AI Agent Platforms Compared: Agentic AI Platform Pricing for Enterprise Buyers

An AI agent platform is software that builds, runs, governs and meters autonomous agents against your company's own data and applications. It differs from a chatbot builder in one specific way: an agent is allowed to take actions in other systems, not just produce text, which is why every platform on this page meters actions separately from seats.

In practice the category has split into four buying shapes, and knowing which shape you are buying matters more than the price:

Traditional RPA vendors are now selling into the same category from a fifth direction, and they price differently again: UiPath prices a tier, then named licenses, then Platform Units metered per model call, which is three layers where most of the platforms above have one or two.

The mistake worth avoiding is buying a metered platform when you wanted a seat-priced assistant. Copilot Studio and Agentforce Flex Credits give you nothing on day one except a builder and a bill that starts when you publish. If what you actually wanted was for 200 people to be able to ask questions about company documents, a seat-priced product is the shorter road.

There is no single best AI agent platform, and any list that names one is usually selling it. The honest answer is that the platform holding your company data already wins, because the integration work you avoid is worth more than the entire seat price difference across this table.

That said, the choice does narrow quickly once you answer one question about your own stack:

Notice that price appears in only two of those rows. That is the real finding of this comparison. Across the seat-priced products the spread runs from $21 to $40, which on a hundred seats is about $23,000 a year, and a single failed integration project costs more than that.

Between $0 and roughly $550 per user per month at list, depending entirely on which unit the vendor sells. Here is what a realistic 100-seat US deployment costs on each platform, at published list prices and before any negotiated discount.

Two of those rows deserve a warning. The Copilot Studio $200 line is real but incomplete: $200 buys 25,000 pooled credits, and credits are consumed at 2 per generative answer and 5 per agent action, so a single agent handling 5,000 actions a month uses 25,000 credits by itself. Credits expire monthly and do not roll over. At 125% of prepaid capacity Microsoft disables custom agents outright, which is a genuine operational risk rather than an overage invoice.

The Glean line is derived, not published. Glean's own pricing page carries no price at any tier, so the only honest public reference point is Vendr's reported median of $98,890 per year across 174 purchases. Treat it as an order of magnitude, not a quote.

Every platform on this page uses one of four meters, and mixing them up is where budgets break.

Per seat. You pay for people who can use the product, whether they use it or not. Gemini Enterprise, Microsoft 365 Copilot, Amazon Quick, Agent 365 and Onyx all work this way. The cost is predictable and scales linearly with headcount, which finance departments like. The waste is real: a broad rollout typically sees a minority of seats do most of the work.

Per unit of work. You pay for what agents actually do. Copilot Studio meters Copilot Credits, Agentforce meters Flex Credits or Conversations, SAP meters AI Units, and Amazon Quick meters agent hours above its included allowance. This is cheap while volume is low and unbounded when it is not. The published rate cards:

Two break-evens are worth memorizing because they decide a real purchase. Agentforce Flex Credits beat the $2.00 Conversation price only below 20 actions per conversation, and the $5 metered Agentforce seat beats the $125 unmetered seat only below 1,200 actions per user per month, which is about 57 actions per working day. Salesforce will not support Flex Credits and Conversations in the same org, so this is a one-way door.

Per bot. Power Automate is the outlier and charges $150 per unattended bot per month, or $215 for a Microsoft-hosted one. That looks expensive next to $200 for 25,000 Copilot Credits until you do the division: $150 of credits at the agent flow rate of 13 credits per 100 actions buys roughly 144,000 actions, so the bot license only wins above that volume.

Flat platform fee. IBM watsonx Orchestrate is the only platform here that does not price per person at all. It charges $6,360 a year for the Essentials edition and $76,320 for Standard, whatever your headcount, up to a ceiling of 4,000 and 40,000 monthly active users respectively. That makes it the cheapest option on this page at any real scale (a $530 monthly equivalent at 100 seats, and $1.59 per employee per year at the ceiling) and the most expensive at ten seats, because there is no way to start small.

The flat model has a trap the others do not. IBM's two published editions are not a volume discount: Standard costs exactly twelve times more than Essentials while including exactly ten times the capacity, so a unit of work bought on Standard costs $1.06 against $0.88 on Essentials, 20% more. Capacity is also capped per person rather than pooled, at 50 messages per monthly active user on both editions, so upgrading buys more users and not more usage per user. The full arithmetic is in our watsonx Orchestrate pricing breakdown.

The fifth model is per outcome, and only one vendor on this table uses it. Intercom Fin charges $0.99 when it actually resolves a conversation and nothing at all when it hands the conversation to a human without resolving it. There is no seat for the agent, no credit pool to forecast, no included allowance and no platform fee. That is a materially different risk position from every other row here: on a seat or credit model you pay for capacity you might not use, and on an outcome model you pay only for work that completed. It also inverts the usual incentive, because a better-performing agent produces a larger invoice rather than a smaller one. Sierra prices on a similar outcome basis but does not publish a rate. The full breakdown, including the $9.99 lead qualification rate that sits ten times above the resolution price, is on our Intercom Fin pricing page.

Three of the fourteen publish nothing, and that is a purchasing signal rather than an oversight.

UiPath sits in a fourth category worth naming: partial disclosure. It publishes one price, $25 a month for its Basic tier, and marks both Standard and Enterprise as contact sales, so the only tier with a public number is the one an enterprise will not buy. What it does publish in full is the consumption side, and that is unusually transparent: 0.2 Platform Units per LLM call on a standard model, 0.4 on a premium model, 0.16 on a basic one, billed in 64,000 input-token increments. The unit rate is public while the dollar value of a unit is not, which lets you compare the relative cost of two designs without being able to price either. See UiPath pricing for the full rate card and the reported contract medians.

Moveworks sits in a fifth category, and it is the one most likely to save you money on any vendor: marketplace-only disclosure. Its own site publishes nothing, moveworks.com/pricing returns a 404 and the live pricing page is a quote form, yet its AWS Marketplace listing carries a real list price of $150 per user per year for the 1,000 to 2,500 user band on a 12 month contract. Verified September 3, 2026. The lesson generalizes: before you accept that a vendor does not publish pricing, check its AWS, Azure and Google Cloud Marketplace listings, because a marketplace listing has to name a number in a way a website does not. See Moveworks pricing for how that published rate compares to what buyers actually pay.

Glean publishes no price at any tier. Checked again on August 26, 2026: glean.com/pricing carries customer stories, connector counts and adoption statistics, and no dollar figure anywhere. Glean's own published numbers on that page are 275 or more connectors, 40 or more supported LLMs, 110 hours saved per user per year and 93% enterprise adoption inside two years. Zillow is cited as running more than 3,400 agents. All useful, none of it a price.

SAP does not price Joule separately. Joule is gated behind a RISE with SAP or GROW with SAP subscription, so the agents are not a line item you can buy. What you can meter is AI Units, at roughly $0.08 to $0.18 per action.

Oracle charges no incremental fee. Its embedded Fusion AI agents are included in the Fusion Cloud applications subscription, which makes them free at the margin and impossible to price standalone.

There is one more piece of price opacity worth knowing about, and it involves Google contradicting itself. Google's Gemini Enterprise editions documentation lists the Business edition seat range as "1-500 users." The Gemini Enterprise product page describes the same edition as "Up to 300 seats." Both pages were live on August 26, 2026. If your headcount sits between 300 and 500 and you are relying on Business, get the number written into the contract rather than trusting either page.

Outside this table, disclosure gets worse, not better

The fourteen platforms above are the horizontal agent platforms. The specialist AI customer service agents that compete for the same budget disclose far less. We probed nine of them on September 2, 2026 and only three returned a pricing page with any dollar figure on it.

Zendesk belongs in the partial-disclosure category alongside UiPath: it publishes the billing unit and states that with AI agents "you pay only for customer requests that were successfully resolved," but it does not publish what one automated resolution costs. Sierra is the clearest case of full silence, which matters because it is also the most expensive option most US buyers will shortlist. We break the reported figures and the per-contact arithmetic down in Sierra AI pricing.

Two related facts about Gemini editions matter for agent buyers specifically. Business is the only edition with no check against "Access the Agent Marketplace," so the cheapest paid Google seat cannot install partner-built agents at all. And Standard and Plus carry an identical check on every single capability row in Google's comparison table: the only documented difference between them is pooled storage, 30 GiB against 75 GiB per seat. Anyone buying Plus to unlock agent features is buying storage.

Four things, and the licensing one catches people out.

A connection to your data. Every platform grounds agents in company content through connectors, and connector coverage varies more than price does. Glean publishes 275 or more, Onyx 40 or more, and Google gates the "full data connector ecosystem" behind Standard, leaving Business with only "select, segment-relevant connectors."

A license that allows the action, not just the chat. This is the trap. In the Microsoft stack an agent grounded only in its own instructions and public web content runs on free Copilot Chat, but the moment it reaches tenant data it needs either a full Microsoft 365 Copilot license at $30 per user per month or pay-as-you-go billing against Copilot Credits. Microsoft Agent 365 has a similar two-gate structure: seeing agents through the AI Reader role is free, acting on them is the paid $15 seat.

Somewhere to put the identity. Agents that act need credentials, and the governance question of who those credentials belong to is now its own product category. Agent 365 exists precisely because enterprises ended up with agents from six platforms and no single registry.

A capacity ceiling you have actually calculated. On Copilot Studio, exceeding 125% of prepaid capacity disables custom agents rather than billing you for overage. Environments linked to a pay-as-you-go Azure meter are exempt from that enforcement, which is why teams with spiky workloads usually hold a smaller prepaid pack and route the surges to Azure.

Most companies evaluating agentic AI platforms do not need a platform yet. They need two or three agents doing specific jobs, and the fastest route to that is installing something prebuilt rather than standing up a builder, hiring for it, and discovering six months later that the first agent still is not live.

Every major platform now ships a catalog for exactly this reason. Google runs the Agent Marketplace, Microsoft merged AppSource and Azure Marketplace into a single Microsoft Marketplace in September 2025, AWS opened AI Agents and Tools in July 2025, and Salesforce renamed AppExchange to AgentExchange in March 2025. The pattern is consistent across all four vendors: build a platform, then admit that customers want to buy finished agents.

The practical test is volume. If you have one repeatable, high-volume job, buy a prebuilt agent for it and measure the result before you commit to a platform seat count. If you have a job nobody sells because it is specific to how your business works, then you need a builder, and the metering model matters more than the seat price. Our agent catalog is organized by the job rather than by the vendor, which is the way most buyers actually search.

One caveat on the build route that is easy to miss: Microsoft's Agent Builder and Copilot Studio declarative agents can never be listed on the commercial Marketplace. If distributing what you build is part of the plan, check the publishing rules before you pick the tool.