By Vijay Gopalakrishnan, Partner, Deloitte India
Agentic AI has been positively influencing businesses across various industries, including finance, procure-to-pay, sales and marketing, HR and IT. Agentic AI deployments have progressed beyond Proof of Concepts (PoC) to a most viable product status, and many are moving to full-scale production, with agent-driven automation of end-to-end processes for a business function. This is achieved through proven technological and functional methodologies, featuring best practices in use case selection, design and implementation, as well as agentic AIOps and user adoption.
Feasible agentic AI use cases
The chosen use cases should be important to the firm’s overall strategy and leadership. It is also crucial to conduct a prompt technology feasibility study for the chosen use cases. The study should focus on aspects such as the technology stack and data to confirm that the agentic AI use cases can be delivered in line with leaders’ expectations. Proven agentic AI frameworks and tech stacks are available for companies to choose from, based on their specific requirements.
Agentic AI solutions must be delivered in a phased manner, with visible outcomes to the business teams and leaders. This will help in securing the right investment and sponsorship for agentic AI initiatives. To accelerate deployment, the delivery process should incorporate proven design practices, such as reusing artefacts across phases. Additionally, efficient development techniques, such as efficient API calls and AIOps techniques, should be employed to ensure a cost-effective and smooth transition into production.
For instance, an FMCG company is using these techniques to implement a Multi-Agent System that automates the entire procure-to-pay process, covering vendor onboarding, three-way PO matching and payments.
Ethical, secure and trustworthy use of agentic AI
Despite agentic AI’s growing prominence, other AI technologies, such as Robotic Process Automation (RPA), GenAI and ML, are widely used based on specific problem statements. For instance, RPA can automate pre-programmed tasks without human-level reasoning, while agentic AI can perform human-level reasoning and decision-making and execute tasks with specific goals. They can work effectively together under the supervision of agents to achieve a team goal.
A Multi-Agent System automates a business process efficiently, which currently involves numerous human interactions within and outside the firm. It involves multiple decisions and tasks executed across various IT systems (internal and external), coordinated and managed by an orchestration layer of human management. For example, in an insurance firm, a Multi-Agent System can automate the outpatient claims processing, which typically involves multiple supervised human interactions and coordination with IT systems.
The agents and models require training with the appropriate techniques and data to ensure that the outcomes are not biased. Proven data privacy and cybersecurity techniques, such as data anonymisation, are to be used to ensure that the solutions are secure.
Agentic AI frameworks with Human-in-the-Loop (HITL) capabilities ensure efficient review and approval of agent outputs at critical checkpoints within the automated business workflow. Additionally, techniques such as Retrieval-Augmented Generation (RAG) ensure that the underlying LLMs produce trustworthy, non-hallucinated responses.
Effective user adoption of agentic AI solutions
Users of agentic AI should help design and develop the solution to ensure it meets their needs and is likely to be accepted. Additionally, companies should demonstrate to users how the agentic AI solution will make their work easier and open up new avenues of interesting work for them, rather than replacing them. This will help users become more supportive of the deployment and adoption of agentic AI.
A case in point would be how the drug researchers at a pharmaceutical firm were able to focus on their core research work, unlike earlier, when they had to spend time and effort manually writing reports on drugs’ compliance with FDA regulations. This is because a Multi-Agent System with a supervisor agent was built to automate the regulatory compliance report writing process, with agents responsible for writing text, drawing diagrams and scanning the intranet, among other tasks. This significantly improved productivity and encouraged greater adoption of agentic AI among researchers.
Agentic AI delivers value when supported by the right technology ecosystem, strong ethical compliance and well-defined use cases. When these elements align, organisations can achieve long-term success and measurable ROI.