Usage Forecast

A usage forecast is a projection of how much of a metered service an organisation will consume over a future period — API calls, tokens, compute hours, storage, seats. It matters commercially because consumption pricing and committed-spend discounts both turn the forecast into a financial position: under-forecast and you pay list rates for the overage, over-forecast and you pay for capacity you never used. The mechanics are straightforward for stable workloads. Take measured usage, decompose it into the drivers that actually move it — active users, requests per user, work per request — and project each driver rather than the total, because a total hides which component is growing. Then apply seasonality and any known step changes: a launch, a migration, a large customer onboarding, a feature reaching general availability. AI workloads resist this in specific ways. Consumption depends not only on how many people use a feature but on how they use it: prompt length, retrieved context size, conversation depth, how often an agent loops, and which model handles the request. A change in any of those moves the bill without any change in user count. Adoption curves are also unusually steep and unusually uncertain, because an assistant that becomes genuinely useful gets used far more than one that is merely present. That argues for forecasting in unit terms rather than in totals. Project the number of conversations, documents or tickets, forecast the cost per unit separately, and multiply. This separates growth from efficiency and makes the assumptions visible, so a forecast that turns out wrong can be diagnosed — was there more traffic than expected, or did each interaction get more expensive — rather than merely revised. Forecasts should be expressed as ranges with explicit assumptions, not as single numbers. The commercially useful outputs are a realistic floor, which is what a commitment should be sized against, and a plausible ceiling, which is what spend caps and budget alerts should be set against. Committing at the expected case is the standard mistake, because the discount tiers reward optimism and the shortfall is rarely recoverable. Finally, close the loop monthly. Compare forecast to actual, decompose the variance into volume and unit cost, and adjust. A forecast that is never checked degrades quietly into a number that gets copied forward, and the first time anyone notices is when a commitment renews or an invoice arrives larger than the budget it was supposed to fit.