Agentic AI has moved from conference hype to a budget line item. This is where the conversation gets more interesting and more uncomfortable. Unlike traditional AI systems that respond to a single prompt, classify a document, recommend an action, or generate a summary, agentic AI systems are designed to pursue goals. They plan, call tools, inspect results, retry failed steps, consult memory, hand off tasks to other agents, and sometimes critique their own work before producing an answer or taking an action.

That extra autonomy is the value proposition. It also introduces the cost problem.

A single chatbot interaction may consume a few thousand tokens. A useful agentic workflow can consume hundreds of thousands or millions of tokens per day because it does more than answer a question. It decomposes the problem, retrieves context, reasons through options, invokes APIs, checks the output, and often runs multiple passes before reaching a result. Therefore, the economics need to be understood at the level of “agent instances,” not just model calls.

For the estimates below, I am using a blended token cost of $3 dollars per million tokens. This is not intended to reflect a single vendor’s list price. It is a blended planning figure that assumes a mix of input and output tokens, reasoning steps, retrieval-augmented generation, summarization, tool calls, memory updates, and occasional use of larger context windows. Some enterprises will pay less through volume discounts or by routing work to smaller models. Others will pay more by using premium models, long-context prompts, web browsing, large document ingestion, and repeated reasoning loops.

The basic formula is straightforward. If an agent consumes 2 million tokens per day, it consumes 730 million tokens per year. At $3 per million tokens, that single agent costs about $2,190 per year in token burn. That number sounds surprisingly low until you multiply it by the number of agents, workflows, and users, plus the surrounding infrastructure required to run these systems safely.

What an agent really costs

In the model used here, the annual token-only cost per agent ranges from about $1,095 to $3,833, depending on the use case.

• A lightweight HR recruiting or onboarding agent that consumes one million tokens per day costs about $1,095 per year.
• A more demanding software engineering agent that consumes 3.5 million tokens per day costs about $3,833 per year.
• Customer support agents are roughly $2,190 per year.
• Legal contract agents cost about $2,409 per year.
• Security triage agents run about $2,738 per year.
• Research agents are around $3,066 per year.

These figures are useful but incomplete. They include only LLM token consumption and exclude orchestration platforms, vector databases, observability, model evaluation, security controls, workflow monitoring, human review, enterprise application integration, data pipelines, audit logging, prompt management, and the engineers needed to build and maintain the systems. In real deployments, I would expect the all-in operating cost to be two to five times the raw token cost. For regulated or mission-critical environments, the multiplier can be even higher.

This is where many agentic AI business cases become less clear. The model call may be inexpensive, but the system around the model is not. An agent that can update a customer relationship management system, approve a refund, generate a purchase order, or recommend a security containment action needs guardrails, permissions, logging, rollback mechanisms, and human escalation paths. These are not optional features. They are the difference between a demo and an enterprise system.

Use cases that make economic sense

Customer support is one of the more obvious use cases. A typical support automation deployment may use eight agents: an intake classifier, a knowledge retrieval agent, a response drafting agent, an escalation agent, a quality review agent, a CRM update agent, a sentiment detection agent, and an analytics agent. At two million tokens per agent per day, each agent costs about $2,190 per year in token burn, bringing the annual total to roughly $17,520. If that system deflects even a modest number of tickets or improves agent productivity, the economics can be attractive.

Sales development is another practical example. A five-agent system for account research, lead enrichment, email personalization, CRM updates, and follow-up scheduling may consume 1.2 million tokens per agent per day. That results in an annual cost of about $1,314 per agent, or $6,570 for the full agent team. This can be compelling if it improves pipeline quality, but it can also be wasteful if agents generate low-quality outreach at scale. The cost of brand damage is not measured in tokens.

Software engineering is more expensive but potentially more valuable. A 12-agent system covering requirements analysis, architecture, code generation, testing, review, security checks, documentation, CI debugging, refactoring, release notes, dependency analysis, and hot-fix support may consume 3.5 million tokens per agent per day. That works out to about $3,833 per agent annually or roughly $45,990 for the full system. Compared with engineering salaries, the token cost is small. The real question is whether the system reliably improves throughput without increasing defects, security vulnerabilities, or maintenance complexity.

Security operations also fit the agentic model because the work is repetitive, time-sensitive, and context-intensive. A 10-agent security triage system could include agents for alert triage, log analysis, threat intelligence, endpoint investigation, network investigation, incident summarization, ticketing, compliance evidence, escalation, and post-mortem. At 2.5 million tokens per agent per day, the annual token cost is about $2,738 per agent or $27,375 for the system. This is easy to justify if it reduces alert fatigue and accelerates response, but risky if the agents hallucinate causality or bury critical signals in confident summaries.

Finance, legal, healthcare administration, market research, HR, and supply chain are also viable:

• A six-agent finance close system may cost about $9,855 per year in token burn.
• A four-agent legal contract review system may cost about $9,636.
• A seven-agent healthcare administrative workflow may cost about $13,797.
• Competitive intelligence using six agents may cost $18,396.
• HR recruiting and onboarding using five agents may cost $5,475.
• Supply chain planning and exception management using eight agents may cost $21,024.

Across all 10 example use cases I’ve mentioned in this section, the model assumes 71 agents and a total annual token burn of about $175,638.

The comparison with traditional AI

The economics of agentic AI should always be compared with simpler approaches. Traditional AI, workflow automation, rule engines, robotic process automation, and non-agentic LLM calls are often cheaper, easier to govern, and more predictable. Agentic AI is usually overkill for tasks like classification, extraction, summarization, routing, or drafting within a narrow context. A deterministic workflow with a single model call can do the job at a fraction of the cost and risk.

Agentic systems make sense when the process requires judgment across multiple steps, dynamic planning, tool use, exception handling, and adaptation to incomplete information. They are valuable when the path to the answer cannot be fully scripted in advance. They are less valuable when enterprises use them as a trendy substitute for basic automation.

The best architecture is usually hybrid. Use traditional automation where the process is stable. Use non-agentic AI where the task is bounded. Use agentic AI only where autonomy creates measurable leverage. That means fewer agents, tighter scopes, explicit budgets, model routing, token monitoring, and human checkpoints for high-impact decisions.

The financial mistake many organizations will make is to treat agents as digital employees with near-zero marginal cost. They are not. They are probabilistic software components that consume tokens, trigger tools, create operational dependencies, and require supervision. The token bill may be manageable. The governance bill may not be.

Agentic AI can absolutely be worth the money. In many cases, the annual token burn for a useful agent team is less than the loaded cost of a single employee. But that is not the same as saying it is cheap. Companies must measure agent cost per completed business outcome, not cost per prompt or cost per model call. In the end, the question is not how much an agent costs. Ask yourself the right question: Does the autonomy an agent provides outweigh the complexity it introduces?

AI governance is an ongoing game of catch-up for enterprises. Model updates and iterations are rolling out at a rapid clip, often making governance frameworks obsolete before they’re battle-tested.

To evolve beyond this paradigm, OpenAI is introducing Active sessions. This new ChatGPT security feature allows users to review and log out of one or more sessions through a simple interface. The feature is now available across all ChatGPT accounts and workspace types, including personal and managed workspaces. Experts call it an important development for the model provider, which currently has 1 billion monthly active users.

Previously, organizations often had limited visibility into where users were logged in, and simply relied on password resets or broad account actions to force re-authentication, noted Ensar Seker, CISO at SOCRadar. “Granular session control is a more efficient and less disruptive approach. From a governance perspective, session transparency improves accountability and supports investigations,” he explained.

A holistic view across session activity

Active sessions allows admins to see known browser and app sessions across ChatGPT, Codex, and API Platform. Specifically, they are able to view device and browser information, approximate location, sign-in date and time, whether a device is trusted, and whether the session is current.

To access the feature, users can go to ‘Settings’ > ‘Security’ > ‘Active sessions.’ They can then log out of specific sessions and remove devices from trusted services. They also have the ability to log out of all sessions (thus ending sessions across devices), however, this action can take up to 30 minutes to complete.

However, OpenAI emphasizes that session details may be “approximate or incomplete,” and that the feature has limits. It does not show or manage connected apps or third-party app sessions, sign-ins through third-party services, Codex CLI sessions, or recently signed-out sessions.

Further, Active sessions cannot be used with accounts linked to an enterprise’s single sign-on (SSO), including security assertion markup language (SAML) and OpenID Connect (OIDC).

Better late than never

While Active sessions is an important security and governance development, experts note that the feature is basic, and was a long time coming.

“The reality of OpenAI offering the ability to end active sessions on ChatGPT by administrators is that it’s something that exists in lots of platforms,” said David Shipley of Beauceron Security. “They should’ve had it sooner, but better late than never.”

From a security standpoint, he noted, OpenAI could do a better job policing ChatGPT to prevent it being used by threat actors to host malware, which is the latest threat to enterprises.

SOCRadar’s Seker also pointed out that this type of visibility and oversight is something that enterprises have expected from SaaS platforms for years. “It allows administrators and users to quickly identify unauthorized access, terminate stale sessions, and reduce the risk of account compromise persisting undetected.”

Iterative upgrades disrupt governance

Last week, OpenAI updated GPT-5.5 Instant in both the ChatGPT app and API to “improve response style and quality,” the company said. It had rolled out GPT-5.5 Instant earlier in May as a successor to GPT‑5.3 Instant, calling it “generally smarter” and prone to fewer hallucinations.

According to OpenAI, the update makes GPT-5.5 Instant “easier to read, more natural in everyday conversations, and better paced in practical help tasks, with fewer overly long or bullet-heavy responses.”

But even with tools like Active sessions, enterprises continue to struggle with governance amidst seemingly continuous iterative model updates. It’s simply not sustainable, said Beauceron’s Shipley: “How do you build an appropriate testing plan with a nondeterministic system?”

SOCRadar’s Seker pointed out that many organizations perform security, compliance, and business validation testing before approving a model for production use. But, “when model behavior changes under the same version family, previously documented assumptions may no longer fully reflect actual performance,” he noted.

“The biggest governance challenge in AI is not model adoption, it’s model change,” Seker said. “Most organizations can evaluate a model once. Far fewer are prepared to continuously evaluate how that model evolves over time.”

This particularly creates challenges for regulated industries where auditability, repeatability, and change management are critical, he said. Even beneficial improvements can introduce governance concerns if organizations are not clearly informed about what changed and when.

Valence Howden, advisory fellow at Info-Tech Research Group, noted that organizations often can’t assess the implications of model iterations against their boundaries, and, worse, are often unaware of them.

While the biggest enterprise challenge was initially tied to which AI model was being used, what that model did, and who owned it, iterative updates can muddy those waters and increase reliance on third party practices and tools that organizations often don’t have the resources for, he noted. 

“Without the ability to opt out [of an update] before it’s incorporated, [enterprises] are basically red-teaming the updates with their clients,” said Howden.

The ongoing game

Security teams today are pushed to their limits because they are expected to manage rapidly evolving models, new features, and changing behaviors, while maintaining compliance, risk management, and business continuity, said SOCRadar’s Seker.

“Governance is difficult because organizations are no longer evaluating a static product,” he said. Rather, they are managing a “continuously evolving service” where capabilities, integrations, and user behaviors can change far faster than can traditional security review cycles.

Info-Tech’s Howden agreed, saying that enterprises’ existing governance practices, especially accountability, are poor, as are their risk practices.

“It’s hard to suddenly become good at things they’re already poor at doing,” he said. “They are also incentivized for speed and innovation, so they ignore governance as a constraint, or don’t want to do it at all.”

How enterprises should respond

Seker advised that, ultimately, organizations should treat AI models as living systems rather than fixed software releases.

Security and governance programs should include continuous validation, monitoring, and periodic re-assessment instead of sole reliance on one-time approval processes, he said. Enterprises should also establish clear vendor change management expectations, including requiring transparency around model updates, behavioral changes, and potential impacts to existing workflows.

“Effective AI governance increasingly depends on visibility into change, not just visibility into risk,” Seker said.

AI could soon lead to systems capable of improving their own performance faster than humans can effectively supervise them, reviving concerns about the industry’s longstanding “alignment problem,” ensuring AI systems reliably pursue human goals, senior Anthropic researchers have warned in a new blog post titled “When AI builds itself.”

Anthropic Institute lead Marina Favaro and Anthropic co-founder Jack Clark outlined three possible futures: growth in AI capabilities may flatten out; AI efficiency gains may continue to grow, but expose bottlenecks elsewhere in software development; or AI systems may become capable of full recursive self-improvement, and build their successors by themselves. It’s that third scenario that’s prompting them to suggest society be ready to hit the brakes on AI development.

“How the alignment problem gets solved — or not — in this future is something we are least certain about,” they wrote. Advanced, self-improving models could follow our needs and wants — or, they warned, “The rare occurrences of misalignment present in today’s models could compound as the models build their successors, growing more frequent but less understood until we lose control of them. It’s possible that we can’t build, integrate, and verify the tools that we’d need to understand which trendline we are actually on.”

While Anthropic’s warning is framed around future AI development, analysts say it highlights governance questions enterprises are already beginning to confront as autonomous AI agents move from answering questions to taking actions.

“The issue is no longer just whether AI gives the right answer, but whether autonomous systems take the right action, at the right time, within the right authority,” said Ashish Banerjee, senior principal analyst at Gartner.

From model governance to agent governance

The warning comes amid growing enterprise investment in agentic AI.

Gartner predicts that by 2028, 15% of day-to-day work decisions will be made autonomously through agentic AI and that one-third of enterprise software applications will incorporate agentic AI capabilities. The firm has also warned that governance shortcomings are already emerging, predicting that 40% of enterprises will demote or decommission autonomous AI agents by 2027 after governance failures become apparent in production environments.

Banerjee said many organizations continue to approach AI agents as advanced productivity tools when they increasingly resemble digital workers operating with delegated authority.

“CIOs should stop treating AI agents as smarter chatbots,” he said. “They are becoming digital workers with delegated authority — and must be governed like privileged users, not productivity tools.”

As agents gain the ability to conduct research, write code, invoke tools, trigger workflows, and make recommendations, enterprises face new risks around unauthorized actions, accountability gaps, data exposure, tool misuse, and insufficient auditability, Banerjee said.

“Human-in-the-loop is not a strategy if the human cannot keep up with the loop,” he said.

Charlie Dai, vice president and principal analyst at Forrester, said Anthropic’s concerns mirror challenges enterprises are already encountering as AI systems gain greater autonomy.

“Alignment becomes operational,” Dai said. “It is about ensuring agents consistently act within policy, not just model accuracy.”

Current governance approaches focus largely on models and data, but increasingly autonomous agents require oversight of runtime behavior, permissions, tool usage, and decision boundaries, Dai said.

Concerns about agent oversight are not limited to AI vendors and industry analysts.

In AI Agent Governance: A Field Guide, researchers from Institute for AI Policy and Strategy warned that “society is largely unprepared for this development” and said “the exploration of agent governance questions and the development of associated interventions remain in their infancy.” The paper argues that advances in autonomous AI agents are outpacing the governance mechanisms needed to oversee them.

Both analysts argued that governance frameworks originally designed for generative AI models may prove insufficient for increasingly autonomous systems. Dai said organizations will need greater oversight of runtime behavior, permissions, tool usage, and decision boundaries as agents become more capable.

Why Anthropic is worried

Anthropic’s researchers argue that those governance questions could become significantly harder if AI systems become increasingly involved in the process of AI research and development itself.

Favaro and Clark stopped short of predicting that fully autonomous recursive self-improvement is inevitable. Instead, they argued that the possibility warrants preparation and discussion among developers, policymakers, and other stakeholders. They also suggested the industry may eventually need mechanisms to slow development if capabilities begin advancing faster than safeguards, while acknowledging that such measures carry risks of their own.

“But if a slowdown simply lets the least cautious actors catch up technologically, it could leave everyone less safe,” they wrote in the blog post.

Forrester’s Dai said the practical implication for enterprises is that governance can no longer depend primarily on human review.

“Supervision becomes architectural, not manual,” he said. Organizations will increasingly need bounded autonomy, embedded guardrails, verifiable execution mechanisms, and fallback controls designed into agentic systems from the outset.

Over the next three to five years, both governments and the private sector will need to rapidly adapt identification and mitigation protocols as adversaries move from AI-assisted to AI-enabled sanctions evasion and proliferation financing (PF), a new research paper warns.

The report, Algorithms of Evasion: The Rise of AI-Enabled Proliferation Financing, from the Royal United Services Institute (RUSI), a UK-based defense and security think tank, defines PF as the use of funds or financial services to acquire, develop or otherwise deal in weapons of mass destruction (WMD). It states, “North Korea and Iran are now developing and deploying AI models to aid with sanctions evasion activities.”

Key findings include the fact that AI is now capable of mass producing high-quality fraudulent documents, as well as automating what the report describes as “the administrative minutia of managing extensive shell company  networks.” AI powered systems, it states, can also “analyze blockchain patterns in real time to dynamically adjust cryptocurrency mixing strategies, effectively evading detection tools.”

In addition, it says, “[tools such as generative AI] which can produce sophisticated fraudulent identification documents, for example, have helped North Korea perpetrate phishing attacks against Western companies.”

Dr. Aaron Arnold, senior associate fellow with the Centre for Finance and Security at RUSI, who authored the paper, said in an email that what prompted it was an uptick over the last year in North Korea’s use of AI to facilitate and enhance its cyber operations, in the form of phishing schemes designed to generate revenue for the country’s ballistic missile and nuclear weapons programs.

He advised enterprise IT managers who need to protect their organizations from becoming victims of sanction evasion activities that “[it] means largely adapting to a landscape where traditional human-focused security boundaries are being bypassed by automated technologies.”

For IT managers, said Arnold, “this might entail incorporating defensive AI, the use of behavior-based analytics, using ‘circuit breakers’ when there is heavy use of API or MCPs, updating personnel training, and hardening identity verification, especially for any remote hiring.” 

Distinction between AI-assisted and AI-enabled activity is ‘central’

Sanchit Vir Gogia, chief analyst at Greyhound Research, said that the RUSI report matters “because it names the right structural shift. AI is not creating sanctions evasion from thin air, it is compressing and scaling methods that already work.”

He pointed out that none of the sanction-evading techniques such as fraudulent documents, synthetic identities, shell companies, hidden beneficial ownership, crypto laundering, and others are new. “What changes is the speed, quality, volume and coordination with which these methods can now be assembled,” he said.

According to Gogia, “the distinction between AI-assisted and AI-enabled activity is central. AI-assisted evasion uses AI for discrete tasks: writing a better email, producing a cleaner document, generating a stronger false profile, translating a pitch, summarizing regulations or preparing a plausible job application. AI-enabled evasion is more serious.”

A ‘structural asymmetry’

This tactic, he said, “begins to coordinate the system itself. It links identity, documents, ownership structures, payment routes, cloud access, crypto wallets, API calls and timing. The difference is not whether AI helps someone fake a document. The difference is whether AI begins to orchestrate the deception.”

That is why the report’s findings should worry enterprise leaders, he noted: “Many organizations still assume the bad actor is mostly human, mostly linear and mostly slow. That assumption is expiring. AI lets adversaries run more attempts, with fewer errors, across more channels, in more languages, with better paperwork and greater patience than most enterprise review processes can absorb. This is not a tale of genius criminals discovering magic. It is the story of ordinary controls meeting industrialized plausibility.”

The evidence today, he pointed out, is strongest around tactics such as identity fraud, document fraud, synthetic personas, remote-worker deception, phishing, social engineering, crypto obfuscation and workflow abuse. “Fully autonomous evasion networks sit on the horizon,” he said. “They are serious, but they are not yet the everyday baseline.”

This distinction matters, said Gogia: “If enterprises obsess over cinematic autonomous agent scenarios while leaving remote hiring, vendor onboarding, payment approvals, and document review full of holes, they will lose in the most prosaic way imaginable.”

The report, he said, also gets the “asymmetry” right. “Offensive actors can learn across the ecosystem,” he said. “They can scrape open information, reuse leaked records, study enforcement patterns, test onboarding forms, inspect public procurement data, watch court filings, probe compliance thresholds and [use the information to] refine their behavior.”

Defenders, by contrast, are hemmed in by privacy rules, fragmented data, explainability requirements, jurisdictional boundaries, conservative operating models and siloed technology estates. “Offensive AI learns broadly,” he said. “Defensive AI often learns from fragments. That is the structural asymmetry.”

He explained that the regulatory landscape also amplifies the problem, in that regulatory bodies “still speak in separate dialects. [For example] the EU AI Act pushes organizations toward stronger obligations for high-risk AI. NIST-style frameworks push risk management, transparency, and governance.”

A trust architecture problem

Financial Action Task Force (FATF) expectations push national risk assessment and counter-proliferation controls, he noted, while banking regulators focus on model risk, accountability and operational resilience. “None of these streams is irrelevant. The trouble is that criminals do not organize themselves around regulatory workstreams. They organize around outcomes.”

What that means, said Gogia, “is that enterprise cannot wait for a clean global rulebook. It will not arrive in time. CIOs, CISOs, compliance officers and boards need a working governance model now. They need privacy-preserving analytics, controlled data environments, audit trails, legal safeguards and clear model-risk accountability.”

He said that enterprise IT managers should treat the situation as a trust architecture problem rather than a narrow sanctions-screening problem. “The uncomfortable truth is that AI is not simply helping bad actors write better phishing emails or forge tidier documents,” he noted. “It is helping them manufacture legitimacy across a chain of enterprise workflows.”

Likely outcome an ‘AI arms race’

Report author Arnold also noted that there are signs that cyber criminals have discovered new AI technologies and abilities that legitimate enterprises could adopt for legitimate applications.

History, he said, “is replete with [criminals] developing novel solutions to tough problems, [which are] later adopted by law enforcement. Much of our anti-financial crime policy is effectively a response to bad actors exploiting systems or using technology in novel ways to perpetrate crimes. In this scenario, I think an ‘AI arms race’ between enforcement authorities and bad actors is the most likely outcome.”

Gogia added, “the baddies are not teaching enterprises how to invent AI. They are teaching enterprises where trust is leaking. That is the lesson worth taking seriously.”

This article originally appeared on CIO.com.