When Does Agentic AI Actually Pay?

It is a mix of both. There is genuine potential in agentic AI, but the maturity of implementations today does not necessarily match the level of excitement surrounding the technology. The underlying concepts are also not completely new. Many of the ideas we now associate with agentic AI have existed in different forms for roughly a decade. What has changed is how these capabilities are being packaged, orchestrated, and positioned.

If you look at how analytics has evolved, there is a clear progression. In the early 2010s, a lot of analysis happened in Excel. Then organisations moved towards dashboarding platforms such as Tableau and Power BI. From there, statistical and machine-learning tools such as SAS and Python became more prevalent, followed by robotic process automation. Agentic AI is another step in that evolution rather than a complete break from everything that came before it.

Functionally, there are similarities between agentic AI and RPA. The difference is that we are now talking about multiple specialised agents or sub-agents working together, often coordinated by a supervising or “master” agent. That creates significantly more flexibility, but it also introduces complexity around reliability, governance, infrastructure and cost.

That is why production success is still relatively limited. At present, only around 10–15% of organisations attempting agentic AI implementations are seeing those initiatives succeed in production. Many organisations can demonstrate something interesting in a proof of concept, but moving from that POC to a reliable production system—and then justifying the return against the cost of operating it is a much more difficult challenge.

So there is a real business opportunity here, but enterprises need to separate the potential of the technology from the current hype around it.

The clearest results today are coming from repetitive, customer-facing workflows, particularly those that sit outside the most tightly regulated parts of the financial-services business.

WhatsApp-based customer journeys are a good example. BFSI organisations are already using these kinds of interfaces to create smoother customer interactions. Contact-centre automation is another area where agentic approaches can work effectively because there are large volumes of repetitive activities that can be understood, automated and measured.

These are relatively good environments in which to introduce agentic AI because the workflow is repeatable and the business outcome can be identified more clearly. If an organisation can automate part of a customer journey, reduce the amount of manual intervention required, improve response times or reduce operating costs, there is a clearer path towards demonstrating value.

The situation becomes much more complicated when you move into core decision-making processes. Underwriting and fraud detection, for example, involve considerably more judgment, regulation and risk. From my experience, only around 30% of underwriting decisions are currently being handled by AI, while approximately 70% still require human judgment.

Fraud detection has similar limitations. An agentic or AI-based system may be able to identify a transaction as high-risk or surface patterns that warrant further investigation, but determining with confidence whether something is actually fraudulent still frequently requires human review. Traditional rule engines therefore continue to play an important role.

This is an important distinction when enterprises evaluate agentic AI. The question shouldn't simply be, “Where can we deploy an agent?” It should be where the characteristics of the workflow allow the technology to create a measurable outcome without introducing disproportionate regulatory, operational or reputational risk.

There are three major issues that compound when agentic AI enters core BFSI processes: regulation, reliability and cost.

Regulation is particularly important because financial institutions cannot simply adopt whichever model produces the strongest technical performance. There are restrictions around where and how data can be processed, which models can be used, and how automated decisions are governed.

For example, RBI guidelines have restricted the use of certain large language models because of concerns around data processing outside India. That can push institutions towards alternative LLMs that may produce weaker results. At the same time, regulators have raised concerns about black-box models and the difficulty of understanding exactly how an automated system arrived at a particular decision.

The idea of a “kill switch” is relevant here as well. If an autonomous system begins behaving unexpectedly, institutions need the ability to stop it. That requirement is difficult to reconcile with highly opaque agentic architectures where multiple agents may be interacting and making decisions.

Reliability creates another barrier. Hallucinations remain one of the first concerns raised by leadership when organisations discuss taking these systems into production. A company can spend considerable time and money developing a solution, but a single hallucinated response in a customer-facing environment can create enough reputational risk to prevent the system from being productionised.

This is particularly significant in BFSI because the consequences of an incorrect answer are not limited to a poor user experience. There can be financial, regulatory and reputational consequences.

That is why traditional approaches have not simply disappeared. Rule engines continue to be used for critical applications such as fraud because they offer greater predictability and control. AI can augment those systems, surface risk indicators and support decision-making, but complete autonomy is much harder to justify when the cost of an incorrect decision is high.

Human judgment therefore remains an important layer. The technology can narrow the problem, identify signals and automate parts of the workflow, but that does not necessarily mean the entire decision should be handed over to an agent.

One of the biggest challenges with agentic AI is that the economics of a proof of concept can look very different from the economics of running the same system continuously in production.

Token consumption is becoming a meaningful part of that discussion. Leadership teams are beginning to look closely at how many tokens individual POCs consume because a seemingly small cost per token can become significant when millions of tokens are being consumed.

A token might cost roughly $0.01, but even a relatively straightforward proof of concept can consume millions of tokens. Once you start scaling that system, adding multiple agents and running it continuously, infrastructure and inference costs become difficult to ignore.

This is leading to a very practical comparison: what does it cost to run the AI system versus what does it cost to have a person perform the work?

In some cases, we have found that paying a data scientist is currently more cost-effective than paying for the infrastructure and token consumption required by the agentic solution. That does not mean the technology has no value. It means organisations have to be disciplined about where they use it.

ROI should therefore not be measured differently simply because the technology is agentic AI. The same principles that applied to earlier generations of analytics still apply. Whether we were using logistic regression, deep neural networks, LSTMs or today's agentic systems, ultimately the business question remains the same: does this increase the top line or improve the bottom line?

On the revenue side, that could mean generating additional business. On the cost side, it might mean reducing fraud, lowering customer-acquisition costs, removing repetitive work or improving operational efficiency.

There is also a growing need for FinOps-style transparency around AI. Enterprises need to understand not just whether a POC works technically, but what it will cost when deployed continuously at production scale. A successful demonstration is not automatically a sustainable business case.

Agentic AI therefore shouldn't be given a special or looser standard for ROI. If the technology cannot demonstrate meaningful revenue growth, cost reduction or another measurable business outcome relative to what it costs to operate, enterprises should question whether that particular use case is ready to scale.

The biggest mistake is trying to automate an entire large process in one go.

Consider an organisation attempting to automate a complete BI workflow. That could include BRD generation, dashboard automation, insight generation and several other activities. If all of those components are implemented as one large agentic AI project, it becomes very difficult to determine exactly where value is being generated.

A better approach is to break the larger initiative into four or five smaller projects that can be measured independently.

Each project should have an identifiable business outcome. You can then determine whether that particular component is creating a top-line or bottom-line impact. Once the individual projects have been evaluated, those results can be aggregated and presented to leadership as the overall business case.

This approach also makes experimentation more manageable. If one component fails to demonstrate sufficient value, that does not necessarily invalidate the entire initiative. The organisation can understand what worked, what didn't, and where additional investment is justified.

It also helps separate technical success from business success. A component may work perfectly from an engineering perspective and still fail to generate enough economic value to justify its production cost.

The rollout strategy should therefore be incremental and measurable: divide the problem, validate each component, quantify its impact and then expand based on evidence rather than attempting a large-scale implementation because the technology itself appears promising.

The best place to start is generally at the edges of a business process rather than at its regulated core.

Peripheral and non-regulated workflows provide organisations with an opportunity to test agentic AI in environments where the risk of failure is lower. Customer journeys, repetitive service activities and other clearly defined workflows can provide early wins while allowing teams to understand how the technology behaves in a real operational environment.

Once those implementations demonstrate value, the organisation has a stronger internal case for expanding agentic AI into more complex areas.

Teams should also be careful about choosing use cases with extremely high token consumption in the early stages. A technically impressive pilot can quickly become difficult to justify if inference costs are so high that they undermine the underlying business case.

Validation is particularly important because agentic AI is still relatively new and continues to behave like a black box in many situations. Organisations shouldn't deploy it simply because agentic AI is currently receiving significant attention. They need to understand whether the system is reliable, whether the economics work, whether the regulatory requirements can be met and whether the business outcome justifies the investment.

The expectation is that meaningful, widespread adoption will take more time. Agentic AI could ultimately become much more self-governing and commonplace, similar to technologies such as GPS or IoT that eventually became embedded into everyday systems rather than being treated as something novel.

Governance may increasingly become embedded within agents themselves, with their behaviour determined by their purpose and operating context. An agent working in a regulated environment, for example, could operate under very different constraints from one performing a low-risk task. Over time, we may also move towards “agents for things,” where agentic intelligence becomes embedded in physical and human-facing systems in much the same way that IoT sensors are embedded in devices today.

But that is not where most enterprises are today. Meaningful widespread adoption is likely still three to four years away, as regulation, governance, reliability and cost structures continue to mature.

For now, the more pragmatic strategy is to start small, stay outside the most heavily regulated core processes, avoid unnecessarily expensive use cases, measure the outcome rigorously and scale only when the evidence supports doing so.