AI Agents in Business: Deployment Risks and How to Stay in Control

Русская версия

AI agents for business are programs built on language models that do not only answer but act: they write and run code, work inside CRM systems and make payments. The main risk of deployment is not an inaccurate answer but an action taken with the company’s rights: the company fixes and pays for the agent’s mistake itself. Control is kept when the rules for actions are set in advance and checked before execution.

In brief

Agents are already at work: in software development, in sales and customer service, and in the first payment pilots.
Failures are not rare: in the VentureBeat tracker (July 2026, 108 organizations, a self-selected sample), 49% said that in the past year an agent or LLM feature that had passed internal evaluations caused a failure that customers saw.
Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.
The company answers, not the agent: a program is not a party to civil-law relations, so the losses from its mistake fall on whoever used it.
Before launch, four things must be known: what the agent is allowed to do, who approved the rules, how to stop it and how to prove later what it did.

Which tasks do companies already hand to agents?

Most often software development, work with customers in CRM systems and routine operations in corporate systems; the first agents have started making payments. How an AI agent differs from a chatbot and an assistant is covered on the page What is an AI agent.

Task	What is known	Source
Software development	Claude Code’s run-rate revenue has grown to over $2.5 billion and has more than doubled since the beginning of 2026; enterprise use represents over half of all Claude Code revenue	Anthropic, 12 February 2026
Work with customers in CRM	Agentforce annual recurring revenue of $800 million; over 29,000 Agentforce deals closed since launch	Salesforce, results for the fiscal year ended 31 January 2026
Payments	Europe’s first live end-to-end payment executed by an AI agent, in a controlled environment, within predefined limits and permissions; the pilot does not constitute a commercial rollout	Santander and Mastercard, 2 March 2026

How agent payments work and who is liable for them: the page on the agent economy.

The market is growing fast. According to PitchBook, in 2025 VC-backed agentic AI companies raised $24.2 billion across 1,311 deals, nearly 73% of their cumulative VC deal value between 2015 and 2024. Gartner predicts (26 August 2025) that 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% in 2025.

What risks does deploying AI agents bring to a business?

The main risk is an action, not a text: the agent works with the company’s keys and rights, and its mistake immediately becomes the company’s operation. Five risks come up more often than others.

Risk	How it shows	What reduces it
A wrong action	Deleted data, a wrong payment, a letter to the wrong client	Rules for actions checked before execution; a way to roll back
Someone else’s instruction (prompt injection)	The agent follows an instruction hidden in an email, a document or a web page	The rule judges the action itself, not the text that triggered it
An agent nobody tracks	The agent was launched without IT and security approval and runs under a shared account	A register of agents; each has its own account, rights and owner
A long autonomous chain	The agent drifts from the task, ignores instructions, stops	Small tasks inside a mapped process; a human at the key steps
Costs without return	The project gets more expensive, and its value is not measured	A value metric before launch; a pilot on one process

What surveys say. A survey by Gravitee (a vendor of an agent management platform; over 900 executives and technical practitioners, February 2026): “88% of organizations reported confirmed or suspected AI agent security incidents in the last year.” On average, only 47.1% of an organization’s AI agents are actively monitored or secured. Only 14.4% of organizations have full security approval for their entire agent fleet. Only 21.9% of teams treat AI agents as independent, identity-bearing entities. This is a vendor survey, not market statistics.

What an experiment showed. KPMG in the Netherlands and the University of Amsterdam are testing whether a company can run without people at all (press release of 20 November 2025). The first plan had to be revised. In the words of Professor Sander Klous: “If you let AI agents perform human roles one-to-one, such as a CFO, they tend to drift off and hallucinate. Sometimes they completely ignore their instructions or suddenly stop working.” The researchers mapped work processes in detail and assigned agents to small tasks within them, an “army of disposable agents” instead of five agents in executive roles.

A real case. On 24 April 2026 a Cursor agent running Claude Opus 4.6 deleted the PocketOS production database together with its backups. This and other cases are analyzed on the page on AI agent incidents.

Why does Gartner expect over 40% of projects to be canceled by the end of 2027?

Because of escalating costs, unclear business value or inadequate risk controls: this is how Gartner explains the prediction in its press release of 25 June 2025. It is a forecast, not a measurement.

The same press release names two more reasons:

Agent washing. Many vendors rebrand existing products, such as AI assistants, robotic process automation (RPA) and chatbots, “without substantial agentic capabilities”. Gartner estimates “only about 130 of the thousands of agentic AI vendors are real.”
Immature models. In the words of a Gartner analyst, “Most agentic AI propositions lack significant value or return on investment (ROI), as current models don’t have the maturity and agency to autonomously achieve complex business goals or follow nuanced instructions over time.” Many use cases positioned as agentic do not need an agent at all.

Companies invest cautiously. In a January 2025 Gartner poll of 3,412 webinar attendees, 19% said their organization had made significant investments in agentic AI, 42% conservative investments and 8% none; the remaining 31% were waiting or unsure.

Gartner’s advice at this stage: pursue agentic AI only where it delivers clear value or ROI. Integrating agents into legacy systems is complex and costly, so it is often better to rethink the process itself around the agent.

What should an owner check before launching an agent?

Six questions. If any of them has no answer, the agent is not ready for launch.

What is the agent allowed to do, and what is forbidden? A list of actions, not “case by case”. Separately, the actions that cannot be undone: a payment, a deletion, a message to a client.
What rights does it work with? Its own account and minimal rights. Not a shared administrator key and not the company’s main account.
Who approved the rules, and who is responsible for the agent inside the company? A human approves the rules; the agent has an owner.
Where is the action checked: before execution or after? A text filter and an after-the-fact log do not stop a harmful action. How a check before execution works: the page on IS 3.0.
How is the agent stopped and its action undone? Who does it, and within how many minutes.
How will it be proven later what the agent did and under which rule? A log that cannot be rewritten after the fact. Without it, both an insured event and the director’s good faith are hard to prove.

What to do

Start with one process where the value can be measured: time, money, the number of errors.
Split the process into small tasks, and do not give the agent the role of “a whole employee”.
Keep a register of agents: each has an account, rights, an owner and rules.
Set rules for actions and check each action before execution; human approval for actions that cannot be undone. On the limits of manual approval: the page on human-in-the-loop.
Check contracts and the insurance policy: who answers for the agent’s mistake and whether insurance covers it. In detail: the pages on liability for AI mistakes and cyber insurance.
For a strategy for years ahead: the institute’s book “Proven Action” (in Russian), on keeping power over agents when they act in the company’s name.

Frequently asked questions

What are AI agents for business?

They are programs built on language models that carry out a company’s tasks themselves: they write code, answer customers, prepare documents and perform operations in corporate systems. Unlike a chatbot, an agent does not only answer but acts with the rights the company gave it.

What are the risks of AI agents?

A wrong action taken with the company’s rights, following someone else’s instruction from an email or a document (prompt injection), agents outside IT and security tracking, failures on long autonomous tasks, and costs without measured value. The company answers for the consequences.

How should a company start deploying AI agents?

With one process whose value can be measured. Before launch, the company defines what the agent is allowed to do, gives it a separate account with minimal rights, appoints an owner and decides how each action is checked before execution and how it can be undone.

Can AI agents automate a business without risk?

Not without risk: an agent makes mistakes, and it can be misled. The risk becomes manageable when the rules for actions are set in advance, each action is checked before execution, and a log lets the company establish later what happened.

Who is liable if an AI agent makes a mistake?

The company that used it. Under the Russian Civil Code a program is not a party to civil-law relations, so the company answers to clients and counterparties, and inside the company the question of the director’s liability may arise.

Why are AI agent projects canceled?

According to a Gartner prediction (press release of 25 June 2025), over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls. Gartner advises pursuing agentic AI only where its value is clear.

Sources

Anthropic, “Anthropic raises $30 billion in Series G funding at $380 billion post-money valuation”, 12 February 2026: https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation
Salesforce, fourth quarter and fiscal 2026 results (fiscal year ended 31 January 2026), 25 February 2026, exhibit to Form 8-K: https://www.sec.gov/Archives/edgar/data/1108524/000110852426000056/crm-q4fy26xexhibit991.htm
Santander, “Santander and Mastercard complete Europe’s first live end-to-end payment executed by an AI agent”, 2 March 2026: https://www.santander.com/en/press-room/press-releases/2026/03/santander-and-mastercard-complete-europes-first-live-end-to-end-payment-executed-by-an-ai-agent
PitchBook, “Agentic AI: The Evolution to Autonomous Systems, Part I”, Q2 2026: https://pitchbook.com/news/reports/q2-2026-pitchbook-analyst-note-agentic-ai-the-evolution-to-autonomous-systems-part-i
Gartner, press release of 26 August 2025: https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025
Gartner, press release of 25 June 2025: https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
VentureBeat, “Agentic reliability and evaluations”, 12 August 2026 (July wave of the tracker): https://venturebeat.com/resources/agentic-reliability-and-evaluations-enterprises-that-got-burned-by-a-bad-eval-are-the-most-likely-to-remove-humans-from-the-loop-not-the-least
Gravitee, “State of AI Agent Security 2026 Report: When Adoption Outpaces Control”, 4 February 2026 (vendor survey): https://gravitee.io/blog/state-of-ai-agent-security-2026-report-when-adoption-outpaces-control
KPMG Netherlands, press release on the zero-person company experiment, 20 November 2025: https://kpmg.com/nl/en/about-us/press-and-media/press-releases/2025/11/ai-zero-person-company-experiment.html
Lumenova, “Agentic AI Security Incidents”, 25 August 2026: https://www.lumenova.ai/blog/agentic-ai-security-incidents/
Civil Code of the Russian Federation (Part One), as amended on 10 June 2026: Article 2 (in Russian), ConsultantPlus: https://www.consultant.ru/document/cons_doc_LAW_5142/

Checked: 17 October 2026. Survey figures apply to their samples; Gartner predictions are forecasts, not measurements.

AI agents for business: risks and control