The AI Agent That Works Is Not Necessarily Worth Running

Why operational controls and measurable outcomes must determine the next phase of agentic AI.

Organisations are discovering that successfully deploying AI does not necessarily mean they have created a successful business investment.

An AI agent can perform its assigned tasks, integrate with enterprise systems and operate continuously without delivering sufficient commercial value to justify its ongoing cost.

That distinction is becoming increasingly important as businesses move from controlled experiments towards autonomous workflows embedded in everyday operations.

Two research releases published this week illustrate the challenge.

Gartner reports that 95% of surveyed infrastructure and operations organisations experienced at least one agent-related incident in the preceding 12 months.

Separately, research commissioned by SnapLogic found that 54% of surveyed enterprise IT decision-makers reported their organisations had abandoned AI projects because the business value was unclear.

These findings come from different populations and measure different outcomes. Neither demonstrates that agentic AI is inherently ineffective.

They do, however, highlight an important issue.

Organisations are becoming increasingly capable of deploying AI agents without necessarily becoming equally capable of operating them economically, reliably and accountably.

For executives and boards, this changes the nature of the investment decision.

Technical success and business success are different things

Most enterprise AI initiatives begin with an identifiable opportunity.

A process takes too long. Employees spend considerable time retrieving information. Customer enquiries require repetitive administrative work. Data moves inefficiently between systems.

AI appears capable of addressing the problem.

A pilot demonstrates that the technology can complete the task, perhaps considerably faster than the existing process.

The pilot is declared successful and the organisation moves towards broader deployment.

The difficulty emerges when the technology becomes part of normal business operations.

The agent may need access to several enterprise applications. Its outputs require validation. Exceptions must be handled. Data permissions need maintaining. Model consumption increases with transaction volume.

Employees may continue checking the work because they do not yet trust the output sufficiently to change their established processes.

The original activity becomes faster, yet the surrounding workflow remains largely unchanged.

This is a familiar enterprise transformation problem.

Improving one activity does not automatically improve the end-to-end process.

The technology can function correctly while the business continues absorbing avoidable cost, duplicated effort and operational friction.

Where the economic case starts to deteriorate

SnapLogic’s November research reports that 64% of surveyed IT decision-makers experienced moderate or significant AI-agent overspending.

The same research found that only 42% believed they could accurately forecast future AI and token expenditure.

These are self-reported findings from a vendor-commissioned survey, rather than independently audited financial outcomes. Nevertheless, they raise a practical concern.

If organisations struggle to forecast the cost of operating AI agents, how confidently can they assess the return on expanding them?

The answer requires more than measuring tokens, API calls or subscription costs.

An economically credible assessment needs to consider the entire workflow.

That includes the cost of integration, infrastructure, monitoring, exception handling, employee supervision, assurance, remediation and ongoing maintenance.

It must also consider whether the expected benefit is genuinely realised.

Take a hypothetical customer service process handling 20,000 enquiries each month.

If automation reduces handling time by 90 seconds per enquiry, the theoretical capacity released is 500 working hours monthly.

That is a meaningful operational opportunity.

It is not automatically a financial saving.

If employees still review every output, the actual capacity released may be substantially smaller. If exceptions increase, some of the benefit may disappear entirely.

If the organisation cannot redeploy the capacity towards more valuable work, the commercial benefit may remain theoretical.

A defensible value case therefore needs to establish what happens to the time released, how quality changes, what the automated process costs and whether the overall business outcome improves.

The correct measurement is not simply the cost of running an agent.

It is the net value of the completed business outcome.

Governance needs to operate at the speed of the workflow

This week’s Gartner findings introduce another important dimension.

Only 15% of surveyed infrastructure and operations leaders reported having automatic circuit breakers capable of stopping runaway agent errors or loops.

That is significant because autonomous workflows can execute actions considerably faster than conventional human approval processes.

An agent with access to enterprise applications may retrieve information, update records, trigger transactions or initiate subsequent processes.

When something goes wrong, the consequences can extend beyond an inaccurate response.

The agent might repeatedly execute the wrong action, consume resources unnecessarily or propagate incorrect information through connected systems.

Traditional governance arrangements often assume that employees initiate actions and that exceptions can be identified through established management processes.

Autonomous systems challenge those assumptions.

They require controls that operate within the workflow itself.

An organisation needs to know which agents are active, what permissions they hold, what decisions they can make and what conditions require intervention.

Higher-risk activities may require explicit human approval. Lower-risk activities can operate within defined limits, supported by monitoring and auditability.

There must also be a practical mechanism for stopping or isolating an agent when it exceeds those limits.

These are not simply technical engineering considerations.

They determine how much operational responsibility the organisation is prepared to delegate and who remains accountable when the system fails.

Autonomy without clear decision rights is not a mature operating model. It is an accountability gap.

The workforce remains central to the value case

Introducing autonomous capabilities also changes how people work.

Employees may move from performing individual tasks towards supervising workflows, managing exceptions, validating outputs and improving processes.

That transition requires more than basic training on how to use an AI application.

People need to understand what the system is authorised to do, where its limitations lie and when human judgement is necessary.

Managers need to establish how responsibilities and performance measures should change.

Business leaders need to decide whether released capacity will improve service quality, reduce costs, support growth or address existing operational constraints.

Without these decisions, organisations risk operating two processes simultaneously.

The original human workflow remains in place while AI introduces an additional layer of activity.

Employees continue checking, correcting and duplicating work because the organisation has not formally redesigned responsibilities.

Adoption may appear healthy because the tools are being used.

The business benefit remains difficult to demonstrate.

This is why workforce enablement must be embedded into workflow redesign rather than treated as a separate communications or training exercise.

What TFx experience tells us

Across enterprise AI enablement and transformation work, one recurring lesson has been the importance of connecting technology deployment with the operating model around it.

TFx leadership experience includes supporting AI adoption across a large international enterprise, with structured enablement, governance considerations and workforce engagement forming part of the transformation approach.

The emphasis was not simply on giving employees access to AI capabilities.

It was on developing the organisational structures, support mechanisms and practical understanding required to use those capabilities effectively.

The same principles apply to autonomous workflows.

An organisation must understand the work before deciding how to automate it.

It must establish accountability before delegating decisions.

It must define expected outcomes before measuring performance.

It must equip employees to operate within the redesigned process rather than expecting them to adapt independently.

Most importantly, it must retain sufficient internal capability to manage, improve and govern the solution after initial implementation.

This is the distinction between deploying technology and building sustainable organisational capability.

A stronger investment test for enterprise AI

For boards, CFOs, CIOs and transformation leaders, the next stage of AI investment should involve a more demanding assessment of operational value.

Before approving expansion, organisations should be able to explain which business process is changing, what performance baseline exists and which outcome is expected to improve.

They should understand the full operating cost, including the activities required to supervise and maintain the solution.

They should identify the executive accountable for the workflow, the decisions delegated to AI and the circumstances requiring human intervention.

They should also be able to explain how the organisation will measure performance after deployment and what conditions would justify redesigning, restricting or discontinuing the capability.

These requirements need not create unnecessary bureaucracy.

Applied proportionately, they can help organisations concentrate investment on opportunities with credible benefits while preventing weak initiatives from consuming resources indefinitely.

A relatively simple internal knowledge assistant does not require the same controls as an autonomous system modifying customer records or initiating financial transactions.

The level of governance should reflect the autonomy, business impact and risk associated with the workflow.

The commercial test, however, remains consistent.

Does the capability deliver a measurable improvement in business performance after accounting for the cost and responsibility of operating it?

From automation activity to sustainable value

The next competitive advantage in enterprise AI will not necessarily belong to organisations deploying the greatest number of agents.

It will belong to those that can consistently identify valuable opportunities, redesign workflows, govern delegated decisions and demonstrate the resulting business improvement.

Some AI initiatives will justify expansion.

Others will need redesigning.

Some may prove technically impressive but commercially unattractive.

Recognising those differences is not a failure of AI strategy. It is evidence of disciplined investment management.

Enterprise transformation has always required organisations to distinguish activity from progress and expenditure from value.

Agentic AI makes that distinction more important because systems can now generate activity, operational dependencies and costs with considerably less human intervention.

The objective should not be to automate everything that can be automated.

It should be to build an organisation that performs better because automation has been applied deliberately, responsibly and economically.

If your organisation can demonstrate that an AI agent works, can it also demonstrate that the business is better off continuing to run it?

Sources

Verified by MonsterInsights