AI Agents Will Make Value Management Impossible To Ignore

Most enterprise software waits to be used. Someone opens the application, enters information, clicks a button and asks the system to do something. The software may automate parts of the process, but the person still decides when the work begins and what they want the system to produce. AI agents change that arrangement.

An agent can receive a goal, decide which steps to take, choose the tools it needs and act across several systems before a person sees the result. It might search a database, call a language model, update a customer record, trigger another workflow and ask a different agent to complete part of the task.

From the employee’s point of view, that may look like one request. From the organisation’s point of view, it could involve dozens of decisions, several technology services and a growing list of costs that aren’t obvious from the final answer. This is where AI value management starts becoming more complicated.

Enterprises already need to understand why an AI initiative exists, who owns it and whether it produces a useful outcome. Agents add another layer because the system can make decisions about how it pursues that outcome. The organisation may know what it asked for without knowing exactly how much work the agent will decide the task requires.

Traditional project reviews and annual return on investment calculations won’t provide enough visibility into systems that operate continuously. Leaders will need a clearer connection between an agent’s purpose, authority, cost and contribution while it’s working. Value management may stop being something organisations do around AI.

It may become part of how enterprise AI operates.

Why Agents Change The Value Equation

Generative AI tools usually respond to a prompt. A person asks a question, the model produces an answer and the interaction ends. AI agents can take that first instruction much further. They’re designed to work towards a goal by planning actions, using tools and adjusting their approach as the task develops.

That difference can sound small until you look at what happens inside the workflow. Imagine an agent responsible for resolving a customer complaint. It may need to identify the customer, review earlier conversations, check an order, read the returns policy, calculate an appropriate remedy and update several records. 

If the information is incomplete, it may search elsewhere or try a different route. Two complaints that look similar to a customer service manager could therefore consume very different amounts of computing power, employee attention and external services. Their outcomes may differ too.

One could be resolved automatically within minutes. Another might produce an incorrect refund, create follow-up work for an employee or escalate a customer who didn’t need escalating. Counting both as completed agent tasks would make their performance look the same when their business impact clearly wasn’t.

This makes agent value dynamic. It depends on what the system did, what the outcome was, what the process consumed and what happened afterwards. Scale will make that distinction harder to ignore. IBM’s 2026 Tech Leader Study found that enterprises expect to deploy an average of 1,661 AI agents by 2027, a 38 per cent increase from current levels.

Managing a few agents through individual teams may be possible. Managing hundreds or thousands through informal oversight probably won’t be.

The Coming Shift From AI Projects To AI Portfolios

Most organisations still discuss AI through projects. A team identifies a use case, develops or buys a system, completes a pilot and decides whether to deploy it. The work has a recognisable beginning, a budget and, ideally, an expected result. Agents won’t always fit that shape.

Some will be purchased as part of existing enterprise platforms. Others will be built by central AI teams, software developers or individual business units. Employees may create smaller agents through low-code tools without treating them as formal technology projects at all. 

Before long, several parts of the organisation could have agents solving similar problems in slightly different ways.

One customer service agent might summarise support cases. Another prepares recommended responses. A third updates records after a conversation. Six months later, a platform update may offer all three functions through one built-in service, yet the original agents could continue running because nobody has been asked to reassess them.

That is how useful experimentation becomes a portfolio-management problem. An agent inventory can show which systems exist, where they run and who created them. It can’t tell leaders which ones still deserve organisational resources and authority. For that, each agent needs a continuing reason to exist.

Some will become essential to important business processes. Others will lose value as workflows change, employees stop using their outputs or better systems replace them. An agent that was worth operating at launch isn’t automatically worth operating a year later. AI lifecycle management will therefore need to include more than technical maintenance. 

Enterprises will have to review whether an agent’s purpose remains relevant, whether another system duplicates its work and whether its current contribution justifies its cost. Without that discipline, organisations may keep funding autonomous systems long after the original business case has disappeared.

Why Cost Management Becomes More Complicated Than Budget Management

Enterprise technology budgets are never perfectly predictable. Still, traditional software gives finance and technology leaders a fairly familiar set of costs to work with. There are licences, implementation fees, infrastructure, support contracts and staff. Even cloud spending, which can fluctuate, usually connects to visible resources and workloads.

Agent economics are less straightforward. An agent may consume:

• Input and output tokens
• Multiple model calls
• External application programming interfaces
• Search and retrieval services
• Data storage and memory
• Monitoring and evaluation tools
• Human review
• Additional work to correct or reverse mistakes

A token is simply a small unit of text processed by an AI model. Providers charge different rates for the tokens a system reads and produces. That makes token economics part of AI cost management, but token prices alone don’t explain whether an agent is economical. A cheaper model may need more attempts to complete a task. 

A more capable model may cost more per call but resolve the work immediately. An agent could also spend very little on AI processing while creating expensive downstream work for employees. The FinOps Foundation’s 2026 research found that 98 per cent of respondents now manage AI spending, up from 31 per cent two years earlier.

It also identified AI cost management as a leading priority for the discipline. FinOps brings finance, technology and business teams together to understand and improve the value of technology spending. Its growing involvement in AI reflects a wider change. Agent costs won’t sit neatly inside one platform bill or technology department.

The more useful measure will be cost per outcome. For a service agent, that might mean the full cost of resolving one customer case without avoidable follow-up work. For a software agent, it could mean the cost of producing code that passes testing and reaches production. 

For a security agent, it may be the cost of investigating a genuine threat without creating unnecessary work around false alerts. This gives leaders a more honest view than cost per token or cost per interaction. It connects resource consumption to something the business was actually trying to achieve.

Why Value Management Must Move Closer To Execution

Most business cases are created before a system is deployed. They estimate what the organisation expects to spend, what it hopes to improve and how long that value should take to appear. Those assumptions are useful. They’re also vulnerable to reality. The process changes. Usage grows faster than expected.

 Employees use the system differently from the way its designers imagined. A model update affects performance. The agent gains another tool or starts handling more complex work. By the time a quarterly review finds the change, the system may have completed thousands of tasks under conditions that no longer match its original business case.

This is why value management will need to move closer to execution. That doesn’t mean putting a finance committee in the middle of every automated action. It means turning the organisation’s value expectations into practical operating controls. An agent could have:

• A spending limit for each task or period
• A maximum number of retries
• Defined outcomes and quality thresholds
• Clear actions it can take without approval
• Escalation triggers when costs or risks increase
• Performance monitoring that includes downstream results
• Automatic suspension when agreed conditions aren’t met

These controls help answer a question that normal reporting often misses: is the agent still operating within the conditions that justified giving it autonomy? An agent may perform accurately but cost too much. It may reduce handling time while increasing complaints. It may complete more work while employees spend longer checking the results.

None of those problems will appear on a dashboard that only counts tasks, users or hours saved. AI value management needs to connect what the agent does to what happens because it did it. That requires operational data, financial information and business outcomes to be reviewed together rather than reported by separate teams several months apart.

The New Questions Enterprise Leaders Will Need To Ask

Agent adoption is often measured through visible activity.

• How many agents have been deployed? 
• How many people use them? 
• How much money is the organisation spending?

Those questions help describe the size of the environment. They don’t show whether it’s working well. A stronger value-management model starts with a different set of questions.

What business outcome is this agent responsible for?

Every agent needs a purpose that can be stated in operational terms. “Improve efficiency” is too broad. Reduce the average time required to resolve a specific type of support case is clearer. So is decreasing the number of invoices that need manual correction. The outcome should connect to a real business priority and have a measurable starting point. 

Otherwise, the organisation has no reliable way to judge whether performance improved.

Who owns its performance?

Technical ownership isn’t the same as business accountability. An AI team may maintain the agent, while a customer service leader owns the process it affects. Finance may track its costs, and risk teams may define the limits on what it can do. Someone still needs authority to decide whether the agent should change, lose access or stop operating.

Without a named owner, poor performance becomes everyone’s concern and nobody’s decision.

What does success cost?

A positive result isn’t automatically an efficient one. Leaders need the full cost of reaching the outcome, including models, infrastructure, external tools, employee supervision and remediation. They also need to compare that cost with other ways of completing the work. The aim isn’t always to choose the cheapest option. 

A more expensive agent may produce a better result, reduce risk or free scarce employees for more valuable work. The organisation needs to understand the trade-off it’s making.

Under what conditions should it lose autonomy?

Autonomy shouldn’t be permanent. An agent may earn more freedom after demonstrating stable performance, acceptable costs and reliable decision-making. The same authority should be reduced if quality declines, the business process changes or the system begins creating consequences it can’t resolve safely.

This turns autonomy into something an agent continues to justify rather than something it receives once at deployment.

When should it be retired?

Every agent should have retirement criteria before it becomes part of normal operations. Duplicate functionality, falling usage and rising supervision costs are obvious signals. So is value decay, where an agent continues working as designed but the work itself is no longer as useful as it once was.

Retiring an agent shouldn’t be treated as failure. It may simply mean the organisation learned what it needed, changed the process or found a better way to produce the outcome.

Value Management May Become The Defining AI Discipline Of The Next Decade

Building agents is becoming easier. Enterprise software vendors are adding them to existing platforms. Developers can connect models to tools through standard interfaces. Low-code products are opening agent creation to employees who may never describe themselves as AI developers.

That abundance will produce useful systems. It will also produce duplication, forgotten experiments and automated activity that continues because stopping it requires more deliberate attention than starting it did.

The opportunity is still substantial. Capgemini estimates that AI agents could generate up to US$450 billion in economic value by 2028 through revenue growth and cost savings across the 14 countries it studied. Yet its research also found that only two per cent of organisations had fully scaled agent deployment at the time of the study.

The gap between potential and scaled value won’t be closed by adding more agents alone. It will depend on whether organisations can manage them as part of a connected operating environment. That means linking authority to accountability, consumption to outcomes and continued operation to evidence of useful contribution.

Some enterprises will approach this mainly as an AI governance problem. Others may leave it to finance, technology or individual business teams. None of those functions will have the complete picture alone.

The organisations that scale agents successfully are likely to be the ones that can bring those perspectives together before autonomous activity becomes too widely distributed to understand as one portfolio.

Final Thoughts: Meaningful Autonomy Depends On Measurable Value

AI agents represent a different stage of enterprise AI maturity. They don’t only generate content or help someone complete a task. They can make choices, use resources and influence work across systems that were previously controlled through separate teams and processes. That autonomy creates opportunity. It also makes value harder to see.

Annual budgets can show what the organisation planned to spend. Adoption reports can show how often agents were used. Project reviews can show whether an initiative met its original targets. None of them can answer, on their own, whether an agent still deserves the access, resources and trust it currently holds.

That will require value management to become a continuous operational discipline. One that can follow an agent from its original purpose through changing costs, expanding authority, evolving outcomes and eventual retirement. The enterprises that benefit most from agentic AI may not be the ones that deploy the largest number of agents. 

They may be the ones that can explain, at any moment, why each agent exists, what it’s contributing and whether it’s still worth keeping. As that management discipline develops, EM360Tech will continue examining the governance, economics and operational decisions shaping the next generation of enterprise AI.