SAP, ServiceNow, and Workday Are Building Toll Booths for AI Agents

Here's a number that's making enterprise software CFOs nervous: a single AI agent can fire off 50,000 API calls in a month. Under the per-seat pricing that has governed enterprise software for 30 years, those 50,000 calls cost you zero extra dollars. Under the new model that ServiceNow, SAP, Workday, HubSpot, and Datadog are building right now, they could cost you thousands.

JPMorgan analyst Mark Murphy has a name for it: a tax on customers. Five of the largest names in enterprise software are installing checkpoints between their platforms and the AI agents that need to access those platforms. Some are metering usage and charging per action. Some are blocking access entirely unless you use their own tools. One is simply setting hard caps that trigger sales conversations when you exceed them.

Enterprise software built itself on a headcount model. You pay per user. More employees using the platform means a bigger invoice. For three decades, that logic held. It was simple to audit, easy to negotiate, and predictable enough that both vendors and buyers could plan around it.

AI agents have upended the assumptions underneath that model. An agent isn't an employee. It doesn't need a seat. One agent automating a finance workflow that used to require eight people can eliminate eight seats from a vendor's invoice without replacing any of the actual work the platform is doing. In fact, the platform might be doing more work than ever, processing more transactions, returning more results, and executing more workflows, but with zero incremental seat revenue to show for it.

The disruption is clearest in high-transaction workflows. An AI agent handling employee IT requests might interact with a service management platform hundreds of times per day per employee it serves. An agent monitoring code deployments might ping an observability tool thousands of times per hour. These interaction volumes are orders of magnitude higher than what a human user generates, and they're happening right now, today, in companies running AI pilots.

The $200 billion enterprise software market runs on subscription contracts that were never designed for this pattern. Vendors know it. They're moving to fix it before their renewal cycles crater.

How Five Enterprise Platforms Are Charging for Agent Access

ServiceNow moved first and most explicitly. At Knowledge 2026 in May, the company unveiled Action Fabric, an integration layer that sits between external AI agents and ServiceNow's platform data. Every agent that wants to read a record, trigger a workflow, or write a result inside ServiceNow has to pass through it.

The pricing logic is metered by action. COO Amit Zavery described it directly: the company will measure how often customers access the Action Fabric, meter that usage, and charge customers for it. Anthropic's Claude is the launch partner, which means if you're running Claude agents against your ServiceNow data, you're already in the pricing model ServiceNow is building out.

The rollout comes alongside a much bigger announcement at Knowledge 2026. ServiceNow unveiled an expanded Autonomous Workforce featuring AI specialists across IT, HR, finance, legal, procurement, and security. These aren't assistants that help humans do work. They complete entire business processes without human involvement. Its internal IT specialist resolves cases 99% faster than a human agent. Docusign is targeting 90% autonomous IT ticket resolution. The City of Raleigh reports a 98% deflection rate on employee requests.

When those numbers become normal across ServiceNow's customer base, the platform will be doing exponentially more work per employee. Action Fabric is how ServiceNow captures revenue from that work instead of losing it to headcount reduction. Zavery put it plainly: "Advisory AI has run its course. Enterprises need AI that senses, decides, and securely acts."

SAP's approach is more restrictive. In April 2026, the company updated its API policy to prohibit third-party AI agents from interacting with its systems outside of SAP-endorsed architectures. The policy bars autonomous or generative AI systems that plan, select, or execute sequences of API calls without official authorization. In practice, that means if your company has built AI workflows using Claude, GPT-4o, or any other non-SAP agent, those agents cannot access your SAP data under the new policy. SAP's own Joule agents remain permitted. Third-party automation that doesn't route through SAP's approved architecture does not.

SAP CEO Christian Klein said customers would not pay to access their own data. That statement drew attention, but the policy itself didn't change after it. What the policy actually does is steer customers toward SAP-built agents as the only authorized route through the gate. This is different from ServiceNow's strategy, and arguably more aggressive. ServiceNow is building a toll road and charging to use it. SAP is closing the road to non-SAP vehicles entirely.

For enterprise buyers with deep SAP deployments, the stakes here are high. SAP's systems often hold the most sensitive and transaction-critical data in a company: ERP, supply chain, procurement, financial records. Restricting agent access to that data to SAP-only tools is a significant constraint on where enterprise AI strategies can go.

Workday CEO Aneel Bhusri has been less specific about implementation, but very clear about intent. Charging for AI agent access, he said earlier this year, offers the company considerable financial upside. The mechanism will be usage-based, tracking how often agents interact with the platform's HR and finance data. An AI agent handling benefits management, onboarding, and payroll exception processing might interact with Workday hundreds of times per employee per month. If agents multiply the number of platform interactions per employee by a factor of 10, the revenue potential expands accordingly, even as headcounts shrink.

HubSpot is further along on implementation. Its AI agents run on a credit system with real production pricing data. Until recently, each Customer Agent conversation cost 100 credits, equivalent to $1.00. Starting April 14, 2026, HubSpot shifted to $0.50 per resolved conversation, aligning revenue more directly with outcomes. A per-conversation model charges you whether the agent solved the problem or not. A per-resolution model charges you only when value is delivered. Credits function as an abstraction layer between a subscription and consumption billing: customers buy a block upfront for predictability, then burn against usage for alignment with actual agent activity.

Datadog's approach is more mechanical. Its model context protocol server, the interface that lets AI agents query Datadog's observability data, allows up to 5,000 daily requests or 50,000 monthly requests before usage slows or stops. For companies running agents that continuously monitor infrastructure, the cap math is significant. 5,000 daily requests sounds like a lot until you consider an agent checking Datadog metrics every 30 seconds across a production environment. At that cadence, a single monitoring agent burns through the daily limit in roughly 41 hours. A fleet of monitoring agents hits the wall in hours. The hard cap functions as a forcing mechanism to bring high-volume AI users into negotiated contracts.

The Buyer-Side Costs and the Analyst Read

For enterprise IT teams building AI agent infrastructure in 2026, these emerging models represent a new category of cost that needs to show up in total cost of ownership calculations.

The risk is the classic cloud scaling problem in a new form. Consumption costs that look manageable at pilot scale become significant at enterprise scale. A Claude agent that makes 1,000 ServiceNow API calls per day during a 10-user pilot might make 500,000 calls per day when rolled out to 5,000 employees. If Action Fabric charges even a fraction of a cent per call, that's a substantial line item.

Most enterprise AI projects are still in pilot or early deployment. The interaction volumes that will define actual costs aren't visible yet. Companies building their agent strategies right now are committing to architectures before they can see the pricing that will apply when those architectures scale.

SAP's restriction model adds a different kind of cost: investment replacement. Companies that have built agent workflows using third-party tools now face either compliance risk or the cost of rebuilding those workflows inside SAP-endorsed architectures. Neither option is cheap.

For a more in-depth look at how enterprise AI governance and rollout decisions interact with these access control developments, the Enterprise AI in 2026 guide covers the governance and cost-planning frameworks that teams need to navigate this landscape.

JPMorgan's Mark Murphy called these controls a tax on customers, and the framing is useful because it names where the money flows: from enterprise buyers who want to use third-party AI agents to the SaaS platforms those agents need to access.

The counterargument is that some of this repricing reflects genuine value alignment. AI agents do consume platform resources. They make API calls, process queries, trigger workflows. If a single agent is performing what 10 employees used to do, charging per action rather than per seat may be a more honest accounting of the value the platform delivers.

What's harder to defend is SAP's approach. Restricting third-party agents outright doesn't meter access, it monopolizes it. Customers who have invested in AI infrastructure built on non-SAP tools aren't facing a new price; they're facing a prohibition. The effect is that SAP's own Joule agents benefit from the policy by eliminating competition at the access layer.

ServiceNow's case is more balanced. Action Fabric is presented as a universal integration layer, not a tool to block competitors. The pricing is metered rather than binary. For Workday and HubSpot, the credit models are somewhere in between: transparent about structure, linking charges to outcomes rather than volume regardless of value delivered.

The real test is whether enterprise buyers push back hard enough to shape the terms. Enterprise software procurement has a long track record of negotiating vendor pricing when the buyer has enough negotiating weight. Companies with large enough contracts can negotiate Action Fabric pricing, credit rates, and cap structures as part of renewal deals. Companies with smaller contracts will likely take whatever the standard pricing is.

What Enterprise IT Teams Should Plan For Now

The practical implications for teams building AI agent strategies in the current environment come down to a few concrete areas.

Map your agents to the platforms they touch. Every enterprise platform that an AI agent needs to access is a potential pricing gate. ServiceNow, SAP, Workday, HubSpot, Datadog, and a growing list of others are all moving in the same direction. Get a clear picture of which platforms your agents depend on before committing to an architecture at scale.

Model consumption at scale, not at pilot. Take your pilot-stage interaction volumes and multiply by the expected production rollout ratio. Then find out what the pricing model is for that volume. Action Fabric pricing, HubSpot credit rates, and Datadog request caps all need to appear in your AI infrastructure cost projections before deployment, not after.

Watch your contract timing. The vendors building these access pricing models are doing so in advance of their renewal cycles. If your ServiceNow, SAP, or Workday contract comes up for renewal in the next 12 to 18 months, the renewal negotiation is when you have the most negotiating power to push back on unfavorable access terms or lock in favorable caps.

For SAP specifically, get clarity on your compliance posture now. If you're running third-party AI agents against SAP data, the April 2026 policy update may already put you in violation. The question of whether SAP will actively enforce the policy against existing customers versus applying it to new deployments is unresolved, but it's not a question to leave sitting.

This year's ServiceNow Knowledge announcements and the parallel developments at SAP are closely connected to what ServiceNow was already building on the auditability and governance side with Project Arc, which adds another layer of context to why the company is moving fast on both access control and metering.

The vendors moving early on access pricing have an advantage: they get to define the category before buyers build expectations around different terms. ServiceNow's decision to make Action Fabric the mandatory entry point for third-party agents gives it structural power, similar to what payment infrastructure providers have when they become the default layer between buyers and sellers.

For AI agent developers and the enterprise teams deploying them, the landscape is shifting from agents that can access anything with an API key to agents that pay platform access fees at each gate. That shift will push buyers toward AI infrastructure vendors that negotiate platform access on their behalf, and toward agent designs that minimize platform calls by doing more per interaction rather than querying constantly.

It will also create pressure on AI agent platforms like Anthropic, OpenAI, and others to negotiate enterprise access agreements that bundle platform costs for their customers. The question of whether an enterprise pays ServiceNow directly for Action Fabric access or pays an AI vendor who bundles that access into its product pricing is already being worked out in enterprise sales conversations.

Every time a new technology pattern creates interaction volumes that existing pricing models can't capture, the incumbents build new gates. It happened with mobile data, it happened with cloud API pricing, and it's happening now with AI agents. The specifics change. The pattern is reliable.

The gates are already under construction. The question now is whether enterprise buyers will negotiate the terms before they've signed contracts that lock them in.