Per-Action Pricing: Examples & Companies

What is it

Per-Action Pricing is a billing unit where each discrete action taken by an AI agent or automation is metered — common in browser automation and agentic workflow tools. Unlike tokens (which meter compute) or seats (which meter headcount), the action unit tries to meter work done: a browser step completed, a data row enriched, a meeting brief generated, or a 15-minute block of autonomous coding.

The unit emerged as automation platforms needed a billing granule that non-technical buyers could understand without knowing anything about the underlying LLM calls or compute cost. Clay was an early adopter, separating its pricing into platform Actions (orchestration, under $0.01 each) and Data Credits (marketplace data cost), so buyers could reason about what it costs to run a workflow independently from what it costs to buy the data. Bardeen applied a similar idea to browser automation: credits map to output rows (standard rows at 1 credit each, enrichment rows at 3 credits), so the bill tracks the lead list delivered rather than the compute consumed.

What makes this category interesting — and challenging — is the wild variance in what “one action” means across vendors. At Cognition, one Agent Compute Unit represents roughly 15 minutes of active autonomous engineering work, a unit designed to price a full agentic task cycle rather than a single step. At Rox, an Agent Action meters a specific sales-agent task — account research, contact discovery, a meeting brief — with the per-action cost floating with the complexity of that task and the AI models it calls. At UiPath AI, the platform’s Intelligent Automation engine meters actions in Platform Units (0.2 Platform Units per chargeable action in its IXP product), a sub-unit nested inside the broader per-seat and per-robot licensing structure. That four-way divergence on the definition of “one action” means the per-action unit is the most definitionally contested billing primitive in this corpus.

How it works

The mechanical question is always: where does the action boundary sit? Vendors answer at three different levels of abstraction.

The pricing formula varies by vendor, but the common structure is:

Bill = (action pool or tier fee) + (actions consumed × per-action rate)

For Clay, the two-meter version is: Bill = Actions base fee + (Data Credits consumed × data credit rate). Actions are a fixed capacity tier; Data Credits are the variable, rollover pool. For Bardeen, the flat-fee version is: Bill = plan fee, with credits consumed until the pool runs out (no overage — you simply stop). For Rox, the renewal-based version is: Bill = base fee × number of times the pool resets per month.

Worked numeric example — Clay, a mid-size GTM team. A Growth team at the $495/month base tier gets 40,000 Actions and 6,000 Data Credits per month. A workflow that enriches 500 prospects — each requiring 1 Action plus 12 Data Credits (phone + firmographic data) — consumes 500 Actions and 6,000 Data Credits, exhausting the included credit pool in a single run. Topping up at the 10,000-credit tier costs $460/month, bringing total spend to ~$955 for that month. The Data Credits are the bill driver; the Actions cost pennies.

Worked numeric example — Bardeen, a lead-enrichment workflow. Bardeen’s own pricing page example: scrape 10 profile links, extract contacts, verify emails with an enrichment step (3 credits each), qualify 4 leads with AI (1 credit each) = ~64 credits consumed in one run. A Premium plan includes 1,000 credits/month, which supports ~15 runs of that workflow before the pool empties. Unused credits expire monthly with no rollover.

Companies using this

Seven in-corpus companies bill in actions: Clay and Bardeen are the canonical GTM/automation examples, while Cognition (Devin), Rox, Automation Anywhere, UiPath AI, and Writesonic each apply the unit at a different granularity — from browser-step to 15-minute agent-time-block.

Patterns observed

Actions are a wrapper around cost opacity, not a replacement for it. Across all seven companies, the action is an abstraction that hides the actual compute and data costs from the buyer. Clay makes this unusually explicit: it separates orchestration Actions (near-zero, for the platform work) from Data Credits (the real variable cost of buying data from 150+ providers). Most vendors collapse both into one unit — which buys simplicity at the cost of transparency. When Rox says “Agent Action cost scales with task complexity and AI model used,” it is describing a floating per-action rate that can’t be quoted as a single number.

The action-boundary decision is the biggest pricing lever. Vendors who define “one action” at a low altitude (Bardeen’s output row) have a legible, predictable meter — buyers can count rows. Vendors who define it at a high altitude (Cognition’s 15-minute ACU) have a meter that more faithfully represents business value, but that is harder to model before running a workflow. Cognition addressed this by keeping the ACU as the enterprise unit while wrapping self-serve buyers in a quota + on-demand-credit structure that softens the unpredictability.

Feature gating and action gating are converging. Rox made the shift cleanly visible: its 2024 beta gated features by tier (a traditional SaaS approach), and by 2026 it had removed all feature gates, making Agent Action volume the only upgrade lever. That means buyers upgrade purely because they want more agent work done — not to unlock a capability — which is the action model operating at its purest.

RPA vendors are retrofitting actions onto an older robot-based pricing frame. Automation Anywhere and UiPath AI both list “actions” as a billing unit alongside the older attended/unattended bot licensing. For both companies, actions appear as a consumption add-on or sub-unit within a larger tier + seat + robot structure, rather than as the headline meter. This reflects how enterprise RPA pricing evolves incrementally — agentic AI work is being metered in actions, but the full bill still runs through the bot licensing scaffold.

Counterexamples & variants

Writesonic is the closest thing to a false positive in this cohort. Its billing_units field includes actions, but in practice, actions are a minor footnote — Growth-plan buyers get an Action Center trial (5 off-page + 5 on-page optimization actions per month), and agentic workflow runs are bounded by a small monthly trial allowance (10 → 50 → 100 by tier). The core value metric at Writesonic is “AI answers tracked daily” across AI search engines, not actions per se. Including actions as a billing unit is technically accurate for the small slice of the product that uses them, but actions don’t drive the bill the way they do at Clay, Bardeen, or Rox. Buyers evaluating Writesonic should weight the daily-answer tracking quota as the real consumption dimension.

The rollover mechanics reveal how confident vendors are in their action definition. Clay’s Data Credits roll over up to 2× the monthly limit, acknowledging that enrichment work is bursty and customers shouldn’t be penalized for buying ahead of a campaign. Rox’s Agent Actions expire at cycle reset with no rollover — which makes sense for always-on monitoring agents that consume actions continuously, but creates a “use it or lose it” dynamic that punishes seasonal sellers. Bardeen takes the same hard expiry approach. The contrast signals a real strategic difference: vendors with rollover trust buyers to smooth their own usage; vendors without it want actions consumed in each cycle’s window, which compresses burn and accelerates renewal.

The human-in-the-loop variants. Automation Anywhere explicitly prices attended bots (human-supervised) lower than unattended bots (fully autonomous). This attended/unattended split is itself a form of action-level pricing: the “action” the bot takes carries a higher price tag when it runs without oversight, because the vendor is absorbing all the reliability risk. No other vendor in this cohort makes that distinction explicit in the price card.

What this means for buyers vs vendors

For buyers

Benchmark your specific workflows before comparing prices, not the headline action rate. A “1 credit per action” plan from Bardeen and a “1 Action per enrichment” plan from Clay can look identical on the surface and diverge 10x in cost once you account for how each vendor counts a multi-step enrichment. The choosing the right usage metric guide recommends running your three highest-volume workflows through each vendor’s counting rules as a literal first step in evaluation.

Watch for rollover policies: they reveal whether you’ll be forced to use up actions on low-value tasks at month-end or whether you can bank capacity for campaign bursts. For predictability, prefer vendors who publish a per-action rate or a per-action dollar ceiling. For scale, prefer vendors (like Clay) whose action cost doesn’t float with an undisclosed model or complexity variable. And read the free tier carefully — at Rox, a full-featured Starter plan with 2,000 actions/month is a genuine product trial because all features are included; at other vendors, the free tier is a crippled demo.

For vendors

The action is the most defensible unit in agentic AI pricing when you can publish the counting rule clearly, but it demands a worked example on the pricing page that lets buyers model their actual workflows. Bardeen’s 64-credit example (scrape, enrich, verify, qualify, export) and Clay’s per-model credit rate table are both examples of this done right. If you can’t write that worked example because your per-action cost floats with too many variables, consider whether your unit is actually an action or a capacity quota. Separating orchestration cost from data/model cost (Clay’s two-meter split) solves the floating-cost problem by putting the variable cost in a separate, rollover pool — a technique covered in the usage-based pricing fundamentals and the tracking and metering usage events guide. Finally, if your agents act continuously — monitoring, researching, alerting in the background — budget controls that let buyers cap daily action burn before the pool empties are table-stakes; without them, always-on agents can consume a month’s allotment in a long weekend.