Why agentic scale needs economic guardrails, not just technical controls.
Enterprise AI is entering a phase where the amount of work performed by machines can increase dramatically without a corresponding increase in human activity.
That creates an opportunity to redesign business operations. It also creates a financial problem that many organisations are not yet equipped to manage.
New research from Menlo Ventures estimates that enterprise AI deployment expenditure reached approximately $177 billion by September 2026, more than five times its prior-year estimate.
The research also reports that AI agents now account for around 80% of token consumption on OpenRouter, one model-routing platform.
These figures should not be confused with independently audited enterprise expenditure or a measurement of all global AI activity. Nevertheless, they illustrate an important shift.
AI consumption is increasingly being driven by systems performing work, rather than individuals directly interacting with models.
For executives responsible for enterprise transformation, that changes the economics of deployment.
A human employee normally operates within established organisational constraints: working hours, approval structures, budgets, responsibilities and management oversight.
An AI agent may operate continuously, initiate repeated actions and consume resources through multiple interconnected systems.
Without appropriate controls, the organisation can increase the volume of activity considerably faster than it increases the value created.
The next challenge is not simply making AI agents more capable. It is ensuring their activity remains economically justified.
The hidden cost of autonomous activity
Consider a business process involving customer requests.
An AI-enabled system receives a request, retrieves relevant information, identifies the issue and determines an appropriate response.
If the information is complete and the request falls within defined parameters, the system can resolve it.
What happens when the information is contradictory?
The agent might retrieve additional documents, call another model, repeat an analysis, query another system and attempt to reconcile the results.
Each action consumes resources. Some may improve the eventual outcome. Others may simply increase the cost of reaching the same conclusion.
In a conventional workflow, an employee may recognise uncertainty and escalate the matter to a colleague.
An autonomous workflow needs equivalent escalation logic, although the triggers and controls must be designed explicitly.
Otherwise, the system can continue processing an exception long after the additional activity has ceased to provide meaningful value.
This is not necessarily a failure of the AI model. It is a failure of workflow design.
The organisation has created an execution mechanism without defining the economic boundaries within which it should operate.
More automation does not automatically mean lower cost
One of the most persistent assumptions in enterprise AI is that automating an activity reduces the cost of delivering it.
Sometimes it does. However, the relevant calculation must consider the complete process.
An agent that reduces manual effort may introduce costs associated with model consumption, orchestration, integration, monitoring, exception handling, auditability and human intervention.
If those costs are lower than the value of the resources released or the additional business outcomes achieved, the investment may be justified.
If they are higher, the organisation may simply have replaced one form of operational expenditure with another.
The distinction matters particularly when scaling.
A workflow that appears inexpensive across a few hundred transactions may behave differently across hundreds of thousands.
Usage patterns, error rates, exception volumes and model-routing decisions can materially influence unit economics.
The business case must therefore remain connected to actual production performance.
The appropriate measure is not simply cost per AI interaction. It is cost per completed, acceptable business outcome.
That outcome might be a resolved customer issue, an accurately processed transaction, a completed compliance review or an operational incident resolved within agreed parameters.
The organisation should understand the full cost of producing it.
The difference between agent deployment and operating-model transformation
Research published this week by the Capgemini Research Institute highlights another important distinction.
Its World Cloud Report for Financial Services 2027 finds that 59% of surveyed financial institutions are piloting or deploying multi-agent AI, but only 10% report achieving enterprise-wide orchestration.
The findings relate specifically to financial services and should not be generalised to every industry. They nevertheless expose a practical challenge.
Deploying individual agents is different from redesigning the enterprise processes through which those agents operate.
An organisation might introduce separate agents for customer service, document processing, financial analysis and internal support.
Each may demonstrate technical competence. But the business still needs to determine how information moves between them, which systems they can access, who authorises consequential actions and how exceptions are managed.
Without those decisions, the enterprise risks accumulating isolated automation capabilities.
Some may duplicate activities. Others may create inconsistent outcomes. Some may introduce dependencies that increase operating complexity.
This is why enterprise architecture and operating-model design cannot be treated as separate workstreams.
The technology needs to reflect how the organisation intends to operate.
Governance must include economic authority
Governance discussions often concentrate on privacy, security, regulatory compliance and responsible AI. These are essential.
Autonomous systems also require economic boundaries.
An AI agent may be technically authorised to retrieve information, invoke services and perform business actions.
That does not mean it should be allowed to consume unlimited resources pursuing a particular objective.
The organisation needs to establish appropriate limits.
Those limits may include maximum expenditure per transaction, permitted model usage, escalation thresholds, retry limits and conditions under which execution must stop.
They should also reflect the business consequences of failure.
A low-risk internal information request may require relatively lightweight controls. A financial transaction, customer commitment or regulated decision may require substantially stronger assurance.
The objective is proportional governance.
Controls should be sufficiently robust to manage risk without introducing unnecessary friction.
Importantly, technical permission must remain distinct from business authority.
An agent’s ability to perform an action does not automatically establish that the action is appropriate, affordable or authorised in the circumstances.
The workforce remains part of the control system
Human accountability becomes more important as autonomous execution expands.
Employees and managers need to understand when they are expected to intervene, which decisions remain theirs and how to challenge uncertain outcomes.
This cannot be achieved through generic AI training alone.
The operating model must define the relationship between human judgement and automated execution.
For example, an employee responsible for resolving exceptions needs access to the information that explains why an agent escalated the case.
A manager approving an automated recommendation needs sufficient context to make a meaningful decision.
A business owner needs visibility into operating costs, service quality and outcomes. Finance needs evidence connecting expenditure to realised value.
Without these capabilities, human oversight can become little more than an additional approval step.
That may increase cost without materially improving accountability.
The objective should be meaningful human control, supported by reliable information and clear responsibilities.
The TFx operating-model perspective
At TFx Holdings, the starting point for enterprise AI transformation is the business workflow rather than the technology.
The objective is to understand how work currently moves through an organisation, where friction exists and which changes can improve measurable outcomes.
This approach draws on experience across enterprise IT and operational transformation, AI adoption, governance and cross-functional stakeholder engagement.
It recognises that deploying a technically capable system is only one component of sustainable change.
The organisation must also establish ownership, workforce readiness, decision rights, appropriate controls and an economic mechanism through which the improvement creates value.
For agentic AI, this becomes especially important.
Before automating a process, the business should establish what successful completion looks like.
It should determine which information is required, what authority can be delegated and which exceptions must be escalated.
It should also identify the cost of the current process and the financial or operational benefit expected from changing it.
Only then can technology options be assessed meaningfully.
This avoids treating AI consumption as an end in itself.
Building the value case around completed outcomes
A practical economic assessment begins with the unit of business value.
Consider an illustrative process handling 100,000 transactions annually.
If the fully loaded cost of completing each transaction falls from £4.00 to £3.20, while quality and service requirements remain unchanged, the gross operating-cost difference would be £80,000 annually.
That is a calculation, not a claimed TFx client outcome.
The net benefit would still depend on implementation expenditure, ongoing support, retained human costs, additional controls and whether the reduction is actually realised in the organisation’s financial performance.
The important point is the measurement approach.
The business evaluates the complete outcome rather than isolated technical activity.
That same approach should continue after deployment.
If consumption increases, the organisation should understand why.
If exception handling becomes more expensive, the workflow may need redesign.
If the model produces better outcomes at higher cost, the business should decide whether the improvement justifies the expenditure.
If the original economic assumptions no longer hold, the organisation must be prepared to change course.
The next stage of AI investment discipline
Gartner’s latest emerging-risk research identifies AI strategy miscalculation among the most significant concerns reported by enterprise risk leaders.
That is a useful signal.
The risk is no longer limited to whether organisations can deploy AI safely.
It increasingly concerns whether they are making the right investment decisions in the first place.
As autonomous systems become more capable, the consequences of weak investment discipline may become more significant.
Organisations will need to understand not only what their agents can do, but what they cost, what value they create and when they should stop.
That requires a combination of workflow design, governance, financial measurement and human accountability.
The objective is not to constrain useful innovation. It is to create the conditions in which innovation can scale without losing economic control.
If your AI agents can continuously consume resources and execute work, who is accountable for deciding when the next action is no longer worth its cost?
