Enterprise agent platforms, Nov 2026 to May 2027: Agent Builder ends November 30. A standard action is still $0.10.

The meter is the product

An enterprise agent platform is the thing a company buys so an agent can work inside the business: a CRM, a service desk, a suite, a cloud runtime, an HR system. The sales name is often a digital worker or an AI employee. The contract is not. On the pages read November 12, 2026, seven vendors were current. Agentspace is the old name on Google’s own blog: the editor’s note says it is now part of Gemini Enterprise. OpenAI’s AgentKit post is still live, and it now says the visual builder is leaving. Workday, ServiceNow, AWS, Microsoft, and Salesforce each have a current pricing or billing page. A vendor that did not survive that check is not in the table. ¹⁰¹⁸²⁰

Salesforce is the vendor that prints the most dollars, so it is the chart. The $2 conversation and the $2 help resolution are the same sticker on different units. The help resolution is 400 credits. The conversation is not credits at all, and the pricing page says the two buying models will not be supported in the same org. Slackbot, which the Q2 earnings release folded into Agentforce ARR, is on the rate card at 40 credits for a standard action and 200 for an advanced one. That is $0.20 and $1.00 at the $500-per-100,000 list. ¹²³

What $0.10, five credits, and 25 assists actually buy

Salesforce sells the action, the seat, and a credit pool that does not roll

The pricing page defines an action as one function the agent executes: updating a record, summarizing a case, answering a product question, or running a custom prompt or flow. Each action is metered on its own. Standard and custom actions are 20 Flex Credits. Voice actions are 30. Unused Flex Credits do not roll into the next term. If you pass the entitlement, the page says there is no overage penalty: the contracted rate is billed monthly in arrears. That is the opposite of a hard stop. ¹

Three seat prices sit beside the credits. The employee add-on is $125 per user per month and says usage for employees is unmetered. The Industries add-on is $150. The Agentforce user license is $5 per user per month and says it requires Flex Credits, so the $5 is the door and the credits are the meter. Foundations is $0 and names the builder, Prompt Builder, Agent Script, Coworker, and Vibes. It does not print a production credit allowance. Max Edition includes the add-on and 2.75 million Flex Credits per org per year. The page does not print the Max subscription price. Divided by 20, the allowance is 137,500 standard actions. At $0.005 a credit it is $13,750 of list-price credit inside whatever the edition costs. It is not a gift of $13,750. ¹

The Coworker user license is on the same page as a per-user SKU. The copy read November 12 lists included searches, the org’s security and access policies, and 30 MB of indexing per user per month. It does not print a dollar. This report does not supply one. Pre-Commit, a second buying model, says it will be widely available later this year and does not name a day. Pay-as-you-go and Pre-Commit apply to Flex Credits. The page says conversation pricing, the add-ons, and Agentforce 1 Editions are not on those two models. ¹

Copilot Studio’s five-credit action is the cheap line. The reasoning example is not.

Copilot Credits are the unit. Learn’s billing page, read November 12, prices a classic answer at 1 credit, a generative answer at 2, an agent action at 5, and tenant-graph grounding at 10. Agent-flow actions are 13 credits per 100 actions. A reasoning model adds a second meter: 10 credits per 1,000 billed tokens, on top of the feature rate. Cached input that the model only reads is not billed. Input it writes, and output, including reasoning tokens, are billed. Computer-using agents are billed at the agent-action rate, and that use is not included in the Microsoft 365 Copilot license. ⁷

Microsoft’s own follow-up, on Claude Opus 4.7, bills 14,100 tokens. At 10 credits per 1,000 tokens that is 141 credits, $1.128 in a pack and $1.41 at pay-as-you-go, before the feature rates for the turn. The same page’s new-conversation variant is 461 credits, $3.688 in a pack, before the generative-answer line. The page says the token counts are illustrative. Eleven Salesforce standard actions, at $0.10, are $1.10. That does not buy the 141-credit line. It is a different meter. ⁷⁸

Enforcement is specific. Learn says a tenant on the prepaid model is enforced at 125% of prepaid capacity, and the environment’s agents are then blocked, including a conversation already underway. An environment that bills the overage to pay-as-you-go is not blocked when the pack runs out. The licensing guide prices Microsoft 365 Copilot at $30 per user per month, billed annually, and says employee-facing Copilot Studio use is included when the person has that license and the agent runs as that user. Fair-use limits apply, and Microsoft says it can change them. The guide does not print the number at which fair use stops. ⁷⁸

ServiceNow and Workday publish the count and withhold the dollar

ServiceNow’s Assist Overview, effective August 24, 2026, defines an assist as the unit for Now Assist skills. The pool is contractual, measured daily, and resets on a 12-month anniversary. Production and sub-production both consume assists. If usage passes the entitlement, ServiceNow may invoice the excess. The overview does not print a dollar per assist. For agents, the rows that matter are the workflow sizes: less than 4 tool actions is 25 assists, 5 to 8 is 50, and 9 to 20 is 150. One action is one tool invocation. Past 20 actions, a new workflow starts. A single information lookup in Agent Studio is 1 assist. So is one AI Gateway call from a registered MCP client to a third-party MCP server. ²¹

Workday’s Flex Credits page says you pay for the work the AI completes, not every token, and not headcount. The worked example is the Self-Service Agent: 1 credit for instant information retrieval, 5 credits for autonomous task completion. Credits are an annual purchase, with complimentary credits sized to the company, and the page says most agents can be tested free in non-production. It does not print a dollar per credit. A quote that answers “what does a Workday agent cost?” with one number is not using this page. ²²

OpenAI’s remaining enterprise agent is a token line, and the builder that is leaving is dated

The AgentKit post is dated October 6, 2025. The update dated June 3, 2026 says OpenAI is winding down Agent Builder and Evals. From November 30, 2026 they will not be available on the OpenAI platform. The same paragraph recommends the Agents SDK for workflows that should continue as code, and Workspace Agents in ChatGPT for natural-language prompting. ChatKit and the Connector Registry are described in the original post. The update does not say the registry is shutting down. The registry, in that post, includes third-party MCP servers. ¹⁰

Workspace Agents are in research preview for ChatGPT Business, Enterprise, Edu, and Teachers. The help article says they are off by default at launch for Enterprise, and admins can turn them on. The product page lists role-based access, audit logs, approval gates before a send or a record change, and a central admin console. The token rate card, for eligible token-based Enterprise agreements, prices Workspace Agents on GPT-5.6 Sol at $4, $0.40, and $20 per million input, cached input, and output tokens. It says the Sol promotional pricing is available at least through November 21, 2026, and that the promotion does not apply to legacy metering. Customers still on a credit card are told to use a different rate card. This page did not price that card. Enterprise seat price is not a number on these pages. It is custom. ¹¹¹²¹³

The one place two vendors almost share a unit

Most of the table cannot be drawn on one axis. Gateway calls can, with a label on the arithmetic. AWS prices AgentCore Gateway at $0.005 per 1,000 API invocations, and the pricing page names ListTools, InvokeTool, and Ping. That is $5.00 per million invocations. Google’s Agent Platform pricing page says one Agent Compute vCPU-hour, at $0.085, corresponds to 15,000 API calls or authorization requests through Agent Gateway during agent execution. A million calls is 1,000,000 ÷ 15,000 × $0.085, which is $5.666…, rounded here to $5.67. Model tokens on Google are a separate bill. AWS model use is also separate. The nearness of $5 and $5.67 is a coincidence of two published ratios. It is not a market price for “an MCP call.” ¹⁴¹⁶

The protocol calendar is a different report. Revision work, the registry, and the July 28, 2027 removal eligibility sit on the MCP outlook and the twelve-month flagship. This table is only what these seven vendors print about their own products. A blank cell is a gap in the sources read on November 12. It is not a compatibility claim. ¹⁸

The badge is ahead of the invoice

Microsoft already issues an agent identity. The tenant admin cannot mint its token.

Copilot Studio’s identity article says new agents get an Entra Agent ID: a service principal with an Agent subtype. Agents created before that existed, in May 2026, keep a legacy app registration and “will be migrated” later. The article does not name the migration day. For a standard-harness agent, the ID is provisioned when the agent is created. For the GitHub Copilot harness, it is provisioned at first publish, not when the draft is saved. A Microsoft-owned blueprint creates the identity. No one in the tenant, including a tenant administrator, can generate tokens with it. ⁹

Scopes on that ID describe Power Platform connector access. They are not Graph scopes such as Mail.Read. Conditional Access on the agent identity is enforced today only when the agent runs in Teams. Other channels still call connectors through the existing Power Platform flow, so an admin can see the scopes and the policy is not yet evaluated on those calls. Custom connectors, MCP servers, and REST API tools do not add API permissions to the Entra Agent ID. Deleting the agent removes the ID. Every ID counts against the tenant’s directory-object quota. ⁹

Salesforce’s hosted MCP still runs as the user. The agent’s own badge is still TBA.

The developer doc for exposing an Agentforce agent as an MCP tool says execution runs in the authenticated user’s context. Sharing rules, field-level security, and the Einstein Trust Layer apply. Only agents built in the Agent Script Builder can be exposed, and they are not exposed until someone configures them. That is the current tool path: the person, not a separate agent principal. ⁶

A help article published September 17, 2026 describes the other path and still uses the future tense. It says Salesforce is targeting November for agent registration and a billing model in which each external agent gets its own identity, narrower permissions, and a meter. Every successful call over MCP or a direct API would be a Headless Platform Interaction. The multipliers “haven’t been announced yet,” the cell is TBA, and “agentic usage isn’t metered.” Salesforce will give 30 days’ notice before a number is added and metering begins. The rate card updated November 3, 2026 prints the same TBA under Customer 360 Platform, Headless Platform Interaction. Sandboxes, scratch orgs, and Developer Edition would not draw that meter. Traditional integrations stay on their current pricing. As of this read, the article has not been replaced by a “now available” line. This report does not turn “Targeting November” into a missed launch, and it does not turn it into a shipped one. ²⁵

The multipliers for this usage type haven't been announced yet, so it is currently listed as “TBA” and agentic usage isn’t metered.

AWS charges the token request unless the agent is already on Runtime or Gateway

AgentCore Identity is how an agent reaches AWS resources and third-party tools, on a user’s behalf or on its own with pre-authorized consent. The pricing page says customers who use it through Runtime or Gateway pay nothing extra for Identity. Everyone else pays $0.010 per 1,000 successful OAuth-token or API-key requests for a non-AWS resource. The same page’s example names Amazon Cognito, Okta, and Microsoft Entra ID as identity providers. Gateway also has a Policy meter at $0.000025 per authorization request, with the first 100 temporal policies per policy engine free of that charge. Natural-language policy authoring is $0.13 per 1,000 input tokens. ¹⁴

Two dates on that pricing page have already passed, and the sentences have not been rewritten. Short-term memory ingestion is $1.00 per GB as of October 6, 2026, replacing $0.25 per 1,000 events through October 5. The v2 committed baseline is still “launching by October 2026,” on a page read November 12. v2 consumption CPU is $0.1276 per vCPU-hour in the worked example. Do not read the October clause as either shipped or cancelled. Read it as a sentence the page has not updated. ¹⁴

Google split the seat from the platform, and named an agent ID in April

The app FAQ says Gemini Enterprise app and Gemini Enterprise Agent Platform are purchased separately. The app is per seat per month and needs a user license. The platform, which the FAQ calls the evolution of Vertex AI, is pay-as-you-go on a Google Cloud account. The FAQ does not print the seat price. Business edition is for teams up to 300. Standard and Plus add higher quota, unlimited seats, third-party agents, and a path to bring external agents in through A2A. Identity controls on Standard and Plus include Cloud IAM, third-party identity systems, and document-level access so a search or an agent does not leak a file the person cannot open. Audit logging, CMEK, and VPC Service Controls are listed for those editions. All editions include Model Armor on prompts and responses. ¹⁷

The April 22, 2026 post is the one that names Agent Identity: a unique cryptographic ID “for complete traceability and auditing,” with Agent Gateway as the control for interactions between agents and data. The same post says both no-code and pro-code agents are managed through one model for identity, security, and auditing, and that Agent Platform supports A2A and MCP. The pricing page read November 12 does not give Agent Identity its own SKU. It prices Agent Compute, and it says Memory Bank billing on the new structure “will commence on September 1st, 2026.” That sentence is still in the future tense on a November read. Same rule as the AWS October line: do not upgrade a stale tense into a fact. ¹⁶¹⁸

What is dated between now and May 2027

Announced means a linked page names the day. Likely means a vendor sentence aims the work at this stretch and does not lock the day, and this report says why that reading is fair. Watch means a later check should confirm the item or drop it. A month with no dated enterprise-platform event stays empty in the announced column. The general six-month outlook covers other product filings. This table is these platforms only. ^3410141523

The Q2 definition is the part a December headline will skip. Salesforce’s August 26, 2026 release says Agentforce ARR exceeded $1.5 billion, up over 240% year over year, and that effective Q2 FY27 the figure includes Slackbot and Headless 360. The same release says Agentforce and Data 360 ARR together reached nearly $3.9 billion, up over 210%. The company defines that combined line as the annualized recurring value of active Data 360 and certain AI subscription agreements. It is not a count of $0.10 actions. 7.0 billion Agentic Work Units had been delivered to date, 3.2 billion of them in the quarter. The release does not define an Agentic Work Unit as a Flex Credit. This report does not either. ³

What to ask before the order form

The next six months do not contain a published price cut for these platforms. They contain one removal, one earnings date, two conferences, and a set of meters that were already live on November 12. A buyer who waits for May to “see where pricing settles” is waiting on a month this source set leaves blank. The decision is which meter you can audit, and which identity the log will name when the agent acts. ¹⁹¹⁰

Five questions, in the vendor’s nouns

Which unit is on the quote: action, conversation, credit, assist, token, or vCPU-hour? Which of Salesforce’s buying models is it, and is Pre-Commit actually orderable yet, or still “later this year”? If the agent is employee-facing, is the person on a $125 Salesforce add-on, a $5 user license plus credits, or a $30 Microsoft 365 Copilot license, and does computer use fall outside that inclusion? What does the audit record show as the actor, and is the meter for a separate agent identity TBA, free through the runtime, or already an Entra object? What happens at empty: arrears at the contracted rate, a block at 125% of the pack, or an invoice for assists? ^157891421

Builders have a narrower cut. If the workflow is code, OpenAI’s own update says the Agents SDK, not Agent Builder, after November 30. If the workflow is a shared natural-language agent, Workspace Agents are the product the same sentence names, and they are off by default for Enterprise until an admin enables them. If the tool is MCP, price the host. Salesforce’s current hosted path runs as the user. ServiceNow’s gateway call is 1 assist. AgentCore’s InvokeTool is on the Gateway line. Google’s correspondence is 15,000 calls per $0.085 vCPU-hour, plus tokens. The MCP brief is the protocol. This page is the invoice. ^61013141621

A token bill for a model you call yourself is still a third invoice. The AI agent cost calculator uses official per-token prices. It does not price a Flex Credit, an assist, or a Copilot Credit. The twelve-month flagship holds the filing and the regulation dates. The label “AI employee,” and why the software is not a legal employee, is a separate guide. ¹²