As AI systems take on more decisions and tasks, companies face difficult questions about cost, responsibility and the people needed to keep them working.
Agentic AI is changing the question businesses ask when they buy technology. For years, the decision centred on what software could help employees do. Increasingly, it concerns what a system can be trusted to do on their behalf, how much freedom it should receive, and who answers when its work goes wrong.
Consider a customer requesting a refund. Conventional software gives an employee the records and tools needed to process it. An AI agent might examine the purchase, interpret the policy, request missing information, and initiate an authorised payment. The customer sees one interaction. Behind it sits a chain of decisions crossing several systems.
That changes what a company is buying. Access to software remains part of the transaction, but so do delegated responsibility, continuing supervision and the expertise needed to handle exceptions.
Agentic AI Will Live Alongside Older Systems
In a recent Forbes commentary, Everest Group founder and executive chairman Peter Bendor-Samuel argues that businesses are moving along two technology paths. Existing systems will increasingly benefit from AI, while environments designed around agents will develop beside them.
The argument avoids an expensive assumption: that cheaper code makes established business systems disposable. A payroll platform contains more than programming. It carries years of policy decisions, connections to other systems, and procedures employees depend on. Replacing it involves rebuilding those relationships and proving that the replacement works.
An agent could make such a platform easier to use without replacing the underlying records or calculations. Elsewhere, companies could design new processes in which agents decide which tools to use and what steps to take.
Anthropic’s engineering guidance draws a useful distinction here. A predefined workflow follows an established sequence; an agent can determine its route according to the task and the information it encounters. The company also advises developers to begin with the simplest workable approach, adding complexity only when it improves results.
For businesses, this means choosing autonomy carefully. A predictable task may be served perfectly well by conventional automation.
Responsibility Continues After Installation
Traditional enterprise software has never been entirely static. It receives updates, depends on changing data, and sometimes fails. Agentic AI adds another source of variation: the system can choose different actions while pursuing the same broad objective.
A successful demonstration therefore offers limited reassurance about everyday performance. A refund agent might handle standard purchases well, then struggle when a customer combines a damaged item, a promotional discount, and an expired return period.
The practical questions become specific. Which records may it read? What may it change? When must it ask a person? Can the company reconstruct what happened and correct an error?
Microsoft’s guidance on agent responsibility identifies additional concerns arising from agents’ ability to invoke tools, retain memory, and interact with other agents. Responsibilities also vary according to how the system is deployed.
The business implication is that supervision must have an owner. Technical teams can maintain the system, but the department using it must define acceptable decisions. A customer service manager cannot hand over that judgement simply because the software comes from an outside supplier.
Human Expertise Becomes Part Of The Product
This continuing involvement complicates the familiar division between software companies and service providers.
Historically, customers could buy a product from one supplier and commission another to install or adapt it. Bendor-Samuel’s argument is that some expertise becomes inseparable from an agentic product’s operation.
There is a practical reason. An engineer who understands a customer’s approval rules, data problems, and unusual cases may be essential to keeping the system useful. That knowledge accumulates through operation; it does not fit neatly into an installation manual.
External consultants will still have work. The boundary between product development, implementation, and ongoing operation, however, may become harder to draw.
This also changes the staffing calculation. Automation can reduce routine effort while increasing demand for people who understand both the technology and the business process. Moving experienced employees away too early could remove precisely the knowledge needed to judge whether the system is performing properly.
Agentic AI Pricing Is Already Changing
Charging for each employee with a software account is straightforward when people do most of the operating. It becomes less revealing when software performs work with limited human involvement.
Counting agents is no obvious improvement. A system could create several temporary agents to complete one assignment. Their number would say little about the quality or commercial value of the result.
Suppliers are already trying alternatives. Salesforce’s Agentforce uses Flex Credits to meter actions, including retrieving information or updating a record. Intercom advertises Fin pricing from $0.99 per outcome, alongside seat charges for its helpdesk plans.
These approaches measure different things. An action is a unit of activity. An outcome is a contractually defined result. Buyers need to understand the difference before comparing prices.
Intercom’s published definition, for example, includes confirmed resolutions, cases where a customer does not request further help after a response, and completed procedures, including handoffs. A billable outcome therefore requires closer examination than the label alone suggests.
For procurement teams, the important figure is the cost of acceptable work. That calculation should include human review, repeated attempts, unresolved cases, and corrections. A low headline price offers little comfort if employees must routinely finish the task.
Better Margins Remain A Proposition
Integrating specialist services with software may also change how these businesses earn money.
Bendor-Samuel suggests that AI could allow relatively small teams to support substantial operations, bringing the combined offering closer to software economics than conventional consulting. It is a plausible direction, but the margins remain uncertain.
The test is whether expertise becomes reusable. If each new customer needs months of individual engineering and intensive support, growth will continue to require considerable human effort.
Computing costs also deserve attention. Anthropic reports that its research systems using multiple agents consumed roughly 15 times as many tokens as ordinary chat interactions. That finding concerns a particular application, but illustrates why more capable automation does not automatically mean inexpensive operation.
Buyers and suppliers will need evidence from sustained use. Successful demonstrations cannot settle questions about support costs, reliability, or profitability.
The Supplier Relationship Gets Harder To Unwind
As providers become more involved in daily operations, continuity becomes valuable. A stable team learns why exceptions exist and which apparent inefficiencies protect the business from mistakes.
That familiarity can improve performance. It can also make changing suppliers difficult.
Companies adopting agentic AI should therefore preserve their own understanding of how the work is done. Decision rules, performance records and operational knowledge need to remain accessible to the organisation, rather than residing entirely with a vendor’s specialists.
The immediate future is likely to be untidy. Established applications will coexist with agents, employees will oversee some decisions and make others, and commercial models will continue to change.
The purchase decision will have to accommodate that reality. Before expanding an agent’s authority, a business needs to know what good performance looks like, who can intervene, and what the full cost of dependable work will be. Those answers will determine whether autonomy becomes useful operating capacity or another system employees must quietly work around.
Stay connected with Business Herald for the latest business news, insights, and updates.
Follow us on Facebook, Instagram, LinkedIn, and YouTube.
Join our growing community on WhatsApp and Telegram for real-time updates delivered directly to you
