Skip to content
TECH CEO Daily
AIGuide

Buying AI agents in 2026: a CEO's guide to costs, pricing models, risks and vendor checks

AI agents act inside your systems, so buying them is closer to hiring a contractor than licensing an app. Here is how to judge cost, pricing and risk.

By · Editor

· 7 min read · Fact-checked

The 60-second brief

  • 1Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 over cost, value or risk controls.
  • 2Vendors now price agents per user, per action, per conversation or per outcome, often in the same contract.
  • 3OWASP ranks prompt injection as the top risk for AI applications; it becomes dangerous when agents can take actions.

The news

In short: an AI agent is software that uses a large language model to plan and carry out tasks, such as answering a customer, updating a record or filing a ticket, with limited human input. When a company buys one, it is not just licensing a tool; it is giving software permission to act on its systems and data. That changes three parts of the buying decision: how cost is measured, how risk is controlled and what to demand from vendors. This guide covers each, using published vendor prices and guidance from Gartner, NIST and OWASP.

The AI agent market in 2026

Start with a realistic view of the market. Gartner said on June 25, 2025 that it expects over 40% of agentic AI projects to be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. Gartner also warned of "agent washing", vendors rebranding existing chatbots, assistants and robotic process automation (RPA) as agents, and estimated that only about 130 of the thousands of agentic AI vendors are real. In the same release it predicted that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024, and that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028.

Both points can be true. Agents are being built into the software companies already use, and many early projects will not survive contact with real costs and controls. Gartner's advice was to pursue agentic AI only where it delivers clear value or return on investment, and noted that many use cases sold as agentic do not need an agent at all. A rule-based workflow or a simple assistant may do the job more cheaply and with less risk.

What AI agents cost: the pricing models

Vendors have not settled on one way to charge, and many large ones offer several at once. Salesforce (CRM) is a useful reference because it publishes its menu. Its Agentforce pricing page lists a per-user add-on at $125 per user per month, Flex Credits at $500 per 100,000 credits with a standard agent action costing 20 credits, and $2 per conversation for customer-facing agents. By our calculation, the credit price works out to $0.10 per standard action. Microsoft (MSFT) lists Microsoft 365 Copilot at $30 per user per month paid yearly. Intercom, a private customer-service software company, prices its Fin AI agent at $0.99 per outcome, which it counts when a customer confirms the issue is resolved, does not ask for more help after Fin responds, or when Fin completes a defined workflow.

How each agent pricing model shifts risk

Each model shifts risk. Per-user pricing is predictable but charges for people, not work. Per-action and per-conversation pricing tie cost to activity, so a busy month or an agent that takes many steps per task can raise the bill quickly. Per-outcome pricing ties cost to results, but the vendor writes the definition of a result. The total cost also includes things that rarely appear on the price page: integration work, data preparation, monitoring, human review of agent output, and staff time spent handling cases the agent escalates. Gartner pointed to integration with legacy systems as a frequent source of cost and disruption.

AI agent risks: what can go wrong

The Open Worldwide Application Security Project (OWASP) ranks prompt injection first in its 2025 Top 10 for large language model and generative AI applications. Prompt injection happens when input changes a model's behavior in unintended ways. It can be direct, from a user's prompt, or indirect, from instructions hidden in a web page, email or file the agent reads. OWASP lists excessive agency separately, the risk created when a model is given too many functions or permissions. The combination is what matters for buyers: an agent that reads untrusted content and can also send emails, change records or move money can be steered into taking actions no one approved.

Using the NIST framework for AI agents

The US National Institute of Standards and Technology (NIST) offers a voluntary structure for managing these risks. Its AI Risk Management Framework 1.0, released on January 26, 2023, organizes the work under four functions: Govern, Map, Measure and Manage. NIST added a Generative AI Profile, NIST-AI-600-1, on July 26, 2024. Neither is a checklist for a specific product, but both give boards and risk teams a common language for asking vendors and internal teams how a system is governed and tested.

The numbers

Agentic AI projects Gartner expects to be canceled by end of 2027
over 40%
Agentic AI vendors Gartner considers real
about 130 of thousands
Salesforce Agentforce per-user add-on
$125 per user/month
Salesforce Flex Credits
$500 per 100,000; 20 per standard action
Salesforce customer-facing agent
$2 per conversation
Microsoft 365 Copilot
$30 per user/month, paid yearly
Intercom Fin AI agent
$0.99 per outcome

AI agent vendor checklist

Before signing, ask every agent vendor these questions. What exactly can the agent do on its own, and which actions require human approval? What identity does it use in our systems, and can its permissions be limited to one task at a time? How does it defend against prompt injection from emails, documents and web pages it reads? What logs does it keep of every action, and can we export them to our own security tools? How are errors and reversals handled, and who pays for an agent's mistake? Where is our data processed and stored, is it used to train the vendor's models, and can we opt out? What is the billable unit, how is it counted and can we audit it?

For CFOs: build the business case on a specific workflow, not on the technology. Measure the current cost per task, such as cost per resolved support ticket, then compare it with the agent's per-task price plus integration, oversight and escalation costs. Run a time-limited pilot with a spending cap and a pre-agreed success measure. Negotiate price protection for renewal, rollover of unused credits and the right to exit if agreed accuracy or resolution targets are not met. Gartner's cancellation forecast is a reminder to keep early contracts short and measurable.

For CISOs and general counsel: treat each agent like a privileged service account and a third-party contractor. Give it the minimum access needed, keep a human in the loop for anything that moves money, changes customer records in bulk or sends external communications, and include agents in incident response plans. Contracts should cover data use, liability for agent errors, security testing and notification if the vendor's model or tools change.

How AI agents change software buying

Traditional software was bought on features and seat counts, then deployed by IT. Agents blur that line. Business teams can switch on agent features inside tools they already pay for, sometimes with a credit wallet that bills as soon as the feature is used. That makes central visibility harder. Companies that handle this well tend to keep an inventory of every agent in use, who owns it, what it can access and what it costs, and require a short risk review before any agent gets write access to a core system.

It also changes negotiating power. When an agent is embedded in a platform such as a customer relationship management system, switching away from it can mean rebuilding workflows, retraining staff and re-testing controls. Buyers should ask early how agent configurations, prompts, logs and data can be exported, and whether the vendor's agents work with other vendors' systems through open interfaces. Portability is easier to win before signing than at renewal.

Finally, expect the price menus to keep changing. Salesforce alone publishes per-user, per-credit and per-conversation prices for the same Agentforce product line, and Microsoft and Intercom use different units again. Contracts that fix the billable unit, cap price increases and give a right to switch units at renewal will age better than ones that lock in today's list prices.

Frequently asked questions

What is the difference between an AI assistant and an AI agent?

An assistant responds to a person's request, such as drafting text or answering a question, and the person decides what to do next. An agent can plan steps and take actions in other systems, such as updating a record or sending a message, with limited human input. Gartner warns that some vendors label assistants and chatbots as agents.

How should we pilot an AI agent?

Pick one high-volume, low-risk workflow with a clear cost per task. Limit the agent's permissions, require human approval for consequential actions, set a spending cap, and measure accuracy, resolution and total cost against the current process for a fixed period before expanding.

Which pricing model is best for AI agents?

It depends on volume and how clearly success can be defined. Per-user pricing suits broad, steady use by employees. Per-action or per-conversation pricing suits variable workloads but needs spending controls. Per-outcome pricing suits tasks with clear, auditable results, provided the contract defines an outcome precisely.

Are there standards for AI agent security?

There is no single certification. NIST's AI Risk Management Framework and its Generative AI Profile provide a voluntary risk structure, and the OWASP Top 10 for LLM and generative AI applications lists common technical risks, including prompt injection and excessive agency. Ask vendors to map their controls to both.

Who is liable when an agent makes a mistake?

That is set by your contract and applicable law, and many standard vendor terms limit the vendor's liability. Negotiate responsibility for agent errors, data misuse and security failures explicitly, and keep humans accountable for high-impact decisions.

What “Fact-checked” means

Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.

What we checked
Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
How
A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
Who
The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, . A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
If something is wrong
“Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error

How we fact-check →

Companies in this story

AI agentsGartnerSalesforceMicrosoftEnterprise software

Earlier coverage of Salesforce

All Salesforce coverage →

Written by

Editor · Technology & Business Writer

Hussein is a writer and business technology enthusiast focused on the intersection of technology, entrepreneurship, finance, artificial intelligence, and digital innovation.

CoversAICybersecurityBig TechSaaSStartupsFintech

About this story. Researched from primary sources whenever they are available and fact-checked before publication.

Published by Tech CEO Daily, an independent publication. Masthead · Editorial standards

Follow Tech CEO Daily on Facebook for the day’s top stories in your feed.

Free newsletters

The technology briefing for people running businesses.

Daily, weekly, bi-weekly or monthly. You choose.

How often

The Daily Brief · Monday to Saturday, 7 a.m. ET

Free forever. One click to unsubscribe. We never sell your email.