AIAgentROI 2026 Q2 Pricing Convergence Index:
Seven Weeks of Vendor Pricing, Synthesized

Last updated: May 5, 2026 · Methodology and limitations · Citations

1. Why We Did This

AI agent pricing is changing faster than the analyst reports tracking it. Vendors are announcing new models mid-quarter. Rebrands are obscuring what was previously a clear product category. Token costs — the floor beneath every platform — fell so fast in early 2026 that pricing structures anchored to last year's economics are already uncompetitive.

Starting March 11, 2026, we began logging pricing for eight major AI agent platforms on a weekly basis. Each week, a researcher visited each vendor's public pricing page, noted the listed tiers, recorded any language changes or model-tier additions, and flagged any earnings disclosures that updated the picture. Over 7 weeks, we accumulated 56 pricing snapshots across 8 vendors — a corpus that lets us say something specific about the direction of this market, not just its current state.

The 8 vendors we tracked were chosen to span four segments: enterprise platform vendors with existing large software relationships (Microsoft, Salesforce, Google, ServiceNow), dedicated CX-layer vendors (HubSpot Breeze, Zendesk), workflow automation (UiPath), and the model-layer API baseline (OpenAI). This spread lets us identify patterns across the stack, not just within one category.

What we found surprised us in a few ways. The speed of model convergence on "outcome-based pricing" was faster than we expected — moving from a niche Zendesk experiment in mid-2024 to a genuine three-vendor standard by April 2026. The timing of platform repositioning across the three majors (Google, Salesforce, Microsoft) within a 9-day window in late April/early May was not something any single news cycle captured clearly. And the continued gap between published per-seat prices and what enterprises actually pay — a 30–60% discount in most cases — means the published "list price" landscape we tracked is a proxy for direction, not a reliable input for budget modeling.

We built the Pricing Convergence Index to give buyers a single, repeatable framework for evaluating pricing structure — not just pricing level. Our methodology and its limitations are documented in the appendix.

2. Methodology

How We Collected This Data

We tracked 8 vendors across the following segments:

• Enterprise platform layer: Microsoft Copilot + Agent 365, Salesforce Agentforce, Google Gemini Enterprise Agent Platform (formerly Vertex AI), ServiceNow Now Assist
• CX and GTM layer: HubSpot Breeze AI, Zendesk AI
• Workflow automation: UiPath
• Model API baseline: OpenAI API (per-token, used as a cost floor comparison)

For each vendor, we logged the following data points weekly:

1. Primary pricing model (per-user, per-action, per-outcome, or hybrid)
2. Published list price per unit at the lowest commercial tier
3. Whether any pricing model changes were announced or implemented that week
4. Any earnings disclosures affecting ARR or pricing guidance
5. Changes to product bundling or product naming that affected price comparability

We supplemented weekly observations with earnings disclosures and press releases. Salesforce Q4 FY2026 results (announced late February), Anthropic's Series G filing (February 12), OpenAI's funding round (March 31), Google Cloud Next announcements (April 22), HubSpot's pricing change announcement (April 2, effective April 14), and Microsoft's Agent 365 GA (May 1) all informed our analysis during the observation window.

The Pricing Convergence Index

We built a scoring framework — the AIAgentROI Pricing Convergence Index — to evaluate vendors on four dimensions beyond raw price. The rationale: a vendor that charges $0.50 per resolved conversation with fully transparent pricing and no minimum commitment is genuinely more accessible than one charging $0.40 but only through a sales-negotiated annual contract. Published price alone doesn't capture buyer experience.

The four dimensions we score (each 0–2.5, total maximum 10):

• Pricing transparency (0–2.5): Is pricing published on a public page? Is unit pricing clear? Are there hidden tiers or "contact sales" gates?
• Outcome-orientation (0–2.5): Does the primary billing unit reflect actual value delivered (resolutions, completions) vs. proxy metrics (seats, calls, tokens)?
• Granularity (0–2.5): Can buyers pay only for what they use? Are there consumption tiers that don't require annual minimums?
• Fairness to small buyers (0–2.5): Is the entry price accessible for teams under 50 people? Are there startup or self-serve tiers?

Scores reflect our assessment as of May 1, 2026, based on public pricing pages. We will re-run this scoring each quarter and publish updates.

3. Key Findings

Finding 1: Three Pricing Models Now Dominate, But Only One Is Growing

The per-user model dominated enterprise software for 30 years and still dominates the platform layer. Microsoft Agent 365 went GA on May 1 at $15/user/month standalone, or bundled into the new Microsoft 365 E7 suite at $99/user/month alongside E5, Copilot, and Entra. ServiceNow Now Assist remains on a per-user-per-year model, running $150–$250/user/year as an add-on to base platform subscriptions. Both reflect a vendor priority: protecting existing seat-license revenue while adding AI capability at the margin.

Per-action pricing — where each interaction or workflow step is billed — is where the market was heading 12 months ago. Salesforce Agentforce launched in this model, initially at $2 per conversation. Agentforce's Flex Credits system charges per-action within a conversation, with credits priced at $0.15 each. UiPath remains in this model conceptually, though its per-robot-per-year structure ($8,000–$10,000/unattended robot/year at list) makes it more akin to per-capacity than per-action in practice.

The fastest-moving category is outcome-based pricing. The table below shows where each vendor stood at the start of our observation window (March 11) vs. May 1.

Outcome-based pricing adoption went from 1 of 8 vendors (Zendesk, which introduced it in August 2024) to 3 of 8 by May 1, 2026, with Salesforce's hybrid tier adding partial exposure. We expect 4 of 8 to offer outcome-based pricing as a primary or co-primary option by end of 2026 — most likely Microsoft, which has signaled agent-level billing in future E7 versions, or ServiceNow, which faces competitive pressure from Zendesk in ITSM resolution workflows.

Finding 2: Outcome-Based Pricing Is the Fastest-Growing Model, and the Adoption Curve Is Steepening

The most significant event in our observation window was HubSpot's April 14 transition. HubSpot announced on April 2 that Breeze Customer Agent and Breeze Prospecting Agent would move to outcome-only pricing, effective April 14. The move cut the nominal per-unit price roughly in half — from $1.00 per conversation to $0.50 per resolved conversation — while changing the billing trigger entirely. Under the old model, you paid for every inbound conversation the agent touched. Under the new model, you only pay when the agent resolves the issue without human intervention.

That's not a modest tweak — it's a fundamental shift in which party bears the risk of agent underperformance. Under per-conversation pricing, a 40% resolution rate means you pay full price for 60% of interactions that ultimately required a human anyway. Under outcome-only pricing, those failed interactions cost you nothing.

HubSpot's framing was explicit about this logic: "Most AI tools aren't priced on outcomes because they can't promise consistent results." The implied positioning is clear — they're betting their agents are reliable enough that outcome-only pricing is self-funding. That's a bold public claim, and it puts measurable pressure on Salesforce and Microsoft to either match the model or justify why their proxy-metric billing (per-conversation, per-user) better reflects value.

Zendesk had already proven the model workable since August 2024, charging approximately $1.50 per automated resolution. According to SaaStr's February 2026 analysis, Zendesk's starting per-resolution rate is $1.50, or $2.00 on pay-as-you-go, with a starter tier included in existing plans. Sierra AI — not in our panel but worth noting — operates exclusively on outcome-based pricing and only collects revenue when an issue resolves without human intervention.

Finding 3: Token Cost Compression Is Collapsing the Platform Floor

The numbers tell the story. GPT-4-level capability cost approximately $30 per million tokens when GPT-4 launched in 2023. By early 2026, equivalent capability runs at approximately $0.06 per million tokens on the most cost-efficient providers — a 500x price collapse over 3 years. More relevant for quarter-over-quarter buyers: LLM API prices dropped approximately 80% between early 2025 and early 2026.

We used OpenAI API published pricing as a week-by-week cost floor baseline throughout our observation period. As of April 2026, the frontier model landscape looks like this:

The implication for platform vendors is significant. If a customer-service interaction uses roughly 3,500 tokens (a 2,500-token context plus 1,000-token response), the raw model cost of that interaction at GPT-4.1 mini rates is approximately $0.0028. Zendesk charges $1.50 for a resolved interaction — a roughly 535x mark-up over raw model cost. HubSpot charges $0.50 — still a 178x mark-up.

These mark-ups are partially justified: integration, tool use, orchestration, SLA guarantees, security, and audit trails all cost real money. But as model costs compress further, the defensible value of the platform layer has to be increasingly non-model. That's exactly what we see in Finding 4.

Finding 4: The Three Largest Vendors Repositioned in a 9-Day Window

The sequence, as we recorded it:

• April 22: Google rebranded Vertex AI to the Gemini Enterprise Agent Platform at Google Cloud Next 2026. The rebrand was cosmetic in some respects — the underlying APIs are the same — but the strategic signal was deliberate: Google wants to compete at the agent-platform layer, not just the model-API layer. The new platform bundles Agent Studio (low-code creation), Agent Hub (deployment and monitoring), and Agent Governance (compliance and audit traceability) under a unified brand. Critically, Google also announced an open marketplace from day one, with Oracle, Salesforce, ServiceNow, Adobe, and Workday already integrated — a move that positions Google's governance layer as cross-platform infrastructure.
• April 15: Salesforce, at TDX 2026 on April 15, announced Headless 360 — described as the most significant architectural shift in the platform's 25-year history. Every feature, workflow, and business logic module is now exposed via API, MCP (Model Context Protocol) tools, or CLI. The practical effect: Agentforce agents can now orchestrate any external model (Claude, GPT-5, Gemini, LLaMA, Mistral) through Salesforce's Trust Layer, rather than being constrained to Salesforce-native models. This is a deliberate repositioning away from "Salesforce is a CRM with AI" toward "Salesforce is an agentic orchestration platform that happens to own CRM data."
• May 1: Microsoft Agent 365 went generally available, bundled into the new M365 E7 "Frontier Worker Suite" at $99/user/month. The product's GA simultaneously established Microsoft's agent control plane as a per-user seat (not per-agent or per-action), and positioned it as a suite play — the value is in the bundle with E5, Copilot, and Entra, not in Agent 365 alone.

Three major repositioning moves in 9 days is not coincidence — it's an industry responding to a shared inflection point. The Model Context Protocol standardization, falling token costs, and Zendesk/HubSpot's outcome-pricing moves are all pushing the same direction: the model layer is commoditizing, and the defensible moat is now the orchestration, governance, and data layer above it.

Finding 5: ARR Growth Rates Are Decoupling from Traditional SaaS Curves

Salesforce Agentforce reached $800M ARR with 169% year-over-year growth and 29,000 deals closed in a single quarter (50% quarter-over-quarter deal growth). For context: Sales Cloud took 5+ years to cross $500M ARR. Service Cloud took 4+ years. Agentforce did it in 18 months. That's not just fast — it means the sales motion is different. Expansion inside existing Salesforce customers, not new logo acquisition, is the primary driver.

Anthropic's run-rate revenue surged from $9B to over $30B between Q4 2025 and early April 2026 — a roughly 3x jump in under 4 months. That rate of growth is not driven by new customer acquisition; it's driven by consumption expansion inside existing enterprise accounts. Enterprises that started with Anthropic Claude for one workflow are expanding to many.

OpenAI closed its $122B funding round at an $852B valuation on March 31. At $2B in monthly revenue (up from $1B per quarter at end of 2024), OpenAI's trajectory implies an annualized run rate above $24B. Enterprise revenue now accounts for 40% of the total.

The aggregate picture is that AI agent revenue — across both the model layer (OpenAI, Anthropic) and the platform layer (Salesforce, ServiceNow's Now Assist, which crossed $600M ACV in Q4) — is compounding at rates that make it a different category of software. The pricing implications are not purely favorable: platforms locked into annual contracts at current prices are capturing revenue based on pricing economics that may look very different in 12 months, as models continue to compress and outcome-based billing becomes standard. Vendors who own the outcome-to-billing workflow today will have a structural advantage as the market matures.

4. The AIAgentROI Pricing Convergence Index

We score each vendor 0–10 across four dimensions (each weighted equally at 2.5 points). The index is designed to answer a specific question: regardless of how much the vendor charges, how well does their pricing structure serve a buyer who wants to pay fairly for what they get?

This is a distinct question from "who is cheapest." A vendor could be inexpensive but opaque (hard to predict costs), or expensive but completely transparent and outcome-oriented (predictable, fair). The index captures the structural quality of the pricing model, not its level.

* Salesforce's score reflects the current hybrid state. The Flex Credits model (opaque, complex) scores 4.5; if Salesforce fully migrates to a published outcome-based tier, we estimate its score would rise to approximately 6.5. We will update this score as pricing documentation evolves.

Score commentary

HubSpot Breeze AI (8.5) — The highest score reflects HubSpot's April 14 pivot to fully outcome-based pricing, which simultaneously increased transparency (you know exactly when you'll be charged), outcome-orientation (you pay only for resolved conversations), and granularity (there's no minimum commitment for SMB plans). The 1.5 on "fairness to small buyers" reflects that while pricing is accessible, HubSpot's Breeze agents are currently limited to customers on Marketing Hub Pro/Enterprise or Sales Hub Pro/Enterprise — not their most accessible plans.

Zendesk AI (8.0) — Similar logic. Zendesk pioneered this model and has had 18+ months of iteration. The slightly lower score on granularity (1.5 vs. HubSpot's 2.0) reflects that their resolution pricing stacks on top of existing Zendesk Suite subscriptions in a way that makes total cost of ownership harder to calculate upfront.

OpenAI API (8.0) — Scores high on transparency (fully published, per-token) and granularity (pure consumption, no minimums). Scores lower on outcome-orientation (1.0) because tokens are a proxy metric that doesn't directly track whether the agent accomplished anything useful. Scores 2.0 on small-buyer fairness because the API is accessible to anyone with a credit card, though agent orchestration adds implementation overhead.

Microsoft Agent 365 (5.5) — Pricing is published ($15/user standalone), but the primary billing unit is a seat, not an outcome or action. The bundled E7 pricing makes it genuinely hard to evaluate the marginal value of Agent 365 alone. Scores low on outcome-orientation (0.5) because the current GA only supports the "on-behalf-of" agent model — true autonomous per-agent billing is still in Frontier preview. Small-buyer fairness is limited by Microsoft's enterprise-centric sales motion.

Salesforce Agentforce (4.5–6.5) — The most complex pricing in our panel. The Flex Credits system requires buyers to understand credits-per-action, actions-per-use-case, and voice surcharges — a multi-variable calculation that even sympathetic analysts have described as unnecessarily convoluted. The legacy per-conversation $2 tier is being phased out. Small buyers are essentially excluded — most Agentforce deployments require existing Salesforce CRM contracts. We flagged this scoring uncertainty explicitly.

ServiceNow Now Assist (2.0) and UiPath (2.5) — Both score near the bottom for the same reason: pricing is opaque, requires a sales conversation, is sold only through annual contracts, and is structured around seat counts rather than outcomes or actions. ServiceNow's pricing model involves bundling Now Assist into tier upgrades (Pro to Pro Plus) at 25–40% cost uplift per user — a structure that obscures the actual per-AI-interaction cost entirely. UiPath's unattended robot licenses at $8,000–$10,000/robot/year are accessible primarily to organizations with established IT procurement teams and multi-year planning horizons.

5. What This Means for Buyers

Our 7-week observation produced five practical conclusions for organizations currently evaluating or renegotiating AI agent contracts.

• 1
  Negotiate outcome-based pricing now, while vendors are still establishing the norm. Outcome-based pricing is not yet the default — 5 of 8 vendors in our panel still bill by seat, action, or robot. That means buyers have negotiating leverage to request outcome-based terms even from vendors who don't publish them. The framing is straightforward: "If your agent performs as claimed, outcome billing costs you nothing more and costs us less when it underperforms. Why wouldn't you offer it?" Salesforce and Microsoft both have the technical ability to implement outcome billing — neither has published it because it hasn't been demanded. RFPs written in the next 6 months that explicitly request per-outcome pricing terms will have more success than those that don't.
• 2
  Avoid annual lock-in at current prices. Our tracking showed significant pricing compression — not just in model costs but in platform economics — between March and May 2026. A contract signed today at current list prices locks in pricing from a moment when token costs are 10x higher than they may be in 18 months, and when outcome-based billing was still a minority practice. The standard advice to "lock in multi-year pricing" runs exactly counter to the direction of this market. Prefer monthly or quarterly commitments where possible. If annual commits are unavoidable, negotiate repricing triggers tied to published API cost benchmarks.
• 3
  Multi-vendor sourcing is rational at current market conditions. 41% of marketing organizations run at least one SDR agent, and multi-agent orchestration involving 3+ distinct vendor agents now accounts for 22% of production deployments — up from 1% in 2024. The standardization of MCP (Model Context Protocol) and the explicit cross-vendor integrations announced at Google Cloud Next 2026 (Oracle, Salesforce, ServiceNow, Adobe, Workday all in Google's Agent Hub on day one) mean inter-vendor orchestration is becoming a first-class use case, not an integration headache. Buyers should structure contracts to preserve the right to orchestrate across vendors.
• 4
  Write your RFP in outcome language, not feature language. Most enterprise AI agent RFPs we've reviewed request capabilities ("can it do X?") rather than outcomes ("what does it cost per resolved ticket?"). This invites vendors to compete on feature lists rather than economics. An RFP that asks vendors to quote a price per successful resolution — defined precisely, including what counts as resolution, what attribution rules apply, and what the SLA is for contested resolutions — will produce both more useful competitive data and more aligned vendor behavior post-contract. It also positions your organization to shift to outcome billing as vendors adopt it.
• 5
  Re-evaluate pricing every 6 months, not annually. The changes we observed in 7 weeks — a major vendor pivot to outcome-only pricing (HubSpot), three platform repositioning announcements in 9 days, and continued token cost compression — suggest that the AI agent pricing landscape is changing faster than annual contract cycles assume. Organizations that review their AI agent pricing at annual renewal only will systematically overpay relative to organizations that treat pricing re-evaluation as a quarterly discipline. Build in a formal 6-month pricing review trigger in any multi-year contract you sign in 2026.