ai governance
most of your work does not need the expensive model
american businesses spent the summer moving everyday ai work onto cheaper models, and told the company that counts their spend that the work held up. the habit is one you can copy.
in september 2026 a corporate card company published a report on what american businesses pay for ai, and it made 2 kinds of headline. one kind said the ai boom was showing cracks, because businesses were buying less of the expensive product. the other said the price of running ai had fallen 41 percent since march1. they are 2 sides of the same invoice, and if you sign off on an ai budget, you are on the good side.
what the report found
ramp is a corporate card platform, so it sees what thousands of american businesses pay for each month. its september report has 2 findings that matter to you. the first is that the vendors cut their prices: openai and anthropic both announced a run of cuts, and the effective price of running ai fell 41 percent between march and september1. the second is that businesses are choosing the cheaper products on purpose.
an ai company sells a range, called models: powerful ones at the top that reason through hard problems slowly and cost the most, and quicker, cheaper ones below that do ordinary work well. in august, more than half of the ai work ramp could see was going to the top of that range. by september it was 45 percent, and companies told ramp why: they had set company-wide defaults away from the expensive models, because the standard ones were "still highly performant and also more cost effective"1.
the move itself is simple. the companies that changed course set a default: the standard model for everything, the expensive one only when a job earns it. a quick summary goes to the cheap model. a hard analysis gets escalated. nobody at a desk has to remember to economise, because the default does it for them. the question it leaves you with is which of your jobs need the expensive answer, and you are the person qualified to answer it.
what to do about it
if your tool lets you pick the model, pick the cheaper one for a week on the routine work, summaries, drafts, tidying, looking things up, and see what you notice. the companies in ramp's report told it the standard models held up, and a week of your own results is worth more than anyone's word.
then write the list. the few jobs that genuinely need a model that reasons for 30 seconds are usually obvious once the routine ones are on paper. that list is your team's default, written in plain english. if it is your it team that picks the model, the list is what you hand them, and it is the same move the companies in ramp's data made this summer.