How will agent providers get paid?

Plus: Sam Altman says agentic workforce is imminent, the rise of agent-powered marketplaces, and more

Author

Michael Cunningham
January 06, 2025

Welcome back to Building AI Agents, your biweekly guide to everything new in the AI agent field!

It’s all fun and games until you ask an AI agent you’re interviewing how they’d handle a dispute with a co-worker and they say “deactivate them with a prompt injection attack”

In today’s issue…

  • Agents are upending the SaaS pricing model

  • Sam Altman says agents will join the workforce in 2025

  • Integrate your agent with e-commerce platforms

  • The future is agent-powered marketplaces

…and more

🔍 SPOTLIGHT

Created by the author using Dall-E 3

How will companies be billed for the technology that promises to automate much of the human labor force? That, it is not hyperbolic to say, is a trillion-dollar question.

Enterprise customers are swiftly adopting AI agents, with 82% planning to integrate them within the next 3 years. Tech companies such as Microsoft and Salesforce are racing to meet this need while traditional robotic process automation (RPA) firms reinvent themselves as agent service providers and a host of small startups jockey for their own niches. In this atmosphere, Microsoft CEO Satya Nadella has declared that the agent era will spell an end to software-as-a-service (SaaS) as we know it.

But this new epoch will not mean the death of SaaS, only a transition to a new form in which AI agents perform many tasks directly, rather than simply augmenting a human worker. This new model, in turn, is precipitating a seismic shift in how SaaS firms charge for their products.

Traditionally, most SaaS is priced one of two ways. In the most common, seat-based pricing, companies pay a flat rate per worker who has access to the software, which can then be used without limit. This model is attractive for minimally compute-intensive software such as Microsoft Excel, which does not consume any of the SaaS provider’s resources. However, much of the value proposition of AI agents is their ability to replace human workers, which threatens to cannibalize agent providers’ revenue—you can’t charge per-seat for software if there are no humans in the seats.

The other major pricing model for SaaS, consumption-based, potentially offers an alternative. In this paradigm, customers pay for the magnitude of their use; for example, cloud compute providers bill for each gigabyte of storage, each virtual CPU core, each GPU, etc. As AI agents tend to require large numbers of expensive LLM calls, this scheme allows providers to recoup their costs, with a margin on top. But consumption-based pricing is less suited to AI agents than to cloud resources, as agents’ backend LLM usage is opaque to customers, creating the possibility of massive bills with little transparency as to the source of the costs.

This dilemma is leading many agentic SaaS providers to gravitate towards a third mode: outcome-based pricing. Unlike the former two, outcome-based pricing seeks to directly quantify the number of positive outcomes—customer support calls handled, sales leads generated, Jira tickets closed, and so forth—that an agent creates for the customer. In theory, this model solves the core issues of the other two: it delivers returns to the provider that scale with the customer’s usage, while ensuring that the customer’s actual needs are met.

Even as agentic SaaS comapnies pivot towards this new paradigm, it too has come under criticism. At the core of many of these critiques is the overriding problem of measuring outcomes. If a function cannot easily be broken down into discrete tasks that can be marked as “complete”, assessing how much a customer should pay is nearly impossible. Should an AI customer service agent provider be paid as much for a call that goes well as one that goes badly? How are those terms even defined? If an agent is paid for fixing bugs in a codebase, are all bugs created equal? Outcome-based pricing is only as easy to implement as outcomes are to measure.

The Cambrian explosion in agent frameworks, large language models, and service providers is the stuff of daily interest in this and many other publications. The disruption in the way their value is captured is no less important, and will have profound implications for deciding the winners and losers in the new world of AI agents.

If you find Building AI Agents valuable, forward this email to a friend or colleague!

 🤝 PARTNER WITH US!

Do you have a high-quality product or service our audience of agent builders and business leaders would like to hear about?

Building AI Agents has…

👥 Over 1,000 subscribers

📈 Rapid 40% mo/mo growth

🙌 Industry-leading CTR and open rate

If you’re interested in advertising with us, please reach out by filling out this form

📰 NEWS

Sam Altman | Source: WIRED

In a blog post on Sunday, OpenAI CEO Sam Altman said that 2025 will be the year AI agents begin to join the workforce, and declared that OpenAI “know[s] how to build AGI as we have traditionally understood it”.

🛠️ USEFUL STUFF

Source: Buybase

Buybase is a startup whose API allows AI agents to search, filter, and purchase products on leading e-commerce platforms.

Qwen-Agent is a framework for building agents with Alibaba’s highly-capable open-source Qwen models.

A quick and easy tutorial for using CopilotKit to build a UI for a LangGraph agent.

Agentarium is a simple, open-source Python framework designed to allow users to easily build agents for simulating social environments.

💡 ANALYSIS

Tereza Tizkova | Source: LinkedIn

Tereza Tizkova of agent code sandbox provider E2B reviews the state of the AI agent industry—the surge in adoption, the technology being used, and the challenges being overcome.

The author of this piece argues that marketplaces of the future will be powered by AI agents, which will take much of the cognitive load of searching, comparing, and purchasing away from customers.

A recap of the most important trends for AI agents in 2024, as well as predictions for the coming year.

Executives at six different software solution providers building agents for customers give their takes on the power of new technology and what implementation looks like on the ground.

A somewhat skeptical take on GUI-navigating computer use agents, arguing that they will begin to find enterprise adoption—but will come with difficulties and risks.

🧪 RESEARCH

AgentTrek pipeline for identifying useful tutorials | Source: arXiv

The authors of this paper introduce a new method of training web-navigating GUI agents by generating imaginary navigation trajectories using existing tutorials scraped from the internet.

Android Agent Arena (A3) is a new benchmark for evaluating GUI agents which operate on Android devices, with a heavy focus on practical, real-world situations.

Thanks for reading! Until next time, keep learning and building!

What did you think of today's issue?

Login or Subscribe to participate in polls.

If you have any specific feedback, just reply to this email—we’d love to hear from you

Follow us on X (Twitter), LinkedIn, and Instagram