Who is accountable when an AI agent acts?
Summary
AI agents are moving from recommending actions to taking them. That affects every part of a business, but marketing is where agents first gain authority over budgets, customer messages and the data behind revenue reporting. Most mid-market companies haven't decided who owns each agent, what it can spend or how its results will be measured. This article sets out six governance questions, three levels of financial authority and seven practical steps to start without an enterprise budget.
Why this matters
Only 21% of organizations have a mature governance model for agentic AI, while 74% expect to use it within two years.
Gartner expects 40% of enterprises to demote or decommission autonomous agents by 2027.
Marketing and sales are where leaders most often see revenue gains from AI, so that's also where an uncontrolled agent does the most visible damage.
Platform updates can add autonomy without anyone deciding to, as Google's AI Max upgrades show.
Only 28% of CIOs consistently measure both the operational performance and the business outcomes of all their agents, and governance is what makes that measurement possible.
Mid-market companies carry the same exposure as large enterprises without the compliance teams to manage it.
When an AI agent can approve its own spend
A marketing employee might recommend raising ad spend by $10,000. In most companies, an established budget or approval process decides whether that happens. If an AI agent can make the same change on its own, those financial controls still have to exist.
Agentic AI means AI systems that can make decisions, use business applications and complete tasks with limited human involvement. It affects every function, from finance to customer service. Marketing is where it reaches revenue first, so it's also where governance decides whether growth can be measured and repeated.
Most companies haven't built that governance yet. Deloitte's State of AI in the Enterprise 2026 report, based on a survey of 3,235 leaders in 24 countries, found that only 21% have a mature governance model for agentic AI, while 74% expect to use it at least moderately within two years.
McKinsey recently explored the banking version of this question in The Rise of the Agentic Coworker. The issues reach well beyond banking. Manufacturers, professional services firms, retailers, property management companies and other mid-market businesses face the same questions, often without enterprise-level resources.
Agentic AI is a whole-business issue
AI has been a productivity tool for most businesses. Employees research faster, automate repetitive tasks and spend less time on administration. That work still needs oversight, but errors are easier to contain when a person reviews the output before acting on it.
Agentic AI carries a different level of responsibility.
A sales agent connected to the CRM might identify prospects, update customer records, send follow-ups and book meetings.
A marketing agent might monitor campaign performance and adjust advertising budgets.
A customer service agent might resolve complaints, authorize refunds or change account details.
A finance agent might initiate payments or update financial records.
Those are decisions and actions, not productivity aids. They affect revenue, expenses, customer relationships and day-to-day operations. The approval logic that governs purchasing, pricing, customer communications and payments applies when an agent does the work.
Automation changes how authorization is enforced, but the need for it stays.
Why governance matters most in marketing
Marketing is usually where a company first gives AI real authority over money and customers. Agents adjust ad budgets, send follow-ups, personalize offers and update the CRM, often inside platforms the marketing team manages on its own.
It's also where the link to revenue is easiest to see. McKinsey's State of AI 2026 survey found that respondents most often attribute revenue gains from AI to marketing and sales. That makes marketing the biggest opportunity and the place where an uncontrolled agent does the most visible damage.
Marketing agents touch money and customers directly
An agent that adjusts spend changes the cost of acquiring a customer. One that rewrites ad copy changes what the brand says in public. One that scores and follows up with leads changes which prospects the sales team speaks to. Each decision feeds directly into revenue reporting.
Individual mistakes may look minor. Repeated across hundreds or thousands of automated actions, they can become material, especially when an agent works from inaccurate customer data.
Marketing automation is moving faster than readiness
Gartner's survey of 402 CMOs found that marketing leaders expect AI-driven automation of marketing work to more than double, from 16% in 2026 to 36% by 2028.
The CMO Survey, run by Duke University's Fuqua School of Business with Deloitte and the American Marketing Association, shows the same trend. Its January 2026 poll of 308 U.S. marketing leaders found that AI use in marketing has more than doubled in two years, and companies expect AI to power more than half of marketing activities within three years. Marketers also reported measurable gains in sales productivity, customer satisfaction and overhead costs.
Execution hasn't kept up. No marketing technology capability in the survey scored above 5 on a 7-point performance scale, and the limits were organizational: budgets, integration, bandwidth and talent. Gartner's 2026 CMO Spend Survey adds that CMOs put an average of 15.3% of marketing budgets into AI, yet only 30% say their organizations are ready to scale it. Gartner says that spending needs the right operating model, governance, data foundations and talent behind it.
Platform automation arrives without a decision
Ad platforms show how autonomy can arrive through an update. Google announced in April that Search campaigns using automatically created assets or campaign-level broad match would be upgraded to AI Max automatically, and those upgrades began on September 1. AI Max widens query matching and generates ad text, so a campaign can stay inside its budget and still behave differently within it.
A budget cap is one control. Governance covers the wider questions of who approved the change, what the system may do within the budget and who checks whether it's producing customers.
Governance makes marketing performance measurable
Marketing leaders carry growing responsibility for revenue growth. The CMO Survey found that most marketers focus on stronger performance tracking as their main way to demonstrate value, while collaboration with finance remains limited and the CMO and CFO partnership has improved only marginally.
Governance establishes what an agent is authorized to do. Measurement establishes whether that activity produces business value. A company needs both, and the first makes the second possible. A named owner, defined spend limits and an audit trail show what each agent did and what followed.
The measurement data shows the gap. A 2026 Harris Poll survey of 685 CIOs, commissioned by Dataiku, found that only 28% consistently measure both the operational performance and the business outcomes of all their agents. In McKinsey's State of AI 2026 survey, 37% of respondents attribute at least some EBIT impact to AI, essentially unchanged from last year even though more organizations are scaling AI.
McKinsey's high performers look different. They were about twice as likely as other respondents to have defined processes for measuring the impact of their AI initiatives, and much more likely to say they work to mitigate unauthorized or unintended AI actions. That's a correlation, not proof that governance causes the results, but it points in the same direction.
Activity reports can show thousands of completed tasks, hundreds of customer inquiries handled or a large volume of sales messages sent. None of that shows whether the activity produced incremental revenue, lower expenses or better profitability. Automation can also add cost through software subscriptions, integration, monitoring, exception handling and error correction. It can still deliver real value, but that value has to be demonstrated through business outcomes, not assumed from the volume of activity.
Why mid-market companies carry the risk without the resources
These risks land hardest on mid-market companies. Those with $5 million to $50 million in annual revenue often run sophisticated technology, including CRM and ERP platforms, marketing automation, ad platforms, accounting software, customer service applications and a growing number of AI-enabled workflows. They rarely have dedicated AI governance teams, risk specialists or large compliance functions.
Technology decisions tend to be shared among an owner, CEO, COO, finance leader, IT provider and several department heads. Individual teams often choose and manage their own software, especially when a new feature promises quick productivity gains.
That works when each application has a defined function and employees remain responsible for the decisions made inside it. It gets harder when AI agents connect across systems. A sales tool may have access to customer data, a marketing platform may be authorized to adjust spending and a finance application may be able to initiate payments.
Each system has its own permissions, monitoring and approvals, so the business may understand what each application does without seeing the full picture of authority granted across its environment.
That's why governance belongs in the leadership conversation. IT is responsible for technical access and security. Business leaders have to decide which decisions can be delegated and how much risk is acceptable.
Why an AI policy isn't AI governance
Many companies have started writing internal AI policies. They typically cover employee use of public AI tools, confidentiality, data protection and acceptable use.
Those policies are useful, but they weren't written for autonomous systems. A typical policy prohibits entering confidential customer information into an unapproved AI tool. It says little about an approved agent that already has direct access to the CRM, ad accounts or customer database. That agent can retrieve information and act without an employee supplying each instruction.
Six questions governance has to answer
Ownership:
Who is responsible for each AI agent, including its performance, permissions and business outcomes?
Authority:
What can the agent decide on its own, and what requires approval?
Access:
What information and systems can the agent use, and are those permissions limited to what it needs?
Oversight:
How is activity monitored, and when must an employee step in?
Accountability:
Can the business establish what happened, why an action was taken and who authorized it?
Control:
Can the agent be stopped, its access revoked and its actions corrected when something goes wrong?
These are familiar management responsibilities. Businesses already apply them to employees, financial controls, purchasing authority and access to sensitive information.
What's different is speed and scale. An AI agent can perform a large number of actions quickly, sometimes across several connected applications. An error that a manual approval step would have caught can spread across records, transactions and customers before it's caught.
Governance has to be built into how these systems operate. A written policy that relies on employees remembering it isn't enough.
Who owns an AI agent in marketing?
Accountability blurs easily when several departments and vendors are involved.
Take an agent that qualifies inbound leads, updates CRM records and sends follow-up messages. Marketing may have selected the technology. Sales owns the customer relationships. IT manages the integration. An outside agency or software provider maintains the automation.
If the agent starts sending inaccurate information to prospects or updating records incorrectly, who finds and fixes the problem? The answer shouldn't depend on which department notices first.
Every AI system that performs a meaningful business function needs a named internal owner. That person doesn't need to understand every technical detail. They do need to understand the system's purpose, the decisions it's authorized to make and the business outcomes it's expected to deliver.
Technical teams, agencies and vendors can support implementation, monitoring and security. They can't take over the company's responsibility for deciding how the technology operates in its business. A third party may run the system, but the business still needs visibility into its permissions, activities and results.
How much authority should a marketing agent get?
Treating every AI application as equally risky is a mistake. An internal assistant that summarizes public market research doesn't need the same controls as an agent that can change ad budgets or send messages to customers. Oversight should match what the system can access, what it can do and what an error would cost.
Gartner makes a related point. In May 2026 it predicted that by 2027, 40% of enterprises will demote or decommission autonomous agents because of governance gaps found only after production incidents. Its analysts say many enterprises treat agent governance as all or nothing, either locked down or fully trusted, and recommend classifying agents by autonomy level instead. McKinsey's banking guide describes the same spectrum, from advisers to executors, with controls rising as autonomy does.
Three levels of financial authority in marketing
For a mid-market business, the simplest way to set an agent's authority is by what it can spend, change or commit. Three levels cover most situations.
No spend authority:
The agent reports and recommends. It might flag an underperforming campaign and suggest moving budget, but people decide.
Spend with approval:
The agent prepares the action and a person approves it before it goes live. Examples include budget reallocations, new ad copy, material changes to bidding targets and new lead follow-up sequences.
Spend within a cap:
The agent acts up to a defined limit. It might shift budget within a daily cap, pause ads that exceed a cost-per-lead ceiling or send pre-approved follow-up messages. Anything beyond the limit goes to an employee.
Dollar limits aren't the only factor in marketing. Agents also need rules for what they can say in the brand's name and which customer data they can use. An agent with no spend authority can still publish the wrong message or expose customer information.
Earning more authority
Adding an approval step to every workflow would erase much of the efficiency automation is meant to deliver. Companies need to decide where human judgment is essential and where automation can run safely inside defined boundaries.
Moving an agent up a level, or raising its cap, should follow demonstrated reliability, testing and appropriate controls. Deloitte found that companies succeeding with agentic AI tend to start with lower-risk use cases, build governance capabilities and scale deliberately.
McKinsey's banking guide likewise recommends releasing autonomy in stages. Unrestricted access shouldn't be the default just because the technology allows it.
Seven steps to start without an enterprise budget
A large governance department isn't required to establish reasonable oversight. The NIST AI Risk Management Framework and ISO/IEC 42001 offer established approaches to managing AI risk. NIST's framework is voluntary. ISO/IEC 42001 is a certifiable standard that applies to organizations of any size. NIST has also started agent-specific work through its AI Agent Standards Initiative, launched in February 2026, which covers agent security, identity and authorization. That guidance is still in development.
A mid-market company can borrow the underlying principles and scale them to its own operations, resources and exposure.
Establish visibility into AI use.
Keep a record of the AI applications and agents used across the business, including autonomous features inside ad platforms, the CRM and other existing tools. Note what each one does, what it can access and which department is responsible for it.
Assign business ownership.
Give each significant AI system a named internal owner who is accountable for its purpose, operating boundaries and performance.
Define decision authority.
Decide which activities can be automated, what spend and messaging limits apply and which decisions need human approval.
Control access to information and systems.
Limit permissions to the data and functions each agent needs for its approved responsibilities. Review them when systems or business needs change.
Monitor actions and exceptions.
Keep records sufficient to investigate significant decisions, spot unexpected behavior and confirm the controls are working.
Measure business performance.
Tie each agent to pipeline and revenue outcomes, including the cost of running and supervising it.
Establish intervention procedures.
Make sure employees know how to suspend an agent, revoke its access, escalate a problem and correct or reverse actions where possible.
These responsibilities can be shared among existing leaders and supported by technology providers. What matters is that they're formally assigned and consistently applied.
Governance is ongoing
Governance isn't a one-time task completed at implementation. AI systems, integrations, data sources and business needs change. Permissions that made sense when an agent was introduced may not make sense six months later. Periodic reviews matter most when an agent gains access to more systems or more autonomy.
The next AI investment decision should include governance
Pressure to adopt AI quickly is real, especially for mid-market companies competing with larger organizations.
Adoption and readiness aren't the same thing. A business can have the technology to deploy an autonomous agent without the controls to supervise it. None of that changes because the work is automated. Customers expect accurate information, transactions need authorization, sensitive data needs protection and management stays accountable for results.
As agents grow more capable, leadership teams need to answer a wider question than which activities can be automated.
Which decisions is the business prepared to delegate?
How much authority should these systems receive?
What evidence will show they're operating responsibly and producing value?
We wouldn't hand a new employee unrestricted access to customer records, financial systems and purchasing authority without defining their responsibilities. It's reasonable to expect comparable controls when an AI agent takes on those responsibilities.
For marketing leaders, those controls also make revenue growth provable. The teams that can show what each agent was allowed to do, what it cost and what it produced are better placed to earn that trust, and more budget, from finance and the CEO.
The businesses that handle this transition well won't necessarily deploy the most agents. They'll know what their agents are doing, keep control of the decisions that matter and be able to show the investment is improving performance.
Which of those three questions has your leadership team already settled, and which is still open?
Frequently asked questions
What is agentic AI governance?
It's the set of decisions that define who owns each AI agent, what it's allowed to do, what it can access, how it's monitored and how it can be stopped. It covers the whole business, but it matters first in marketing, where agents touch budgets, customers and revenue data.
Who is accountable when an AI agent makes a mistake?
The business is, and inside the business it should be one named person. An agent can't be held accountable, so each one needs an internal owner who understands its purpose, its limits and the outcomes it's expected to deliver. Agencies and vendors can run the system, but they don't take over that responsibility.
What's the difference between an AI policy and AI governance?
A policy tells employees how to use AI tools. Governance controls what an approved agent can do once it's connected to the CRM, the ad accounts or the customer database. Most policies were written for the first situation and say little about the second.
How much spending authority should a marketing agent have?
Start with the lowest level that does the job. The three levels are no spend authority, spend with approval and spend within a cap. An agent moves up only after it has shown reliable results, and rules for messaging and customer data apply at every level.
Where should a mid-market company start?
With an inventory of every AI tool and agent in use, including autonomous features that arrived through platform updates. Then assign a named owner to each significant one and set its limits. The seven steps in this article can be shared among existing leaders, so no dedicated governance team is needed.
How do we know AI agents are producing revenue?
Governance records what each agent was allowed to do, what it cost and what it produced, and those records are what make measurement possible. Without them, activity reports show volume but not whether revenue grew. Only 28% of CIOs in the Dataiku and Harris Poll survey consistently measure both operational performance and business outcomes.
Sources
McKinsey & Company, The Rise of the Agentic Coworker: A Banking CIO's Guide to the Next Wave of AI, October 26, 2026
Gartner, Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure, May 26, 2026
Deloitte, State of AI in the Enterprise: The Untapped Edge, January 2026
Dataiku and The Harris Poll, Global AI Confessions Report: CIO Edition 2026, September 24, 2026
McKinsey & Company, The State of AI in 2026: On the Road to ROI, August 25, 2026
Gartner, Gartner Survey Reveals Marketing Leaders Expect AI Automation of Marketing Work to Double by 2028, May 11, 2026
Gartner, Gartner 2026 CMO Spend Survey Finds CMOs Allocate 15.3% of Marketing Budgets to AI, But Only 30% Are Ready to Scale AI Capabilities, May 11, 2026
Duke University Fuqua School of Business, Deloitte and the American Marketing Association, The CMO Survey 35th Edition, March 31, 2026
Google, We're Upgrading Dynamic Search Ads to AI Max, April 15, 2026
National Institute of Standards and Technology, AI Agent Standards Initiative, February 2026
National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023
ISO and IEC, ISO/IEC 42001:2023 Information technology, Artificial intelligence, Management system, December 2023
Comments