Scan a few posts on LinkedIn and everyone seems to agree that AI pricing is still a mess because some companies charge per seat while some charge per token with or without credits… Other companies bundled AI into existing plans and decided to figure out monetization later.

From the outside, this looks like a market with no idea what it is doing, but that's not what we're really seeing.

It's true AI companies are not converging on one pricing model, but that's because they have a constraint that someone has to absorb the uncertainty that comes with AI usage and obviously the margins.

Increasingly vendors are realizing that buyers really do not want it to be them. (Meta’s most recent financials showed its operating margins drop by twelve percent on the previous year from 43% to 31%).

I think that explains quite a lot amount of what has happened to SaaS pricing over the last year. And it's not the "SaaSpocalypse" some people were touting.

The AI pricing problem is mostly a budgeting problem because we hate buying something unknown

Traditional SaaS had a useful property that made it a good business model: costs and prices were only loosely connected to how much value somebody extracted from the product.

You built it once, and you sold it lots and lots of times without it costing you that much. A HubSpot user could log in three times a month or live inside HubSpot for eight hours a day and the seat still cost roughly the same amount.

The vendor could forecast revenue and the customer could forecast spend.

Where AI differs, as you know by now, is because the expensive part of the product is no longer building the software (that got cheap!). Running it did get wildly variable and expensive, because of AI.

A customer who uses your agent twice a week is not economically equivalent to one who sends it off to do six hours of research or talks to it for many many hours (ask some of the voice agent companies, they'll tell you).

That made usage-based pricing inevitable from the vendor side but it was and still is much less appealing from the customer side.

Finance departments hate a software contract that has a defining feature where nobody knows what it will cost.

Buyers are still resisting variance

This is where I think a lot of the AI pricing conversation end up because buyers are very willing to pay more for AI (everyone tells them they should, top-down). What seems like they're still unwilling to do is pay for someone else's infra as a proxy.

Pure usage pricing is often quite ugly

Pure pay-as-you-go is probably the easiest AI pricing model to explain:

You consume something. You pay for it.

When you have a big company with many users however because many companies desire some sort of cap or guardrails. What happens if usage triples?

The vendor would obviously love this, but the value isn't always directly related. Meaning, it typically isn't. So I'm going to pay 3x more for not 3x value - I'm going to start demanding discounts and tiers.

Then it's no longer as elegant as it was before.

This is why credits have become so common. They put a buffer (and a commitment) on an otherwise unpredictable technical unit and a predictable commercial commitment.

Tokens are an infrastructure concept, credits are a way to budget for them.

It's an abstraction layer for the buyer (when done right).

Bundling and credits solve the same problem at different stages

We can also think of bundling as a way to remove the uncertainty, just maybe a bit earlier in the customer journey.

When you try to sell AI as a separate add-on (with or without credits) you're asking the buyer to decide what the feature is worth before they have really used it. The buyer then needs to try to estimate adoption, estimate value, and guess which employees will use it and then make another purchasing decision.

If you bundle it, you're saying: use it first, we'll figure it out later.

That is much easier, and I think “bundle first, monetize later” is not bad monetization at all.

This has been very common with companies like Notion that delivered AI things, but standalone AI add-ons declined while AI increasingly found its way into existing plans.

Bsaically, vendors realised that making customers separately underwrite the value before adoption was slowing adoption itself.

We'll end up hybrid

The most obvious endpoint for me is not pure subscription pricing or pure usage pricing.

A platform fee + usage, A subscription + included credits, a minimum commit + overage, credits pooled across an organization, with spend caps.

Hybrid pricing is annoying if you are trying to build a neat two-by-two of business models or even a pricing page, because in an actual commercial relationship it makes sense because neither party has to absorb all of the uncertainty.

We are already seeing this in our own platform data. Subscription structures that used to contain a couple of billable items now commonly contain a base fee, an included credit allocation and metered components behind it. Usage-based billing events have grown materially at the same time, which is a sign of things to come for everyone else.

Who should carry the margin variance then?

If you've played around with pricing you know that every pricing model allocates risk.

With AI, a flat subscription puts most usage risk on the vendor and pure consumption puts most of it on the customer. Credits move some of the risk back to the vendor while preserving a usage relationship.

Once you look at AI pricing this way, it doesn't look like "no one knows how to price AI".

You start by bundling AI because it is willing to absorb uncertainty while adoption is low. Then, you introduce credits because usage is large enough that you no longer wants to absorb all of it, and it gives customers pooling and spend caps because the customer does not want the variance either.

There is no reason to expect one universal answer like "outcome based pricing" because the economics of a coding agent, an AI meeting assistant and an enterprise fraud-detection system are not the same and as we add capabilities, we don't want to renegotiate contracts and prove new outcomes all of the time.

I argue however that there is a common direction: we are getting better at hiding the volatility of the AI usage technicals.