI spend forecast under three scenarios On the current trajectory monthly spend reaches about 447 thousand dollars. Rolling the feature out to everyone reaches about 662 thousand. Following the optimization plan holds it at about 238 thousand. 0 $175k $350k $525k $700k $447k $662k $238k NowQ1Q2Q3Q4Q5
Current trajectoryFeature rolled out to everyoneWith the optimization plan

Illustrative product view. Figures are examples.

Your model, not ours

It feeds the model you already built.

Nobody is rebuilding a mature planning model around a vendor's interface. Camaze is designed to be the driver source underneath it.

More detail

Load your plan at whatever level you hold it. Take actuals, forecasts and drivers back out into Excel, Google Sheets, Snowflake or over the API, on your own cycle and in your own structure.

  • Plan loaded at your own level of detail, no restructuring
  • Excel and live Google Sheets output
  • Nightly Snowflake tables at full detail
  • REST API covering actuals, forecasts, budgets and findings
Scheduled reports 6 running
ReportRunsLands in
Monthly close pack1st of the monthExcel, emailed to finance
Chargeback journal1st of the monthERP export, journal-ready
Board AI summaryQuarterlyPDF and slides
Cost by teamEvery MondayGoogle Sheets, live
Raw cost and usageNightlySnowflake
Budget variance digestEvery FridaySlack, #finance

Illustrative product view. Figures are examples.

Bring next year's AI assumption.

We will build the driver-based version of it on your own usage during a 30 minute call.

In practice

Where it fits in your calendar

No new cycle and no new process for the FP&A team.

1

Monthly close

Attributed actuals and decomposed variance arrive on your close calendar, with an accrual where provider billing has not yet settled and a true-up when it does.

2

Rolling reforecast

The forecast updates continuously and can be pulled on your reforecast cycle, with a range and with prior accuracy attached so the range can be weighted.

3

Annual planning

The AI line goes in as scenarios rather than a single number, with assumptions documented. When it is challenged in the review, the assumptions are visible.

FAQ

Questions people ask

Do we have to move our planning model into your tool?

No, and we would advise against it. Camaze is designed to sit underneath the model you already have as a driver source. Load your plan, take drivers and actuals out into Excel, Sheets, Snowflake or the API, and keep your model where it is.

What level can we plan at?

Company, department, product, project or individual workload. Most teams hold the plan at department level and use workload detail for explanation rather than for planning, which keeps the model manageable.

How do you handle prepaid credits and committed spend?

Both are modeled as what they are. Credit drawdown is reported as consumption rather than as zero cost, so the real run rate is visible before the balance is exhausted. Committed spend is applied to the forecast including the point at which the commitment would be used up.

Can we see how accurate previous forecasts were?

Yes. Prior forecasts are retained and scored against actuals, by horizon. That tells you how much weight an intra-month projection deserves compared with a four-quarter one, which is usually a large difference.

What about workloads that do not exist yet?

They go in as scenarios. You supply expected volume and workload shape, and Camaze prices it against current model pricing and comparable existing workloads. It is an estimate, it is labeled as one, and it is built on your own unit costs rather than a list price.

Model the AI line properly.

Book 30 minutes and we will build a driver-based forecast on your own usage during the call.

5 minute setup. Read-only. No engineering time needed.