Three Ways I Think About AI Spend

People ask me this one a lot. What's different about forecasting AI costs versus forecasting a normal software subscription. I get it, on paper it sounds like some brand new cost category finance has never dealt with before.

Turns out it's not as wild as people expect. A lot of modern SaaS tools already charge based on usage instead of a flat fee, so AI spend isn't reinventing the wheel as much as people think.

I break AI spend into three buckets.

• The first is stable and recurring. Specific AI tools with a fixed cost, could be monthly, quarterly, or yearly. This one is easy. It behaves just like traditional SaaS.
• The second is general usage. This is what most people picture when they think of AI spend, teams using tools like Claude, GPT, or Copilot day to day. I track this at a weekly or monthly level with the team, since it moves with how much people are actually using it.
• The third is project based, and this is the one that gets missed the most. A lot of companies are quietly building small internal tools to solve specific problems, almost like running a tiny dev shop inside finance or ops. Each of those projects has its own cost, its own timeline, and its own mix of people, tooling, and tokens.

Once you split spend into these three buckets, your forecast gets a lot more accurate. You stop trying to force one formula onto three very different kinds of cost.