In my previous article, I graded my own year-old predictions about agent pricing against what has actually happened. I found that, while the predictions largely held up, comparing vendors on price has not gotten any easier, thanks in part to the rise of incompatible systems of AI credits. Here, I look ahead to assess where agent pricing might be headed over the next two years, and what enterprises should do about it now.

(Note: My firm, Moor Insights & Strategy, provides advisory services to many tech companies — including, from this article, AMD, AWS, Google, IBM, Intel, Microsoft, Nvidia, Salesforce and SAP.)

Where Agent Pricing Goes From Here

As I mentioned in my previous article, every vendor I have spoken with has said that agent pricing remains a work in progress. So we know that changes will be coming, but this year’s predictions are a bit more challenging as we see continued product innovation, government regulations and enterprise best practices continuing to evolve.

Expect More Sovereignty, IP Protections And Local Use

Microsoft CEO Satya Nadella recently went viral when he shared a pointed insight in a post on X , which he has since framed as a “Reverse Information Paradox.” What got the headlines was the notion that users are paying for AI twice: first for the use of whatever tool they choose, and second by sharing valuable information that makes the models better. It’s not the first time this has been brought up, and I think it is something that holds back adoption of agents as a whole.

Over the past year we have seen many infrastructure vendors bring AI-ready systems and solutions to market, but at the same time we are seeing open-source tools and open-weight models improve. Enterprises know that their data, expertise and know-how can be a source of differentiation and want to protect it — as do the relevant national, local or sector-specific regulatory bodies. As we see the capability for local agents increase, the existing cloud-based leaders will need to take more measures to guarantee these protections. This movement will likely influence pricing models and offerings over the next two years. This is part of a broader sovereignty push across the industry; my colleague Michael Leone recently wrote about how IBM is turning sovereignty into a product in its own right, and I expect agent vendors to face similar pressure.

Expect More Aggressive Bundling by Vendors

While the AI-credits model promotes better unification of pricing and billing from the vendor’s perspective, as I discussed in the preceding article it is not prescriptive enough to help customers decide which products to use, and that has been a major challenge over the past two years. The first round of generative AI investments led to rapid product releases, and often we saw competing products from the same vendor. I am encouraged that vendors have started to consolidate product lines to reduce confusion, but it is likely that there will also need to be some product bundling for specific customer situations.

For example, an enterprise where agents will be built and maintained by professional developers will need a different set of products than one seeking to enable broader knowledge-worker agents. I also think we will see bundles used to promote partnerships or other technology affinities; for example, you may see a company like SAP partner with a specific model provider on a bundle, or AWS partner on an industry-specific SaaS play. While I am not certain we will see these bundles marketed visibly, I do expect that sellers and their partners will leverage bundles and discounting plays to accelerate the selling motion.

Expect Accuracy To Command A Premium

A major issue raised by customers is dealing with inaccurate results or unintended outcomes. The variability associated with agent use can be refined by using better models and context approaches, but that’s still no guarantee of clean results. However, we are starting to see different methods to evaluate agent results before they are delivered. Initially, these evaluation technologies were just another LLM acting as another set of eyes (so to speak) on the request, but now we are starting to see other concepts come to market. For example, AWS’s Kiro is using neurosymbolic AI to ensure that requirements are well translated. We are also seeing companies like Weights & Biases, through tools such as ARIA , loop agents across many experimental runs to refine and optimize results. If these new technologies become more proven, we predict that increased accuracy will command a premium price. What is unclear is whether the premium for evaluations will be billed as a higher-end service, such as a higher number of credits or tokens per request, or just a flat fee for “certified results.”

How Vendors Can Reward Staying On-Platform

As I stated in the previous article, I also believe that vendors will offer significant pricing advantages for customers that keep more of the agent lifecycle within a platform and its associated vendor products. This is analogous to how cloud providers offer a pricing advantage when a customer chooses an infrastructure instance using that CSP’s own custom silicon versus chips from Intel, AMD or Nvidia. My rationale for this is twofold. First, there is not a lot of differentiation between different vendors’ AI capabilities, and despite AI-based migration tools, platform plays are inherently sticky. So, revenue growth for agents will very likely be driven more by additional dollars from existing customers than by net-new customer capture, and the vendors will want to incentivize that growth as quickly as possible. I’ve already seen an early version of this play out in Microsoft’s Copilot Cowork pricing , which rewards staying inside the Microsoft stack.

Second, we are already seeing examples of this where companies like AWS and Google Cloud are offering very attractive pricing and features for data storage. This provides customers with something analogous to a low-cost Snowflake competitor. This can also be used as an enticement to get an enterprise’s data into the cloud and provide agents (and possibly models) with better context and potential performance improvements. I also think this will extend beyond lower-cost hardware and software infrastructure into how AI credits get counted. For example, it may be beneficial for a Salesforce customer to use MuleSoft for integration versus an MCP call.

Could Agent Scheduling Bring Back Surge Pricing?

Another incentive may center on when people choose to execute agents. The advent of technologies like OpenClaw or scheduled tasks within Claude enables agents to run without human intervention. This also means that agents can be scheduled to run at certain times when there may be less strain on the infrastructure. I expect vendors to test both levers: nudging customers toward off-peak execution windows first, then layering in outright surge pricing once usage volume makes the added complexity worth it. As we begin to see humans spending less time “in the loop,” infrastructure optimization starts to look a lot like the time-sharing technologies of the mainframe era.

What Should Enterprises Do Now?

Given the continued rapid pace of innovation, it will be very hard to get ahead of things in the agentic space. That informs these three recommendations for enterprises looking ahead at agent pricing in 2027:

  1. No long-term deals, or at least make deals with major flexibility. I have been cautioning customers against multi-year deals lately for two reasons. The first is that product roadmaps are still fluid, leading to some abandoned or significantly disrupted upgrades. But more concerning to me is the fluid nature of pricing models and the continuing decrease in inferencing costs.
  2. Build a cost-engineering function within your business. Last year I recommended that enterprises consider deploying more observability and FinOps capabilities to help determine a cost baseline. However, given the volatility of pricing and the increased number of reports of surprise billing, better tooling is not enough. You also need staff who have a deep understanding of how agent costing works and what guardrails need to be in place to mitigate risk and encourage the right adoption patterns. I laid out a fuller version of this playbook in “ Scaling the Agentic Enterprise ,” covering both the technology and the business considerations that enterprises must get right.
  3. Work with vendors you know, and fill in gaps with partners. While I have been a proponent of choosing hybrid and best-of-breed solutions for decades, the current agent market may be an exception, at least in the short run. Besides the pricing issues, enterprises will simply need a lot of help to achieve agentic success. Leveraging existing vendor and partner relationships may be the best route to maximize existing investments and leverage existing skills for initial projects, especially when it comes to working with vendor and partner forward-deployed engineering teams. Sizing up which of the current generation of enterprise agentic assistants actually offers forecastable pricing is a good place to start that conversation.

So, Has Agentic Pricing Gotten Better?

Frankly, no. Things are arguably worse than they were a year ago from an enterprise customer point of view. And while I acknowledge and have deep sympathy for product teams trying to work out pricing, we simply are not there yet.

That said, I do have some optimism looking ahead. My sense is that what’s driving the pricing challenge is twofold. The first issue is a lack of usable adoption metrics, though there are multiple factors that suggest headway can be made: forward-deployed engineering initiatives, reduced roadmap churn, credits- and action-based pricing, and better cost controls should provide the right signals to improve the current state. But, second, we also have to take into consideration the financial bet that vendors have made on AI, and the implications of any pricing misstep on profitability and stock price. So, whatever progress does get made will need to be balanced against vendor goals and market expectations.

Moor Insights & Strategy provides or has provided paid services to technology companies, like all tech industry research and analyst firms. These services include research, analysis, advising, consulting, benchmarking, acquisition matchmaking and video and speaking sponsorships. Of the companies mentioned in this article, Moor Insights & Strategy currently has (or has had) a paid business relationship with AMD, AWS, Google, IBM, Intel, Microsoft, Nvidia, Salesforce and SAP.