Agentic AI is a transformative technology that enables autonomous decision-making and execution within defined parameters. By combining advancements in generative AI, orchestration frameworks, and transaction-ready networks, it is revolutionizing how businesses operate. Shailesh Dhuri, CEO of Decimal Point Analytics, in a conversation with Tech Achieve Media, shares how the company is leveraging this cutting-edge innovation to streamline workflows, enhance data insights, and address complex challenges across industries. Currently, the company is pushing the boundaries of what agentic AI can achieve, thus setting a benchmark for technological integration and business value.
TAM: What technological advancements or market needs have catalyzed the advent of agentic AI?
Shailesh Dhuri: The concept revolves around three converging “rails.”
- Generative AI (LLMs): The first rail comprises large language models (LLMs), often referred to as generative AI. Over the past two years, LLMs have seen significant advancements, from ChatGPT 3.5 to O3 and DeepSeek. These advancements have notably reduced hallucinations and operational costs, making the technology more accessible and reliable.
- Affordable Orchestration Frameworks for Agents: The second rail involves cost-effective orchestration frameworks for agent-based systems. Open-source solutions like LangChain, AutoGen, and CrewAI have emerged as robust “plumbing” for building agentic AI platforms. These frameworks have simplified development and reduced infrastructure costs, enabling more widespread adoption.
- Transaction-Ready Networks: The third rail is the rise of transaction-ready networks. For instance, Visa recently launched an AI-powered payment network. This network allows users to deploy agents capable of performing transactions autonomously.
For example, imagine you want to book a flight ticket but don’t have time to monitor prices continuously. With this system, an AI agent can track prices for you. The moment the ticket price drops below your set threshold, even at 2 AM while you’re asleep, the agent can use your credit card to book the ticket on your behalf.
These innovations, advanced LLMs, streamlined agent orchestration frameworks, and transaction-ready networks, are creating a fertile environment for the rise of agentic AI, unlocking new possibilities for automation and efficiency.
TAM: As AI systems take on more autonomous roles, how can organizations ensure robust governance to prevent unintended consequences?
Shailesh Dhuri: The way to think about it is to treat every agent as both a micro-intern employee and a micro-risk. In a company, when you have interns, you assign them tasks and allow them to make some decisions, but you provide them with tight guardrails. For example, you don’t allow them to make significant decisions independently; instead, you ask them to check with you if a situation exceeds their scope. The same principle should apply to agentic AI.
We need to provide agentic AI with role-based permissions and perform continuous audits. When working with inexperienced staff, you might review their work every four or five hours, whereas with experienced staff, you may only need weekly check-ins. Similarly, agentic AI requires continuous monitoring, and you should have a readily available kill switch. This kill switch ensures that if the AI begins to misbehave, it can be stopped quickly and efficiently.
Fortunately, there are robust frameworks in place to guide this process. For instance, the U.S. NIST has developed a generative AI profile with four key parameters: go on, map, measure, and manage. Companies can use these parameters to define how the AI will operate, evaluate its performance, and manage its integration effectively. Essentially, the same processes used to oversee human employees should also apply to agentic AI.
A layered oversight approach is also crucial. For example, the EU AI Act has established obligations for high-risk and low-risk systems, including requirements for real-time action logs and post-hoc independent audits. These guidelines provide valuable insights into how to maintain control over agentic AI.
Additionally, it is vital to adopt a sandbox-before-production approach. This means rigorously testing AI systems in a controlled environment with as many edge cases as possible, using synthetic data for simulations. AI systems should operate in shadow mode until they meet both performance and ethical standards. Only then should they be gradually introduced into production. Skipping this step and deploying AI directly into production is highly risky and should be avoided.
Finally, all policies should operate under a policy-as-code model. This means converting your policies into machine-readable and enforceable code that can be implemented directly within the agent runtime, rather than relying on static documents like PDFs.
TAM: What are some of the significant hurdles businesses face in integrating agentic AI into existing workflows?
Shailesh Dhuri: What you referred to as a historical issue, I like to call “legacy spaghetti.” Most organizations have built systems and processes that were appropriate for their time. However, these systems are often complex and tangled, making them ill-suited for today’s needs.
Agentic AI requires clean APIs to interact seamlessly with core business systems, but these systems were never designed for autonomous API calls. As a result, organizations must untangle this “spaghetti” and clean up their systems, which is a daunting and time-consuming task. This is why many companies initially deploy agentic AI at the periphery of their operations rather than at the core. Cleaning up legacy systems is a significant hurdle.
Another challenge lies in data granularity and entitlements. Often, agentic AI either accesses too much information or too little, because organizations typically define data access policies in broad terms rather than with the granularity required for agentic AI. This lack of precision necessitates substantial preparatory work before AI frameworks can be effectively implemented.
Then there’s the skills gap. New roles are emerging, such as prompt engineers, product owners, and security architects. These roles take on a new dimension in an era where software is not just presenting information but also making autonomous decisions. For example, a security architect’s responsibilities change dramatically when the software transitions from supporting human decision-making to independently acting on its own. Addressing this skills gap is critical, and it’s something my clients are actively working on with help from Decimal Point.
Finally, there’s the issue of regulatory ambiguity. When new technologies like agentic AI emerge, especially in regulated industries, existing rules, such as Basel III, HIPAA, or GDPR, don’t offer clear guidance because they were created before this technology existed. Companies must engage with regulators to interpret these rules, which often delays the approval pipeline.
TAM: How do you see agentic AI working in tandem with technologies like quantum computing, blockchain, or IoT?
Shailesh Dhuri: Each of these technologies has unique use cases and applications, and most of them complement one another. Take blockchain, for example. Blockchain acts as a store of records, and it doesn’t make decisions, whereas agentic AI does. Blockchain is ideal for storing historical data securely and tamper-proof, ensuring permanence.
Many companies are now combining blockchain with agentic AI for innovative applications, such as self-custody wallets and smart contract payments. In Western markets, businesses are exploring synergies between blockchain and agentic AI platforms to unlock new possibilities.
Now, let’s look at quantum computing. Just a few days ago, IBM announced an advancement where AI agents can call quantum solvers. Traditional digital computers, which agentic AI currently operates on, struggle with certain optimization problems. IBM is addressing this by making their quantum computers accessible to AI agents, enabling faster and more cost-effective solutions for challenges like portfolio optimization and risk analysis. This helps AI agents make better decisions.
Blockchain provides a secure record for agents, quantum computing enhances decision-making, and IoT enables real-world sensing. For example, in logistics, production, or even warfare, more sensor data improves the decisions AI agents can make. With advancements like DeepSeek’s small models running on devices as compact as Raspberry Pi, integrating IoT with agentic AI even at the network edge is becoming feasible. Together, these technologies could pave the way for entirely new computing platforms in the next four to five years.
Now, regarding industries leading the way in agentic AI development, let’s explore some examples and lessons:
Consumer Finance
Companies like Visa and Klarna are leveraging agentic AI for low-complexity, high-volume tasks that require rapid validation. Examples include ticket booking or low-value loan approvals, which AI agents handle instead of humans. If something goes wrong temporarily, the losses are minimal, and corrective action can be taken quickly. The key takeaway here is to start with low-risk, high-volume processes to build confidence in the technology.
Institutional Banking
Firms like JP Morgan and Morgan Stanley have extensive in-house data reserves and strong relationships with regulators. This allows them to create sandboxes and demonstrate to regulators that agentic AI can operate safely. Their competitive edge lies in pairing AI agents with their institutional data to gain defensible advantages. For these companies, their long history and wealth of proprietary data provide a unique edge in leveraging agentic AI.
Manufacturing
In discrete manufacturing, IoT-enabled, sensor-rich environments are paving the way for innovations. Companies like Siemens and SymphonyAI are co-locating compute power at the edge. Previously, sensor devices were limited to basic on/off decisions based on a few parameters. Now, with the integration of AI models into devices like Raspberry Pi, decision-making can happen at the edge with low latency. This reduces reliance on human operators and allows for cost-effective, local computation, a game-changer for latency-critical applications.
These insights across sectors, consumer finance, institutional banking, and manufacturing, illustrate how agentic AI is being adopted. Each industry presents unique opportunities and challenges, but the lessons learned are invaluable for others venturing into this space.
TAM: How is Decimal Point Analytics leveraging agentic AI to create unique value propositions for its clients?
Shailesh Dhuri: I’d like to share some of the fascinating work we’re currently engaged in. We’re in the era of “Trump tariffs,” and many of our customers are asking, “What do these tariffs mean for my business?” For instance, if new tariffs are announced, how do they impact operations or strategies? To address this, we are developing agents capable of providing quick, reliable answers.
These agents scrape customs data, apply reasoning through economic models, and analyze network effects. While still under development, they will soon be able to generate risk scores based on the latest tariff announcements, offering clients real-time insights into how such changes might affect their businesses.
Another exciting initiative is DPA’s DNA in research. We are building an agent that ingests filings, news items, and even social media feeds from a company and its employees. This allows us to produce a quantified sentiment report and a concise three-page board brief. These reports are invaluable for clients evaluating loans or equity investments in a company. Already, we’ve seen preparation time for analysts reduced by 70%, while the depth and quality of insights have significantly improved.
The third area involves data preparation on a massive scale. We collect data on approximately 15,000 listed companies worldwide. This includes filings, executive appointments, insider trading activities, earnings announcements, and upgrades or downgrades. Over the last few months, we’ve deployed agents to optimize this process.
As of today, we are proud to say our output is the fastest and cleanest in the world. No other company is currently matching our speed and data accuracy. While competitors will likely catch up in the next three to six months, we’re ahead of the curve right now.
These are just a few examples of how we’re leveraging agentic AI and LLM-based frameworks to push boundaries and deliver innovative solutions.
