The AI Budget Problem
AI will doubtless have some really good use cases. There are already things AI can do that would never be practical or cost effective to do any other way. It will probably make life better at least as much as it makes it worse. At this point in time though, I believe the evidence supports a position that it is being significantly oversold and yet, many of us are buying.
I hear wildly optimistic statements about AI every day. Most of them are from senior managers convinced that this new magic (Clarkes Third Law) will allow them to transform their cost pressured businesses into something sustainable. This might yet prove to be true, but I sometimes feel like the only CEO not banking on it. Why? Here’s one undeniable reality.
AI is not free. There is both a societal and environmental cost to AI, but hard data is limited so we will ignore this now and just consider direct cost to the end user. Whilst there are limited versions of many models you can access at no financial cost, they will all, sooner or later have to find a way of making money and unlike many other products, AI is not getting cheaper, but more expensive with each new generation.
Almost no foundation model builders are making money from AI at the moment – it’s all speculative venture capital based on the potential of future returns. Everyone is trying to capture market share and worrying about monetisation later. This means that there is a truly epic and fast-growing debt that is going to have to be paid. Paid for by you, the user and if not today, then tomorrow.
Consider the challenges of building an economic case for change based on the use of an AI tool that you know will have to get much more expensive, but you have no idea by how much. If you are basing your cost model on a fixed monthly subscription then you are living in a dream. No AI platform will be sold in the long term as an all-you-can-eat buffet. Instead, you will be eating tokens. A different number of tokens for every task – but without any meaningful way to forecast how many tokens any job will take.
There’s no way to anticipate how many tokens a prompt will actually burn, which makes any kind of budgeting a non-starter. It’s like going to the supermarket and committing to buy a gallon of milk, not knowing if it’ll cost you $5 or $50. Ed Zitron Why Are We Still Doing This?
More complicated? More tokens. More iterations? More tokens. A mistake that needs to be corrected, more tokens. Agentic AI – loads more tokens. The cost differential between a currently available all-you-can-eat subscription plan and a token-based API for any kind of AI power user, will be huge. How can you write your transformation business case without knowing what the foundational component required to enable it, is going to cost? One thing you can be sure of is that it will cost more, a lot more, than it does today. This isn’t speculation, its fundamental economics. To base your business case on the idea that AI consumption will get cheaper over time, ignores the reality of the current cost model. I expect new hardware will mitigate this in the future but basing a financial model on something yet to be invented can perhaps be most kindly described as ‘aspirational’.
Microsoft is currently promoting the idea that companies should report the costs of ‘digital labour’ in the same way they include analogue labour (that’s you and me). Why is this useful? Once you start to conceptualise the idea of a digital workforce, it seems only reasonable that you should also buy software licences for that workforce to use. The cost of that digital workforce will increase over time as the costs of your software licences increase. However, analogue people costs increase at roughly the rate of inflation, software costs increase at significantly above inflation.
Software and Hardware companies have not become the largest corporations in the world through their generosity. If a piece of software can save you £100, you should expect the people who sold it to you to want a very significant proportion of that. They will need it to pay for the pile of debt they have incurred building the tools that let you save the money. That’s not unreasonable, unexpected or unfair. But not knowing what the tool will cost tomorrow makes it near impossible to forecast cost/benefit. The business case that justifies the running costs of your AI enabled transformation is significantly less predictable than a spin of the roulette wheel. Are you really prepared to stake the success of your financial transformation on that?
This doesn’t mean that AI is not a good solution to your problem, or an opportunity worth exploring – it might be. But you should not base your financial model on today’s costs or licencing options. The foundation model builders are almost certainly going to have to increase costs to get to breakeven. Either that or the platforms will become increasingly enshitifed and of lower value. It would be wise to check your numbers still work with AI cost growth of several times. Given software costs rise much faster than inflation, you also shouldn’t expect your costs to fall over time. If you don’t think like this, the ‘lower value human capital’ you potentially replace might turn out to have been better value than you thought.
Giles Letheren, Chief Executive Officer