How I think about reducing AI costs

AI inference costs are something I've been writing about for a while. It's clear that for many companies this is becoming a huge problem: AI spend per employee per month. Source: Ramp AI Index via a16z, 12 August 2026. I've heard from a lot of readers that reducing this cost is becoming a hot topic internally. So, here's how I think about this problem at a high level. Audit costs You need to have a good handle on what is driving your bill. I've met a lot of companies who have a pretty fragmented understanding of the costs - they can be siloed over many different teams, business units and roles. At a minimum, you need to know - business-wide - how much you are spending on AI and importantly, key stats on how that breaks down. There are two main dimensions on this. Firstly, the model in use. I see a lot of people running ancient, poor value for money models. The tech debt is real here! For example, GPT-4o is $2.50/$10 per million tokens, but is drastically worse than GPT-5.6 Luna, which…