Salesforce CEO Marc Benioff recently came closer to answering a multibillion-dollar question when he said seat-based pricing – with some caveats – was becoming the norm for its AI agents after flirting with pricing based on consumption and per-conversation payments.

While there is much debate about how users will get money back from investing in AI agents, the stakes are also high when it comes to how they pay for them. Gartner has forecast that agentic AI could drive around 30 percent of enterprise application software revenue by 2035, surpassing $450 billion, up from 2 percent in 2025.

Vendors supporting the pillars of enterprise for most large organisations – names like SAP, Oracle, Workday, and Infor – all see AI agents as a big part of their plans.

Salesforce has bet big on AI agents to boost productivity among its users, and it expects commensurate returns. “We’re talking about 3x, 4x the ability to multiply the monetisation on customers because, by the way, they’re getting 3 or 4x or 10x more value from our products,” Benioff told investors last week.

But with concern raging about how many jobs AI might replace, the CRM giant was presented with a dilemma should it continue with seat-based licensing: the more customers benefit from AI’s promised productivity gains, the less they might end up paying as they cut staff.

Benioff said customers were keen on its Agentic Enterprise License Agreement (AELA) introduced in October, which offers a flexible menu of options and reusable credits in a single package, based on the number of seats. That’s now the preferred option for licensing.

“When we first started with Agentforce, we were talking about [charging] so much per conversation. It was this type of pricing, maybe transaction-based pricing, usage-based pricing, but customers have pushed for more flexibility,” he said.

The switch back to seat-based pricing is a sign that, in the current phase, users are not buying AI agents to replace people, and that they want some certainty in the pricing model before investing in the much-promoted technology, according to Jan Cook, senior software licensing expert with Gartner.

“Customers still adopt a cautious approach to investing in GenAI because of the unpredictability of pricing models,” he said. “Vendors have wrestled with pricing, trying to find a balance between profiting and achieving a return on their own investments, with ensuring that pricing provides value to customers.”

While users might welcome the return to seat-based licensing to ensure costs are predictable and help sketch out return on investment calculations, there are some caveats for most vendors, he warned.

“They still have an underlying currency called AI units. Most vendors are moving in that direction. They’ve got seat-based, per-user pricing, but that normally comes with some secondary license metric. For example, a license might come with a provisioned limit of credits. If that’s exceeded, you would have to top up those credits. In other cases, seat-based licenses come with ‘fair use limits,’ meaning it’s not truly unlimited, but it allows vendors some ability to control the impact of potential overuse of GenAI solutions.”

The approach allows tech departments to get what they want in terms of more predictable costs, avoiding signing blank checks to spiralling usage until they understand how employees respond to the new tech. Meanwhile, vendors retain underlying control to ensure software is not overused and they end up losing money on calls to large language models, which they ultimately pay for at the back end.

“The big question always was, are we looking to save money by replacing humans with AI, or are we looking to arm the business with AI technology to make ourselves more effective at what we do, and thus giving us a competitive advantage? It seems to be the latter. GenAI pricing has been around long enough: if there was going to be a mass seat shrinkage issue that would have emerged [by now] and customers would be renegotiating contracts on lower seat volumes. We haven’t seen that to any significant extent,” Cook said.

Both more predictable costs and viable use cases are spurring the steady growth of AI agent implementations. “It’s not a sharp spike in adoption, but what we see is a very predictable, steady adoption trajectory for most vendors, which was slow in the beginning, but it is steady,” he said.

Others see a rocky road ahead for the adoption of AI agents. Forrester predicts that using AI for financially driven layoffs can backfire: 55 percent of employers regret laying off workers because of AI. Meanwhile, more people in charge of AI investment expect it to increase headcount (57 percent) than decrease it (15 percent) over the next year.

Earlier this year, the analyst firm warned that for vendors, “the era of experimentation is over, and the era of monetisation has begun,” while users face an increased risk of vendor lock-in, and “the unglamorous work of process redesign,” if they are to make gains from their AI investments.

In this context, it was hardly reassuring when Benioff told investors that the introduction of AI agents into its software represented a “very high margin” opportunity for his business.

Lisa Singer, Forrester principal analyst, said that the downside of the seat-based approach to AI licensing is that if end user uptake is below expectations, “the buyer might be paying for users that are minimally using the products.”

In the near term, vendors will continue to use a range of commercial models for AI agents, she said. Hybrid models will include a flat fee plus usage-based pricing on actions or workflows. “Copilots will continue as seat-based since their usage is tied to humans and generally use is predictable. Workflow automation agents will migrate to pricing based on usage or outputs or outcomes,” she said.

Users can also try to get a fair share of any benefits from AI without the vendors’ hunger for margin taking an unwarranted slice of the pie.

“Users can ensure they share the gains through outcome-based pricing where the vendor and the customer share in the productivity gains, revenue uplift, and so on. This pricing model will require a great deal of information sharing to track the gains and attribute them to the agentic/agent offering. Transparency and a great deal of trust will be required on both sides,” Singer said.

This is the challenge for in-house tech leadership. Not only do they have to implement a new technology with a seemingly infinite number of things that could get in the way of the business value – user behaviour, data quality, software failure, model hallucinations, take your pick – they also have to keep an eye on a meter that has been rewired and most likely skewed toward the vendor’s margins. And they said AI would make life easier.