The agent tax just made your AI roadmap a three-landlord lease

ServiceNow shipped Action Fabric. DataDog capped its MCP server at 5,000 calls a day. SAP started blocking unendorsed external agents against a $200B install base. Workday and HubSpot moved to usage-based agent pricing. JPMorgan's Mark Murphy named it before the analyst notes did — "essentially a tax on customers using outside AI agents."

This is not five product teams arriving at the same idea by accident. Unmetered agent traffic wrecks capacity planning and per-seat pricing simultaneously. A tollgate fixes both in one control plane. The interesting question isn't whether the rest of enterprise SaaS follows. It's what your P&L looks like when they do.

The cost stack grew a layer while you were testing prompts

A week ago, the agent cost model had two lines: model inference and orchestration. Now there are three. Model providers are raising prices, not lowering them — GPT-5.5's headline 2x is somewhere between 49% and 92% net once you account for completion-token effects, depending on your prompt-to-output ratio. Anthropic killed flat-rate subscriptions. GitHub is rebuilding for 30x agentic load, which is what metered pricing produces at scale.

Underneath that, the data layer is now metered separately. A planner that fans out three discovery calls before the real query — which is how most ReAct loops actually behave in production — burns DataDog's daily cap before lunch. I've watched this happen in a staging account. Trace one user prompt through ServiceNow's Action Fabric and you'll count six metered events for one piece of work: planner, three tool calls, a reflection step, synthesis. The pricing page shows a small per-action number. The invoice does not match the pricing page.

And the buyer's ROI doesn't cover any of it. Pete Stavros at KKR told Milken the actual AI earnings lift across the portfolio is 5%, not the 50% that underwrites trillion-dollar lab marks. Uber's CTO admitted publicly they "blew through" the AI budget after turning on agentic tools. a16z's seed cohort is spending $300K a year on agents in lieu of engineers, with Cursor bills exceeding salaries — token spend and headcount rising together, which is the exact opposite of the "AI replaces people" thesis that justified the spend.

Two landlords who can independently raise rent. A buyer measuring single-digit returns. That is not a margin compression scenario. That is the margin.

Anthropic already cut the deal you'll have to decide about

Claude Cowork shipped a first-class connector into ServiceNow's Action Fabric the same week the metering went live. The two-tier integration model is now visible: first-party agents get native pipes and presumably better pricing, custom agents get the standard API or get blocked. SAP's endorsement-only posture is the most aggressive version. Workday and HubSpot are softer but moving in the same direction.

The rational play for a team with custom agent logic is to wrap it inside a Claude Cowork session and inherit the preferred connector. That is an agent-platform tax paid to avoid the integration-layer tax. Pick which toll you prefer — but understand you're picking.

The counter-positioning is real. AWS CEO Matt Garman went on record that incumbents "trying to protect what they have could get into trouble." OpenAI is now free to serve any cloud after the Microsoft restructuring closed exclusivity through 2032. Grok 4.3 launched at $1.25 input / $2.50 output per million tokens — roughly half of Sonnet 4.6 with a 1M context window. The price floor moved while the access floor rose. Whoever builds the neutral metering and routing layer between model providers and gated enterprise data owns the Stripe-of-agents opportunity. That is the asymmetric bet to source right now.

What a builder actually does this week

The forcing function is instrumentation. Every agent tool-call needs a structured metric emitting source system, tier (free API vs. Action Fabric premium vs. metered MCP), and estimated dollar cost per call. Without it, the first signal of monthly burn arrives as a ServiceNow invoice, and by then the conversation with finance is already a defensive one.

Three concrete moves this sprint, in order:

Add per-vendor quota and rate-limit awareness as first-class config in the agent orchestrator. DataDog's 5K daily cap silently throttles to look like model errors. Cache reads aggressively — reads count against the same meter as writes, and caching is the cheapest intervention available.

Run the ablation nobody has run yet: task success rate on the standard API tier versus the Action Fabric premium tier for your top three workflows. Until someone measures the capability delta against the per-action cost, you're paying premium prices on a hunch. The number you're looking for is dollars per successful task, broken out by vendor and integrated SaaS. Most eval harnesses don't measure this. That gap is the bottleneck.

Inventory every pipeline that depends on SAP, ServiceNow, Workday, or HubSpot data and flag anything routed through external agents. SAP's endorsement-only policy means some of those break without warning, and the teams that map their exposure now will negotiate from a position the teams that don't will not have.

The boring reading of this week is that enterprise software vendors found a new SKU. The honest reading is that the AI feature defending your ARR is the feature compressing the margin you were defending ARR to protect — and the people who get to keep arguing about feature parity are the ones whose unit economics already hold.