Judge AI Spending by Its Return

AI is better treated as capital deployed for an output than as labour whose effort is inherently valuable. For a business, the useful measure is leverage. How much more valuable output can the same investment produce, or how much less investment can preserve the result? More agents, tokens or generated code don’t answer that question because activity isn’t the same as value.

That distinction matters when a dependable monthly subscription gives way to usage-based costs. A business can do a great deal with a fixed $200 bill, but variable token consumption forces it to choose where AI is worth applying. Expensive models and long-running agent loops should be directed towards tasks where their added capability changes the business outcome, rather than used simply because more computation is available.

Token maximisation reverses that discipline by making consumption look like progress. Spending $1 million a month instead of $200 for a 5 per cent performance improvement may produce a technically better result, but it would deliver dreadful leverage unless that small gain created extraordinary value. The relevant comparison isn’t how much work the AI performed. It’s whether the improvement earned the capital consumed.