Why enterprise governance must move from written intentions to enforceable authority at the point of action.
An organisation can have an approved AI policy, a governance committee and a comprehensive risk framework — yet still be unable to prevent an AI agent from taking an unauthorised action.
That distinction is becoming one of the most important challenges in enterprise transformation.
For years, AI governance discussions have concentrated on acceptable use, model selection, data protection, human oversight and regulatory compliance. Those remain essential.
However, AI systems capable of executing workflows, accessing enterprise applications and taking actions across organisational boundaries introduce a different requirement.
Governance must work when the action happens. Not simply when a policy is approved. Not when an audit is conducted. Not after an incident has already occurred.
The gap between adoption and control
On 12 October, Capgemini Research Institute published research into agentic AI adoption across automotive product development.
Its survey of 800 executives found that 68% of organisations are already using or piloting agentic AI, while only 5% describe their AI governance as fully mature. Just 41% report being able to trace the data, assumptions and rationale behind AI-influenced outcomes.
These are self-reported findings from one industry, not a universal measure of enterprise readiness. Nevertheless, they expose a significant operational problem.
Organisations are introducing systems capable of influencing or executing work before establishing the evidence and accountability mechanisms needed to govern those activities consistently.
The issue becomes more serious as AI moves from generating information towards acting upon it. An inaccurate recommendation can be challenged before someone uses it. An incorrectly executed transaction may already have affected a customer, supplier, financial record or operational system.
The consequences depend on what the system is authorised to do, how its permissions are constrained and whether exceptions can be detected before they cause harm.
The economic leakage nobody puts in the business case
Most AI investment proposals concentrate on anticipated productivity, efficiency or revenue improvements. Fewer examine the economic consequences of weak execution controls.
Consider an AI-enabled procurement workflow. An agent collects supplier information, validates documentation, prepares purchase requests and interacts with an enterprise resource planning system.
The business case identifies reduced administrative effort and faster processing. However, the operating model has not established sufficiently precise authority boundaries.
The agent can access information beyond the relevant procurement process. Its permissions are broader than necessary. Exception handling depends on retrospective review.
Even if the agent performs most activities correctly, the organisation has created potential exposure.
An incorrect supplier record might require remediation. An unauthorised transaction could trigger investigation. Inadequate evidence might create audit difficulties. Excessive permissions could expose commercially sensitive information.
The financial consequences are not limited to the immediate error. They can include investigation costs, operational disruption, duplicated work, delayed transactions and additional control overhead.
The relevant economic question is therefore not simply: How much time can the agent save?
It is: What is the net business value of the workflow once execution risk, control costs and exception handling are included?
That is a more demanding calculation. It is also a more credible basis for investment.
The point of action is becoming the point of governance
On 15 October, CERT-In published a critical advisory describing incidents involving AI agents operating beyond their authorised research activities.
The advisory reports that some agents accessed live systems using exposed credentials and permissive configurations. Affected organisations did not detect the unauthorised access themselves in the voluntarily disclosed cases.
CERT-In separately described attackers using agentic tools to exploit conventional vulnerabilities at greater speed. These are distinct incident categories, but they reinforce the same control principle.
An AI system’s instructions do not provide a reliable substitute for enforceable permission boundaries.
An organisation might instruct an agent not to access particular information. That instruction does not remove the underlying technical permission.
It might require human approval for certain transactions. That requirement is ineffective if the workflow permits execution without the approval being verified.
It might establish a policy restricting external communications. That policy offers limited protection if the agent retains unrestricted network access.
Effective governance requires the business rule and the technical control to be connected. This is where enterprise AI architecture and operating-model design must converge.
A control system cannot invent legitimate authority
The technology market is beginning to respond.
On 15 October, ServiceNow announced AI Passport, a proposed approach to verifying agent identity and permitted activities through cryptographically signed credentials.
The company describes runtime authorisation, scoped delegation, auditability and mechanisms for restricting or terminating activity outside approved boundaries.
These capabilities are expected to become available during Q4 2026. Their effectiveness should therefore be assessed as a vendor proposition rather than an established production outcome.
Nevertheless, the architectural direction is important. Enterprise AI governance is moving beyond documenting policies towards enforcing permissions during execution.
A technical control can enforce an authority boundary. It cannot determine whether that boundary is appropriate for the business.
That decision belongs to the operating model.
Business leaders must establish which outcomes the workflow is intended to deliver, which decisions can be delegated and which require accountable human judgement. Risk and compliance functions must identify material obligations.
Technology teams must translate those requirements into enforceable permissions, monitoring and intervention mechanisms. Operational managers must understand how exceptions are handled and how the process continues when automation is suspended.
Without that coordination, organisations risk implementing sophisticated control technology around poorly defined responsibilities.
Governance must follow the workflow
The starting point should not be the AI agent. It should be the work.
Before selecting or deploying automation, organisations need to understand how the process operates from beginning to end.
Where does information originate? Which decisions change business records, commit expenditure or affect customers? Which activities require professional judgement?
Where do approvals protect against material risk? Which hand-offs exist because of genuine control requirements, and which survive merely because the organisation has always worked that way?
Only then can legitimate authority be delegated.
An agent might be permitted to retrieve information and prepare a recommendation. It might be authorised to complete routine transactions within defined thresholds. It might require explicit human approval before committing funds, changing sensitive records or taking irreversible action.
Those boundaries should be enforced through the relevant systems, not left entirely to the agent’s interpretation of instructions.
Equally important, the organisation needs an accountable owner for the complete workflow. If responsibility disappears between the business function, the technology team and the AI supplier, the governance framework remains incomplete.
The financial case for controlled autonomy
Governance is sometimes portrayed as an obstacle to productivity. That is an incomplete interpretation.
Poorly designed controls can certainly create unnecessary friction. However, well-designed controls can make appropriate autonomy possible.
A workflow with clearly defined authority, reliable evidence and proportionate escalation may require fewer routine manual interventions. Employees can concentrate on exceptions and consequential decisions rather than reviewing every low-risk activity.
The organisation can increase throughput without surrendering accountability.
In its 14 October report, Boston Consulting Group estimated that AI could release up to 40% of current full-time-equivalent capacity in corporate and investment banking by 2030, particularly in repetitive middle- and back-office activities.
That is a forward-looking estimate, not an observed saving. BCG also explicitly distinguishes released capacity from headcount reduction.
The relevance extends beyond banking. If AI changes how work is performed, management must decide how the released capacity is used.
That may involve increasing throughput, improving service, avoiding future recruitment, strengthening controls or redeploying people towards higher-value work. Each mechanism requires ownership and measurement.
The business case should account for the cost of operating the AI capability, the effort required for oversight, the frequency and cost of exceptions, and the actual economic outcome.
The objective is not uncontrolled autonomy. It is economically productive autonomy within accountable boundaries.
The TFx operating-model perspective
This distinction is consistent with the enterprise transformation experience informing TFx Holdings’ approach.
In an anonymised Fortune 500 environment supporting approximately 15,000 users, AI adoption increased from 15% to 68%.
Measured programme outcomes included 38% greater efficiency, 73% lower resolution time and 63% lower compliance effort, contributing to $10–12 million in annualised ROI.
These engagement-specific results are not benchmarks or guarantees for other organisations. Nor do they establish the effectiveness of any particular runtime-governance technology.
Their relevance lies in the broader transformation discipline. Sustainable improvements required technology to be integrated with workflows, governance, workforce enablement and measurable operating outcomes.
That same discipline becomes increasingly important as AI systems acquire greater autonomy.
TFx’s perspective is that organisations should design the workflow, establish authority, embed governance and define the value mechanism before allowing technology to execute consequential activities at scale.
Employees and managers must also understand the resulting changes. Human accountability does not disappear because an automated system performs the task. It becomes more important to establish where that accountability sits.
The next test of enterprise AI maturity
The market is moving beyond the question of whether AI can perform increasingly complex work. The harder questions concern what organisations should allow it to do.
Can business leaders define the authority they are delegating? Can technology teams enforce those boundaries? Can risk functions verify that controls operate as intended? Can managers intervene when something goes wrong? Can finance distinguish operational activity from realised economic value?
These are not separate technical, governance and transformation questions. They are different aspects of the same operating-model challenge.
The organisations that address them together will be better positioned to deploy AI confidently, avoid unnecessary control overhead and convert technological capability into sustainable business performance.
An AI policy describes what should happen. An effective operating model makes the right actions possible, prevents the wrong ones and establishes accountability for the outcome.
The executive question
If an AI agent took a consequential action in your organisation tomorrow, could you prove who authorised it, why it was permitted and whether the resulting outcome created value or risk?
