In the Information Technology (IT) and Business Process Outsourcing (BPO) services domain, if there was one buzzword in 2024 it would be generative AI (gen AI) . In 2025 another buzzword has taken over, agentic artificial intelligence (AI).
Agentic AI is basically a set of systems that are designed to have autonomy, decision-making abilities and intelligence to execute workflows based on various inputs and achieve defined objectives for a user or an enterprise. From handling support tickets to drafting reports, the promise is that agents will do it all.
Enterprises’ interest in adopting such systems is rapidly increasing, and that extends to usage of such systems in IT and BPO services deals too. The underlying premise is that agentic AI will not just drive productivity and cost savings but be instrumental in realizing more strategic outcomes.
For example, if integrated with an enterprise CRM and email platforms, it can detect that there has been no activity in recent days on a potential sales opportunity, triggering a tailored incentive and sending that autonomously to the potential client (see this blog for more details).
Reach out to discuss this topic in depth.
Over 2023 and 2024, enterprises learnt the importance of doing pilots quickly in the context of generative AI, and we are now seeing that happen for agentic AI too. The idea is to test the feasibility, outcomes and overall Return on Investment (ROI) from agentic AI.
However, while the excitement is justified, many enterprises are now walking into these deployments with a dangerously narrow view of the cost involved, thereby giving a misleading picture of the potential ROI.
Most business cases focus on the obvious costs:
- Licensing Large Language Models (LLM) and building integrations
- Provisioning for expected cloud compute workloads
- Data quality improvement
- Hiring engineers
These are visible, measurable, and easy to scope. What is missed though are the “hidden costs” that are lurking beneath the surface and get understood post-deployment, in the form of budget overruns, friction, and unmet expectations.
The cost of change is human
Agentic AI does not eliminate the need for humans. Realizing an effective agentic AI enabled workflow requires strong human involvement at various levels. For instance, process architects will need to design new workflows. Relatively newer roles such as AI ethicists and prompt engineers might also need to be integrated into the ecosystem.
Humans in the loop will need to be trained to work effectively with the agents (how to provide inputs, validate outputs/transactions, handle exceptions, etc.). Costs for such phases need to be accounted for.
An even bigger cost component that is overlooked is the cost of organizational change management. Agentic AI redefines how people interact with technology in their organization. That shift doesn’t happen automatically and often comes with the risk of people feeling insecure about their own roles. Keeping that in mind, the need for effective organizational change management is even more pronounced as compared to other transformational initiatives such Enterprise Resource Planning (ERP) implementation.
Errors do happen and have a cost too
Another overlooked area is the cost of error. Agents can hallucinate or misinterpret ambiguous tasks. Despite training and quality assurance processes, some errors do occur in a production environment. The consequences can be serious, sending incorrect emails, escalating issues unnecessarily, or generating inaccurate reports. It is important to factor for some error or failure rate in production while analyzing the positive outcomes from agentic AI.
Why hidden costs matter
When a comprehensive assessment of these hidden costs is not done, the result is a misleading ROI projection and unplanned costs down the line. To avoid surprises, enterprises should approach agentic AI like any major transformation. That means budgeting for:
- Process redesign
- Training and change adoption
- Post-deployment monitoring, debugging, auditing, and improvement
- Risk management
- Buffering for liability
- Provisioning for unexpected cloud compute workloads
The last point can be the trickiest, as cloud providers often subsidize AI workloads initially with credits and incentives. Over time, as those credits expire, the increase in costs could be exponential. Organizations should factor in the true cost of cloud compute (Graphics Processing Unit (GPU) usage, API calls, etc.) while determining the ROI.
Final Thought
Agentic AI can be a game-changer, but only if organizations look beyond the code and compute. True success lies in preparing the business, not just the model. The costs may be hidden, but the impact of ignoring them won’t be!
If you found this blog interesting, check out our Agentic AI: True Autonomy Or Task-based Hyperautomation? | Blog – Everest Group, which delves deeper into another topic regarding Agentic AI.
If you have any questions or want to discuss the evolution of Agentic AI in more depth, please contact Rahul Gehani ([email protected]) and Abhishek Sharma ([email protected]).
