Full narration of the briefing. Click a section heading to jump the player to that moment.
We are going to spend this episode on why this negotiation is harder than the ones we usually teach, because if you walk in expecting an Oracle renewal you will be surprised in ways that cost money. I am Tom, Daniel is with me. There are five difficulties. None of them means you should not buy.
All of them change how you should buy, and most of them argue for the same thing, which is buying smaller and shorter than instinct suggests.
The first difficulty is the one we feel most acutely as an advisory firm. There is no published benchmark for what comparable enterprises commit or what discount they achieve. In a Microsoft or an Oracle negotiation we can tell you where a number sits in a band because there is decades of precedent and a large peer set. Here the agreements are recent, private, and few.
So when someone tells you a specific discount is market for AI commitments, ask them where the number came from. We will not quote one, because we cannot source one honestly.
The second is that your own demand is genuinely hard to forecast, and not because anyone is careless. A single new feature shipping into a popular product can double consumption inside a month. A team discovering a better prompt can halve it. Adoption in this category tends to be lumpy rather than linear, arriving in steps as individual teams go live.
Every other software forecast we build rests on headcount or transaction volume, which move slowly. This one rests on engineering decisions that have not been made yet.
Third, the thing you are buying changes during the term. New models arrive, existing ones are retired, capabilities improve and the price to accomplish a given task tends to fall. That is good for you in the medium term and awkward in a contract, because you are committing money against a product specification that will not be the same in eighteen months. Traditional software moves slowly enough that a three year term is a reasonable bet on a known thing.
Here a three year term is a bet on something you have not seen yet.
Fourth, a difficulty that hides as an advantage. On paper switching model providers is easy, because it is an API call and the alternatives are real. In practice your prompts, your evaluation suites, your guardrails and your output handling are all tuned to one model's behaviour, and that tuning is the actual asset. So the switching cost is not integration, which is genuinely low.
It is revalidation, which is not. Be honest with yourself about that before you build a negotiating position on portability, and we will come back to it next episode.
Fifth, and it is organisational rather than commercial. In most software categories procurement owns the vendor relationship and can see the spend. Here consumption is driven by engineering decisions, often across several teams, and the bill arrives after the fact. We regularly meet organisations where nobody could say which application was responsible for the largest line on the invoice.
If you cannot attribute your own consumption, you cannot manage it, you cannot forecast it, and you certainly cannot negotiate about it credibly.
Put those five together and they argue for the same posture. Commit smaller than your expected case, because your forecast is weak. Take a shorter term than instinct suggests, or build in review points, because the product will change. Protect the downside on rates rather than locking them.
And instrument your consumption before you negotiate, because attribution is the foundation of everything else. Next time, Claire and Daniel take the other side of this, which is that you have more leverage here than in almost any mature category.
Redress Compliance works on contingency: our fee is 25 percent of what we save you. Nothing saved, nothing paid. Independent, buyer side only, never vendor funded. Want Redress to contact you? Reach out and we respond the same day.
Talk to a Anthropic negotiator