Last year I completed research about the emerging pricing practices for agentic solutions, which I captured in two articles, “Demystifying AI Agent Pricing” and “Headwinds and Tailwinds Driving the Future of AI Agent Pricing.” The trigger for that research was the disruption of the per-user-per-month packaging we had come to expect from chat-based AI solutions. It was the power of reasoning models, combined with semi-autonomous agentic harnesses, that effectively broke the business model for per-user pricing. AI vendors needed to make changes to maintain their margins as aggressive AI build-out continues.
My initial goal with this research was to dig a bit deeper and understand the challenges facing enterprises as they deploy more agent-based solutions. Would it be easy to use price as a selection criteria? Are product stacks comparable enough for a true apples-to-apples comparison? And what suggestions could I provide for enterprises heading into the next budgetary cycle? It turns out this is an almost impossible task without deep discussions and negotiations between customers and vendors. Even something as simple as AI credits — the new billing metric being used by many vendors including Salesforce, Microsoft, and SAP — is not easy to define, compare, or forecast.
Simply put, the industry is still figuring this out. In fact, in every vendor briefing about pricing I have received this year, the only area of consensus is that pricing remains a work in progress. My core takeaway one year in: Pricing has not gotten easier to compare, and the rise of AI credits is a big part of why. Given that sobering realization, my focus one year into my work on agentic pricing is to assess what was predicted, provide insights on the current market, and update my predictions for the next 12 months. This is part one of a two-part series: This piece grades last year’s predictions and sizes up today’s market; part two covers what’s coming next and what enterprises should do about it now.
Agentic Pricing Predictions From 2025
Overall, my predictions from “Where Is AI Agent Pricing Headed? Predictions for the Next 3 Years” from last year were solid, but unfortunately that does not mean that comparing different solutions is any easier now, and in fact it is arguably harder in some cases given the advent of AI credits. Also as I suggested last year, the ability to predict and forecast future costs without metering or better FinOps technology remains a challenge.
Prediction 1: Time-Based Pricing Will Merge Into Consumption-Based Pricing
- Rationale: Time-based (per-seat, per-month) pricing lacks viability for agents given inconsistent usage. Vendors would move toward “reserve pricing,” a monthly fee for a reserved capacity block plus overage charges beyond it, with pooling strategies letting teams share a consumption allocation rather than paying per user.
- What has actually happened: Largely confirmed. Every vendor we examined now has some form of metering for agent usage that’s separate from a base price. However, the billing for that metering is often based on somewhat obscure units, sometimes referred to as “credits.” While credits make capacity more fungible across a vendor’s stack, they are also more complex and less predictable. (Billing for text- versus voice-based interactions is an example of this.)
Prediction 2: Action-Based Pricing Has Potential
- Prediction: A per-action model, comparable to what Salesforce was signaling, has real viability, particularly where agent cost can be measured against the cost of a human doing the same task. This works best in domain- or industry-specific solutions, with pricing that scales according to task complexity and replacement risk.
- What has actually happened: Partially confirmed as a viable model, but still a single-vendor pattern. Salesforce has now adopted the Flex Credit “standard action,” which costs a flat 20 credits ($0.10) regardless of how many tokens or LLM calls it took. This improves cross-product fungibility but is also less transparent and predictable than customers want.
Prediction 3: Consumption-Based Pricing Will Not Change Much
- Prediction: The core consumption model (tokens, compute, API calls) would remain largely intact for enterprise applications, but vendors would increasingly bundle multiple products (IDEs, platforms, observability tooling) into simplified, multi-product packages rather than changing the underlying metering.
- What has actually happened: Confirmed. AWS’s stack is still priced almost entirely in literal, granular consumption units (vCPU-hour, per-1,000-invocation, per-token), essentially unchanged in structure. Google prices its infrastructure layer the same way. The bundling half of the prediction landed even more clearly than expected: All three tiers of vendors have consolidated and rebranded their agent platforms into single umbrella SKUs within the last 12 months. The underlying meter didn’t change; the packaging around it did, exactly as predicted.
Where Agent Pricing Stands Today
So, what does this landscape look like today? As I stated earlier, vendors still understand that this remains a work in progress. In fact, we are now dealing with more options than before. This continued experimentation puts an extra burden of research on enterprises trying to figure out what will work best for them at the right price.
Credits-Based Systems Have Challenges
In an attempt to simplify pricing across a wide range of AI and SaaS products, some vendors have introduced “credits.” Unfortunately, the simplicity this achieves within that vendor’s own ecosystem has also produced incompatible cross-vendor comparisons (see the graphic below as an example). While making an apples-to-apples choice difficult, the bigger controversy with credits is the lack of predictability over time. Without a true countable metric to attach to (such as users, servers, or minutes), credits can be priced somewhat arbitrarily and used to incentivize what the vendor wants to achieve rather than the customer’s desired outcome. This could yield good or bad results for the customer.
