The agents are here: But at what price?

Welcome to Runtime! Today: How Zendesk is approaching one of the biggest shifts in enterprise software pricing in years, Meta gets into hot water over the claimed performance of a new AI model released over the weekend, and the latest funding rounds in enterprise tech.

(Was this email forwarded to you? Sign up here to get Runtime each week.)

No owners, only spenders

The forced march to agentic AI as the pulse of enterprise software isn't just a technology shift, it's a business-model shift. If we are to assume agents are the future of enterprise software — and why wouldn't we, given how many enormous tech companies searching for growth have been telling us so over the last year — they will forever change the way business software is bought and sold.

Instead of buying seats for individual users of enterprise software, companies that adopt agents will pay based on the activity those agents generate, sort of similar to how they've paid for cloud infrastructure services over the last 20 years. Zendesk CEO Tom Eggemeier thinks that its customers should only pay for agentic activity that actually gets something done.

• "What we're offering from an AI agent or bot perspective is we only get paid when the AI agent actually solves a customer's business problem," Eggemeier said in a recent interview.
• The customer-service software company calls them "automated resolutions," and they are calculated in a few different ways.
• The meaning of what counts as a "resolution" differs slightly based on the activity, but Zendesk uses large-language models to evaluate the outcome of a conversation (which Eggemeier wryly acknowledged is kinda funny) as well as a quality-assurance tool and a random sampling of outcomes to verify if companies using Zendesk's software were actually able to solve problems on behalf of their customers.
• "We're finding out that we're agreeing in more than 90% of the cases it actually is a customer resolution," he said.

Zendesk rolled out the new pricing strategy last month along with several new AI agents for its software, which helps companies manage and respond to internal and external issues. Right now, it's a little like buying a package of 10 visits to the personal trainer, but the strategy will continue to evolve over the rest of the year, Eggemeier said.

• Depending on which tier they're using, Zendesk customers will get 100 free automated resolutions and have the option to buy packages of 100 after that.
• Other customers might decide to pay as they go, he said.
• Salesforce rolled out consumption-based pricing last year with the launch of its Agentforce product, which costs $2 per conversation, and while it's not clear whether or not that applies to any conversation or just successful ones, it does offer 1,000 conversations as part of its free tier.
• ServiceNow customers get access to differing levels of what it calls "assists" through their Pro+ or Enterprise+ subscriptions, and they can buy more as needed but the criteria for what counts as an assist varies from service to service.

Agentic AI adoption remains a work in progress: Salesforce doesn't expect to generate "meaningful" revenue from AgentForce until next year, and Eggemeier said about 10 percent of Zendesk's customers have adopted generative AI to some extent. Zendesk hopes that number hits 20% by the end of the year, he said, in part because AI use is becoming a mandate inside many companies.

• "I think there's a lot of top-down pressure right now in customers, that they see what the C-suite or the board thinks about agents and [those leaders] probably see two or three big opportunities" such as improving the efficiency of their software engineers or holding hiring costs flat, Eggemeier said.
• Shopify CEO Tobi Lutke told employees Monday that they'll have to prove AI can't handle a task or project for which they've requested more headcount or budget before the company approves that expense.
• And just as the smartphone redefined how companies interact with their customers, who demanded a more flexible user experience, Eggemeier thinks generative AI will have the same effect of resetting expectations.
• "We think customers don't care that we helped a company take five workflows out, or we don't think customers care that a human agent was able to do their disposition notes; what they care about is getting their problem resolved or helping them with an opportunity," he said.

Working the refs

Companies evaluating new technologies need some sort of arbiter to help them judge the performance, but tech companies have been playing games with benchmarks since the invention of benchmarks. Meta came under fire Monday after releasing a new version of its Llama large-language model that at first glance appeared to meet or exceed the performance of competitive high-end models.

However, the version used to generate the benchmark score was an "experimental" version that wasn't actually released to the general public over the weekend, according to TechCrunch. After discovering the bait and switch, LM Arena, which hosts models for benchmarking, posted that "Meta’s interpretation of our policy did not match what we expect from model providers,” and modified its results, The Verge noted.

Enterprise companies evaluating LLMs are unlikely to take results from a site like LM Arena at face value, but there's obviously some benefit to word-of-mouth chatter that positions a new model at the top of the competition. At least Meta's Llama models are available with open weights, which allows companies to get a better understanding of how they work.

Enterprise funding

Tailscale raised $160 million in Series C funding to build out what it calls "identity-based networking," based around a managed version of the open-source WireGuard VPN technology.

Cyberhaven landed $100 million in Series D funding for its "data lineage" technology, which helps companies protect their data by understanding how it flows across their networks and customers.

Redpanda scored $100 million in Series D funding as it builds out a platform for managing and securing data needed to run AI agents.

Portland's own Hydrolix raised $80 million in Series C funding to expand its streaming data lake technology to new use cases and cloud providers.

Anecdotes scored $30 million in additional funding to close its Series B round at $55 million, with plans to expand its compliance software with AI agents and new partnerships.

Adaptive Security landed $43 million in new funding for its security software, which was designed to combat AI-powered deepfakes and other social engineering attacks.

The Runtime roundup

Microsoft canceled another data-center project, this time in Ohio, as Microsoft's Noelle Walsh confirmed the company is "slowing or pausing some early-stage projects" in light of what appears to be the beginning of the end of the AI data center boom.

Meanwhile, the Trump administration plans to promote coal as the solution to the data-center energy crisis, because everything this administration wants to do was ripped from headlines in the early 20th century.

Nvidia will probably be able to dodge some of the worst tariffs on its DGX servers and HGX motherboards because most of the U.S.-bound shipments of those products come from Mexico, according to Bernstein Research and CRN.

Cloudflare acquired Outerbase for an undisclosed amount as it attempts to beef up its database portfolio.

Thanks for reading — see you Thursday!

This post was updated to correct the amount of funding raised by Tailscale.