Enterprise AI agents in 2026: the platforms compared, by what you already run

The platforms worth a shortlist in September 2026 fall into four groups by what you already run: systems of record, work suites, automation platforms and clouds. Each entry gives what you must already run, how a person approves or stops what an agent does, what the vendor says about your data, the price in the unit the vendor charges, and one limit the vendor documents.

Meronym builds custom AI agents, sells none of the platforms in this list and has no referral or partner deal with the companies that do. We read every vendor's own pricing pages and documentation on 27 September 2026. We didn't run trials, so what each platform can do is what its maker documents, quoted and linked, and all prices are in US dollars.

What makes an AI agent an enterprise AI agent

An enterprise AI agent changes records in the systems an organisation runs on, such as its CRM, ERP, HR system or service desk, under an identity an administrator controls, with a person approving what matters and a record of everything it did. A chatbot answers a question about an invoice. An enterprise agent does the work: Oracle's Payables Agent, for one, can "match to POs/receipts" and "route for approval/payment" inside Oracle's finance system, where the approval rules and the audit trail already live.

Few organisations run them at scale yet. In McKinsey's 2026 State of AI survey, fielded in May and June 2026, 40% of respondents from organisations with more than $1 billion in revenue said they were scaling AI agents, up from 27% a year earlier; at smaller organisations the share stayed at 22%. In Gartner's 2026 CIO survey, only 17% of organisations had deployed AI agents, though more than 60% expected to within two years.

The risk sits in control more than in models. Gartner predicted in May 2026 that by 2027, 40% of enterprises will demote or decommission autonomous AI agents "due to governance gaps identified only after production incidents occur". So most of the questions below are about who can see and stop an agent, not about the model inside it.

How we chose

We considered 49 platforms and products: the 43 named on the pages that rank for this search and three related ones ("agentic ai platforms", "agentic ai platform" and "agentic ai solutions"), and six big-vendor platforms none of those pages names: SAP's Joule Agents, Oracle's AI Agent Studio, Microsoft's Foundry Agent Service, Glean, Claude Enterprise and OpenAI Frontier.

We didn't check 27 of them: eleven research projects and experiments, BabyAGI and Auto-GPT among them; six analytics tools; four developer tools (GitHub's Copilot cloud agent, the Copilot SDK, NVIDIA's NeMo Agent Toolkit and Kagent); and six named only by their own makers or in passing in a forum reply. We read the other 22 on their makers' own pages, and a platform stayed in the list only if it passed five rules:

1. It's generally available, not a preview or a limited release.
2. It states how you pay: a published price, a unit you pay by, or the package it comes in.
3. It changes something in a business system, rather than only answering.
4. Its documentation says how a person approves or stops what it does.
5. It's sold to whole organisations, as part of a system of record, a work suite, an automation platform, a cloud or an enterprise AI assistant.

Fifteen platforms passed, from thirteen vendors: Microsoft and Google each have two, a suite for every employee and a cloud for developers. OpenAI Frontier didn't: OpenAI launched it on 5 February 2026 "to a limited set of customers", and we found no statement since that it's generally available, and no price. Six open-source frameworks are in a table of their own, because they're code your developers run rather than a product you buy, so rules two and five don't fit them.

This isn't the guide for a small team choosing its first agent for one job, or for customer-service agents such as Fin, Zendesk's and Sierra's; our guide to AI agents for business sorts those by job. Where a vendor's page left something out, the entry says so. The last entry, a custom agent, is ours, and it carries the same fields as the rest.

Nine questions to ask before you buy enterprise AI agents

Every entry below answers these as far as the vendor's own pages do.

1. Which system does it change, and under whose identity? An agent acting as the signed-in user can do only what that person can. An agent with an identity of its own needs its own permissions and an owner.
2. What must we already run? ServiceNow's autonomous IT agents come only with ITSM Prime, its top IT package. Gemini Enterprise takes your own and third-party agents on Standard, Plus and pay-as-you-go but not on Business, and Google says the pay-as-you-go edition is "currently available to a limited group of customers". Microsoft says Agent 365 works best with E5.
3. What's the unit, what counts as one, and what does one cost? Actions, credits, model calls, runs, seats and messages don't convert into each other, so ask how many units one of your own requests uses.
4. What happens when the units run out? Microsoft switches custom agents off unless you've added a pay-as-you-go meter, UiPath refuses production runs, ChatGPT Enterprise pauses advanced features, self-serve Claude Enterprise stops, and Salesforce and SAP bill the excess.
5. How does a person approve or stop an action, and is that the default? Claude asks before any action by default, Glean's web app and Microsoft's Cowork pause before steps that write or send, ServiceNow's agentic workflows ask the user to confirm by default, SAP requires a person's approval on high-risk processes, and Workday's Talent Management Agent keeps a person on every output. On the rest, approval is a step you set up, and in Copilot Studio's agent flows and Google's Agent Development Kit that step is still in preview or experimental.
6. Where is the record of what it did, and who can read it? UiPath and SAP say every agent action is logged. Ask how long the log is kept, and whether your auditors can read it without the vendor.
7. Where do our prompts and records go, and does anyone train on them? Eleven of the thirteen vendors here state in writing that your data won't train the models behind their agents, though some say so only by default or only of outside providers' models. ServiceNow goes the other way: it shares data with its own AI development programme until an administrator opts out. Where your data is processed and how long it's kept vary more; each entry's Data line gives what the vendor states.
8. Can it govern agents from other vendors? Nine of the thirteen vendors here offer a way to, Amazon's in preview. Ask which vendors' agents it can find, and whether it can stop one or only list it.
9. Do our other vendors' rules let it in? SAP and Workday now route outside agents through their own gateways.

Enterprise AI agents inside your system of record

If one system holds the records a job changes, its maker's agents are the first place to look. They act under the permissions, approvals and audit trail that system already keeps. What they do less well is reach outside it.

Salesforce Agentforce: best for work that starts and ends in Salesforce

What it does. Salesforce's agents work inside your Salesforce org, for service, sales and employees. Its service agent, Help Agent, "acts instead of just answering": its actions "manage cases, schedule appointments, and update orders". Agentforce Engagement works new leads by email and passes interested ones to a rep.

Needs you to run. Salesforce on Enterprise, Performance or Unlimited edition; for service, Service Cloud's pricing page sells Agentforce for Service from its Core edition, $195 a user a month. Flex Credits also need Salesforce Foundations, which is free.

Control. Help Agent hands a customer to a person with the full context, and the Engagement agent can be set so a rep approves every email before it's sent.

Data. Prompts sent to outside models go through Salesforce's Trust Layer, under which those models neither store them nor train on them.

Price. Flex Credits at $500 per 100,000, which is 20 credits, or $0.10, per action; or $2 per conversation for customer-facing agents. Help Agent can instead be billed at $2 per resolution, charged only when it resolves an issue without a person; that option needs Salesforce's self-service add-on licences, at extra cost. For employees, Agentforce add-ons cost $125 per user a month for unmetered use. No trial is stated.

For what it can't reach, see our Salesforce integration services page.

ServiceNow AI Agents: best for IT, HR and service teams already on ServiceNow

What it does. ServiceNow's agents "act autonomously to get work done" across IT, customer service and HR, inside the ServiceNow AI Platform. They work through flows, subflows, scripts and skills, and reach other systems over MCP. Moveworks, which ServiceNow finished buying in December 2025, now ships as a layer in each IT service management package.

Needs you to run. The ServiceNow AI Platform. In IT service management, the autonomous agents come with the top package, ITSM Prime.

Control. Agentic workflows ask the user to confirm by default, and ServiceNow's developer guide recommends that mode for tools that create, update or delete records; read-only tools can run on their own. AI Control Tower inventories "any AI agent, model, and MCP server from first or third parties", and ServiceNow says it can spot an agent working beyond its permissions and "shut it down in real time".

Data. ServiceNow shares customer data with its own AI development programme until an administrator opts out: its documentation says you "must opt out of data sharing" to stop it gathering customer data for its AI and data products. Each instance is opted out separately, which can take up to five business days, and GCC and self-hosted instances never share. The pages we read don't say where prompts are processed. AI Control Tower comes with governance workflows for the EU AI Act and NIST's AI Risk Management Framework.

Price. Not published, and the package is the unit: the agents come inside ServiceNow's Foundation, Advanced and Prime packages, quoted per customer ("Get Custom Quote"), and AI Control Tower is sold as a package through the account team. We found no trial.

For what it can't reach, see our ServiceNow integration services page.

Workday's agents: best for HR and finance work that already runs in Workday

What it does. "Workday's AI agents do real work across HR, finance, IT, and more", Workday says: a Self-Service Agent for employees' requests, and agents for HR service, recruiting, talent, scheduling, pay and payroll. Sana, now part of Workday, coordinates them from one conversation. The Talent Acquisition Agent sources and grades candidates while "leaving the hiring decision to recruiters".

Needs you to run. Workday. Its agents are bought with Workday Flex Credits.

Control. Workday says its agents act for employees and managers "all with appropriate human oversight and visibility into those agent decisions and actions", and its Talent Management Agent keeps "a human in the loop on every output". Its Agent System of Record manages each agent's skills, including agents from outside Workday, and every outside agent must route its calls through Workday's Agent Gateway.

Data. Workday says it never shares customer data to train third-party public models.

Price. Flex Credits, bought once a year in bulk and spent at the rates on Workday's rate card, last updated 2 September 2026. The Self-Service Agent uses 1 credit per action, and a task uses more actions the more complex it is. Other agents are metered by the piece of work: 6 credits to score a CV, 10 for a payroll question, 500 to review and redline a contract. The Flex Credits page still gives 5 credits per action for autonomous task completion, a rate the card doesn't list, so ask which one your contract uses. Each customer gets a yearly allotment of free credits based on company size, and testing outside production is free. Workday doesn't publish what a credit costs.

For what it can't reach, see our Workday AI agents page.

SAP Joule Agents: best for finance, supply chain and procurement work in SAP

What it does. Joule Agents "execute multistep tasks across SAP and non-SAP systems", drawing on SAP's process knowledge, and connect to other systems and agents over MCP and A2A. Joule Studio builds custom agents, in low code or in code with frameworks such as LangChain and LlamaIndex.

Needs you to run. An SAP cloud subscription, which includes Joule itself at no extra cost. Agents are Premium AI, paid for in AI Units.

Control. "All agent traffic flows through the agent gateway, so agents can only execute what the end user has permission to do directly." SAP says "High-risk processes always require explicit human approval", and that "workflow and change logs record all agent activity". SAP AI Agent Hub keeps an inventory of agents across your systems, SAP's and other vendors'. Outside AI agents may use SAP's APIs only through routes SAP endorses, under its API Policy.

Data. SAP says it doesn't share your data with outside model providers to train their models, but may use it to improve its own AI features where your contract allows.

Price. 0.02 AI Units per agent action for SAP's own agents in 2026, where an action is "a single step an agent takes to complete a task". AI Units are bought once a year. As you near your purchased volume you can switch agents off in SAP for Me; keep them running and the excess is billed at SAP's excess-use rate. SAP doesn't publish what an AI Unit costs, and we found no trial.

For what it can't reach, see our SAP AI agents page.

Oracle AI Agent Studio: best for Oracle Fusion customers, at no extra cost

What it does. Oracle ships "the 1,000-plus AI agents delivered through Fusion Applications" across finance, HR, supply chain and customer experience, among them the Payables Agent above, which reads invoices from email, portals, EDI and PDFs. AI Agent Studio is where you extend those agents or build agentic applications that run inside Fusion.

Needs you to run. An Oracle Fusion Applications subscription.

Control. Agentic applications run "while inheriting Oracle Fusion Applications' security, governance controls, approvals, and auditability controls", and Oracle lists human-in-the-loop approval workflows among AI Agent Studio's features.

Data. The agents work on Fusion's own records under Fusion's security and audit controls. Oracle said on 23 October 2025 that in Fusion "your data is not used to train the underlying LLMs", and its AI brief that for models hosted on Oracle's cloud, "all data stays securely within Oracle's cloud infrastructure".

Price. "Available at no additional cost" with Fusion. Fully custom agents, built from scratch rather than from Oracle's templates, are a separately priced offering on seat-based and employee-based pricing.

Agents for every employee, in the suite they already use

These put agents in front of every employee, inside Microsoft 365, Google Workspace, ChatGPT, Claude or Glean. They reach a whole workforce at once, and seat counts drive most of the bill.

Microsoft 365 Copilot, Copilot Studio and Agent 365: best for organisations that run on Microsoft 365

What it does. Microsoft 365 Copilot comes with ready-made agents for research, analysis and meetings. Copilot Cowork plans multi-step work across Microsoft 365, Dynamics 365 and Fabric. Copilot Studio builds your own agents that act through flows, APIs, MCP servers and more than 1,400 connectors. Agent 365, generally available since 1 May 2026, is Microsoft's "control plane for agents".

Needs you to run. A qualifying Microsoft 365 business or enterprise plan; standalone Copilot Studio needs an Azure subscription. Microsoft says Agent 365 "works best when using Microsoft E5 as a pre-requisite".

Control. Cowork pauses for your approval at checkpoints, such as before it sends an email, and Copilot Studio's agent flows can add multistage human and AI approvals, in preview. Agent 365 covers "agents built across platforms and runtimes, including SaaS, cloud, and local agents", using Microsoft Entra for access, Purview for data and Defender for threats.

Data. Microsoft says business data in Copilot stays encrypted and isolated and is never used for AI training, and Cowork keeps it in your tenant.

Price. Microsoft 365 Copilot costs $30 per user a month, paid yearly, and Cowork is billed by use on top. Copilot Studio is sold in Copilot Credits: $200 for 25,000 a month, or $0.01 each pay-as-you-go, and an agent action uses 5. On prepaid credits, custom agents are switched off at 125% of the prepaid amount, unless a pay-as-you-go meter is set up to take the overage. Agent 365 costs $15 per user a month on its own, or comes in Microsoft 365 E7 at $99 per user a month with Copilot, E5 and Entra Suite. Business plans have a one-month free trial, and a Copilot Studio trial can build agents but not publish them.

For what it can't reach, see our Microsoft 365 AI agents page.

Gemini Enterprise: best for Google Workspace organisations, or one agent app across both suites

What it does. Gemini Enterprise gives every employee one app for search, chat and agents: Google's own, third-party ones and ones people build without code. Its Workflow Builder, generally available since 3 September 2026, runs steps and actions through connectors to Gmail, Calendar, Drive, Slack, Jira, ServiceNow, SharePoint and Outlook.

Needs you to run. Nothing for the Business edition, which works with Google Workspace or Microsoft 365. The pay-as-you-go edition needs an invoiced Google Cloud billing account.

Control. Workflows can include human approval steps, and agents can use only the connectors, tools and permissions an administrator provisions. Connecting your own MCP servers stays off until an administrator turns it on. Standard and Plus can bring in your own and third-party agents, and every edition manages its agents from one central view.

Data. Google says prompts and outputs aren't used to train its models or any other customer's, and Standard and Plus add data residency.

Price. Business from $21 per seat a month, for up to 300 seats; Standard and Plus from $30 through sales; or a pay-as-you-go edition with no seat fee and usage billed for tokens, memory, compute and storage, which Google says is "gradually rolling out and currently available to a limited group of customers". Each seat edition has a 30-day trial. On Standard and Plus, daily quotas are pooled across the edition's users: each Standard licence adds 160 assistant queries a day to the pool, and each Plus licence 200.

For what it can't reach, see our Google Workspace AI agents page.

ChatGPT Enterprise: best for organisations standardising on ChatGPT

What it does. ChatGPT Work is "an agent that can take action across your apps and files" and stay with a project until it's done, working in Slack, Teams, Drive, SharePoint, email, calendars and CRMs. Workspace agents are shared agents a team builds once and runs in ChatGPT or Slack.

Needs you to run. An Enterprise contract through OpenAI's sales team. API use is billed separately.

Control. Workspace agents: "For sensitive steps, like editing a spreadsheet, sending an email, or adding a calendar event, you can require the agent to ask for permission before moving forward." Enterprise administrators decide which tools and actions each group can use.

Data. OpenAI says it doesn't train its models on your business data by default, and Enterprise lets you set how long data is kept; its business page sends data residency to sales.

Price. Enterprise is priced by quote. Beyond each seat's included usage, credit-based agreements bill work in credits (agent mode uses 30 a message), and we didn't find what a credit costs; on token-based agreements, usage is charged in US dollars at your contracted rates. No trial is stated.

Claude Enterprise: best for handing whole tasks to an agent that asks before it acts

What it does. Claude Cowork "completes tasks you can steer from anywhere": it works across your files and tools, including Google Drive, Gmail, GitHub, Microsoft 365 and Slack, and hands finished work back for review. Anthropic is folding the Cowork name into Claude itself, starting with its Pro and Max plans. Claude Code takes coding tasks.

Needs you to run. A Claude Enterprise plan, with at least 20 seats self-serve or 50 through sales.

Control. "By default, Claude asks before taking an action." For connector tools that write, members approve each task unless an owner turns on a setting that's off by default.

Data. Anthropic says there's no model training on your content by default, and Enterprise can keep inference in the US, at 1.1 times the token price.

Price. $20 per seat a month, billed annually, plus all usage at API rates: the seat itself includes no usage. Anthropic lists trials for sales-assisted Enterprise, not for self-serve.

Glean: best for organisations whose knowledge is spread across many apps

What it does. Glean indexes your company's content across more than 275 apps, with each person's permissions enforced, and its agents plan, search and act on it. You can "Trigger agents from key events, route tasks between agents, and connect to external systems".

Needs you to run. Glean, on per-user Enterprise Flex seats.

Control. Write tools "Pause before they run, instead of executing immediately", by default in Glean's web app. For scheduled agents, an administrator chooses which write tools may run without confirmation, and the rest are blocked from background runs.

Data. Glean lists SOC 2 Type II, ISO 27001, ISO 42001, HIPAA and GDPR, and discounts its credits if you host it in your own cloud. Its model documentation says the model providers it contracts with "are configured with zero-retention commitments, so customer data is not stored or used for training by the model vendors"; on a Customer Key deployment, you set up the providers yourself.

Price. Per-user seats plus FlexCredits: an agent run uses a variable number, about 7 for a typical run and about 114 at the 90th percentile. Glean publishes neither the seat price nor the credit price, and we found no trial.

Automation platforms that now run agents

If your organisation already runs robotic process automation, its maker now sells agents that run inside the same processes, next to the robots and the people.

UiPath: best for organisations whose processes already run on UiPath

What it does. UiPath's platform runs agents, robots and people in one process, from structured steps to exception-heavy casework. Processes "survive failures, pause for approvals, and resume without losing state", and every agent action, human decision and policy applied is recorded at the level of the business transaction.

Needs you to run. UiPath's Standard plan or above, in UiPath's cloud or on your own premises. There's a 60-day free trial.

Control. A person's decision can be a step in the process, and which steps are agentic, fixed or human is set at design time and enforced at run time. UiPath offers "One control plane for every agent, native and third-party."

Data. UiPath says its AI Trust Layer never trains a model on your data and keeps nothing with outside model providers, and it offers data residency in eight cloud regions.

Price. Platform Units, charged per model call: 0.2 per call on standard models, 0.4 on premium ones, or 0.2 per run if you bring your own model; some plans count agent units instead. A checking model adds its own units to each run. UiPath's sales team quotes both the Standard plan and the price of a Platform Unit.

For what it can't reach, see our UiPath consultant page.

IBM watsonx Orchestrate: best for a published monthly price, on IBM Cloud, AWS or your own servers

What it does. IBM calls watsonx Orchestrate "A comprehensive agent management platform to build and manage all your AI agents in one control plane." It builds agents and agentic workflows, and can "automatically scan for agents" in other environments, "such as Amazon Bedrock", and bring them under the same controls.

Needs you to run. Nothing named. It runs as a managed service on IBM Cloud or AWS, or on your own premises.

Control. IBM's developer docs point to agentic workflows when "Human-in-the-loop interaction is needed for approvals or decision-making". It connects to other agents over A2A and MCP.

Data. It can run on your own premises, and data isolation, a dedicated data space and a HIPAA-ready option come with Premium. IBM's service terms say that, unless you choose to submit feedback, IBM "does not use your content or model outputs to train any foundation models or embedding models".

Price. Essentials from $567 a month, for 4,000 monthly active users and up to 200,000 messages; Standard from $6,805 a month, for 40,000 users and up to 2,000,000 messages; Premium, which adds data isolation, by quote. IBM calls these prices indicative. There's a 30-day free trial.

Cloud platforms for building your own

These are for your own developers, building agents on the cloud you already use. AgentCore, Foundry and Agent Platform each host your agents, connect them to your tools and let you require a person's approval before a tool runs. You bring the agent's logic, and the model is billed separately.

Amazon Bedrock AgentCore: best for teams that build on AWS, with any framework and model

What it does. AgentCore runs production agents built with any framework, including CrewAI, LangGraph, Google's ADK and OpenAI's Agents SDK, on any model. Its Gateway turns your APIs and Lambda functions into tools an agent can call, and connects to MCP servers and to Salesforce, Slack, Jira, Asana and Zendesk.

Needs you to run. An AWS account and developers: Node.js 20 or later for its command-line tool, and Python 3.10 or later for agent code.

Control. AgentCore Policy checks every tool call at the gateway against your rules, and can require "that a recorded human approval exist before a consequential action", according to AWS's FAQ. AWS Agent Registry, in preview, catalogues agents and tools built on AgentCore, on other providers or on your own premises, with approval workflows for what gets published.

Data. It runs in fifteen AWS Regions, and every AgentCore service can connect privately to your own network.

Price. Consumption, with no minimum: the Runtime costs $0.0895 per vCPU-hour and $0.00945 per GB-hour of memory on its first version, or $0.1276 and $0.0169 on the newer v2, and the Gateway $0.005 per 1,000 calls. Models aren't on AgentCore's price list; they're billed wherever they come from. New AWS customers get up to $200 in Free Tier credits.

Microsoft Foundry Agent Service: best for Azure teams building agents in code

What it does. Foundry Agent Service builds and runs agents on Azure, from prompt-based agents to hosted agents built with Microsoft Agent Framework, LangGraph or other open-source frameworks, which Foundry runs in containers with its own scaling, monitoring and identity. Published agents can go to Microsoft Copilot, Teams and the Entra Agent Registry.

Needs you to run. An Azure subscription and developers.

Control. Microsoft's guidance is to "Require approval for high-risk operations, especially tools that write data or change resources", and to log approvals and tool calls.

Data. Microsoft says prompt agents can run on a private network, and hosted agents can use your own virtual network, with each session in an isolated sandbox connected to it.

Price. No charge for creating or running Foundry-native agents. You pay for model tokens and for tools, such as $0.033 per Code Interpreter session and $14 per 1,000 web searches, and for hosted agents by the container hour. New Azure accounts get a $200 credit for 30 days.

Gemini Enterprise Agent Platform: best for Google Cloud teams that want an identity and a registry for every agent

What it does. Google's Agent Platform, which replaced Vertex AI on 22 April 2026, builds agents in a low-code studio or in code with the open-source Agent Development Kit, and runs long-running agents "for days at a time". Agent Registry "indexes every internal agent, tool, and skill", and Agent Gateway connects agents to tools under one set of policies.

Needs you to run. A Google Cloud project with billing switched on, and developers.

Control. ADK's Tool Confirmation can pause a tool until a person says yes, and the platform can enforce natural-language policies on tool calls. Every agent, "whether built on Agent Platform or sourced from our partner ecosystem", gets a trackable identity.

Data. Google says it won't use your data to train or fine-tune any model without your permission. Gemini models cache inputs and outputs in memory by default, which you can switch off.

Price. Pay as you go: the Agent Runtime costs $0.085 per vCPU-hour and $0.009 per GiB-hour after a monthly free tier, idle time between turns isn't billed, and models are billed at their own prices. ADK is free and open source, and new Google Cloud customers get $300 in free credits.

For agents that work on your warehouse data, see our BigQuery AI agent page.

Open-source agent frameworks

If your developers would rather own the agent's code, these are the frameworks the ranking pages name. They're free; you pay for the models and for wherever you run them. In a Reddit thread that ranks second for "agentic ai platforms", developers compare them. The replies name Microsoft Agent Framework, GitHub's Copilot SDK, LangGraph and OpenAI's Agents SDK, and one ties the choice to its cloud: a Microsoft stack of Azure, Foundry and Microsoft Agent Framework, adding "We also used langgraph but it's slowly disappearing".

When no platform reaches the job

A custom agent: best for a job that crosses platforms, or has to act in SAP or Workday on their terms

What it does. Meronym, the publisher of this guide, builds it. We scope one workflow in two weeks, write down what the agent may do alone and what it has to escalate, and give you a fixed quote and your engineer's name before any build. It's built on the platform and cloud you already run, reaches SAP and Workday through the routes those vendors set, and write actions wait behind an approval gate until the agent passes an evaluation on your own reviewed cases. This guide does not cite a published client result for our entry.

Needs you to run. Nothing new: it's built on the systems and cloud you have.

Control. What it may do alone and what it escalates are written down before the build, and it registers wherever you govern your other agents.

Data. Records are processed in your cloud account or a private deployment you approve, and nothing is copied to a Meronym environment to train or tune a model.

Price. After the two-week scope, $4k–$12k and credited against the build, a pilot on one workflow is $8k–$35k over four to six weeks, and a production agent $16k–$82k over eight to fourteen. The running cost is quoted apart.

The agentic AI platforms side by side

One row per platform, to take into a vendor meeting. The last column is the limit to ask each vendor about first.

The vendors that want to govern everyone's agents

Nine of the thirteen vendors in this guide now offer a place to see and govern an organisation's agents, including agents they didn't build; Amazon's is still in preview. None of the pages ranking for this search sets them side by side.

Each of these is also a bid to be the platform all your agents report to. Pick one, and make it the one that already sees the systems where agents change money or people's records. Running two means keeping two lists of which agents exist and what each may do, and reconciling them.

Two vendors also decide how anyone else's agent gets in. SAP's API Policy, as we read it on 27 September 2026, bars API use by "(semi-) autonomous or generative AI systems that plan, select, or execute sequences of API calls" except through "SAP-endorsed architectures", and SAP's FAQ on the policy says it covers the on-premise portfolio as well as the cloud. Workday's documentation says: "All third-party agents interacting with Workday APIs must route their API traffic through Agent Gateway using Agent System User (ASU) security." So an agent you build or buy elsewhere reaches SAP or Workday only through the door each vendor provides, under the identity it issues. Ask any vendor that promises to act in those systems which door it uses.

What enterprise AI agents cost

Oracle's own agents cost nothing on top of Fusion. The other fourteen platforms charge in eight different ways:

• Per action: Salesforce, SAP, Copilot Studio, and Workday's Self-Service Agent.
• Per model call: UiPath.
• Per agent run: Glean.
• Per seat: Agent 365 and Gemini Enterprise's seat editions.
• Per seat plus usage: Microsoft 365 Copilot, whose Cowork is billed by use, ChatGPT Enterprise and Claude Enterprise.
• Per monthly plan, capped by users and messages: IBM watsonx Orchestrate.
• Per hour of compute: AgentCore, Foundry's hosted agents and Agent Platform, with the model billed separately.
• Inside a package quoted per customer: ServiceNow.

Five of them publish the unit but not its price: Workday's Flex Credits, SAP's AI Units, UiPath's Platform Units, Glean's FlexCredits and ChatGPT's credits.

What happens when the units run out differs too:

• Microsoft switches Copilot Studio's custom agents off when a tenant reaches 125% of its prepaid credits, unless a pay-as-you-go meter picks up the overage.
• UiPath won't run agents in production without purchased units.
• OpenAI pauses ChatGPT Enterprise's advanced features when the shared credit pool is used up, unless a workspace owner allows overages or buys more.
• Anthropic stops work for the whole organisation on self-serve Claude Enterprise when prepaid credits reach zero, until an owner buys more.
• SAP lets you switch agents off in SAP for Me as you near your purchased volume; keep them running and it bills the excess at its excess-use rate.
• Salesforce bills use beyond your entitlement at your contracted rate, monthly in arrears, with no overage penalty, and unused Flex Credits don't carry into the next term.

Ask for the rule in writing, because it decides whether an overrun stops the work or only costs more.

To see how far apart the units sit, take one month: 10,000 employee requests, each needing four actions. Four is a round number for the sum, not any vendor's figure, and an "action" isn't the same size everywhere: a Salesforce action, an SAP agent step and a UiPath model call measure different things.

Three rows rest on a stated basis. Workday's assumes every request goes to the Self-Service Agent at the rate card's 1 credit per action; its other agents are metered per request, CV, requisition or document. Glean's is a floor: on Glean's own figures a tenth of runs use about 114 credits or more, so those 1,000 runs alone come to about 114,000, whatever the other 9,000 use. IBM's turns on headcount: Essentials covers 4,000 monthly active users, so if more people than that send the 10,000 requests, the plan is Standard.

Six platforms aren't in the month. Gemini Enterprise, ChatGPT Enterprise and Claude Enterprise charge by the seat, so their month follows headcount rather than requests, with ChatGPT's and Claude's usage on top. AgentCore, Foundry and Agent Platform charge for compute time and bill the model separately, so theirs depends on how long each request runs and which model it calls.

Only four of the nine rows can be put in dollars from public pages, and for one of them, Oracle, the answer is nothing on top of Fusion. Microsoft's rate table also charges no credits when the person using the agent already has a Microsoft 365 Copilot seat, which is $30 a month each, so the cheapest route depends on how many of your people hold seats. Before you compare quotes, ask each vendor how many of its units a typical request of yours uses, and what one unit costs.

Buy, build, or both

Buy first. For most jobs the right first agent is the one inside the system where the job already happens, because it inherits that system's permissions, approvals and audit trail. Add a control plane once agents act in more than one vendor's systems. Building starts to pay under three conditions:

1. The job crosses systems no single platform reaches. List every system the agent must read or write. If none of the platforms you run connects them all, that part of the job remains with your team or needs a custom integration, often with an integration layer such as MuleSoft integration services underneath.
2. A vendor's rules decide the route. An agent that acts in SAP or Workday from outside has to use the routes set out above. Our pages on SAP AI agents and Workday AI agents cover what that takes.
3. Your volume makes the unit price cost more than owning the agent. Multiply your monthly actions by the unit price and by twelve. At $0.10 an action, 40,000 actions a month on Agentforce come to $48,000 a year. Compare that with a build, $16k–$82k for a production agent of ours, with the running cost quoted apart.

A job whose steps never change needs no agent at all: that's fixed workflow automation, the work of business process automation services. An incorrect action in a regulated process needs its own rules, logs and sign-off whichever platform you choose; our page on AI agents for compliance covers what that takes.

If your job meets one of the three conditions, that's the work behind our AI agent development services: the custom agent in the list, scoped in two weeks to a fixed quote. For a programme that spans several platforms and business units, see our enterprise AI development services. If you're still deciding which platform should govern the rest, that's agentic AI consulting.

Your next step

Shortlist two or three platforms for your enterprise AI agents from the systems you already run, and put the nine questions above to each vendor in writing. Then run a month of your own past requests through the one that answers best, with a person approving every write, and count the units it used.

If no platform reaches the job, or it has to act in SAP or Workday from outside, send it to us instead: use the Scope an agent button at the end of this page.