What is agentic AI, really? Do CFOs really need to play around with LLMs every day? Will proprietary ERP systems become obsolete? We, like you, have a lot of questions about AI. To get some answers, we sat down with Ashok Krish, head of the AI practice at Tata Consultancy Services, in one of the global IT company’s New York offices. Below are some highlights from a wide-ranging 45-minute conversation.
Many vendors are using the terms “agentic AI” and “agents” in their product announcements. Can you give us a rough and ready way to distinguish between them?
Agentic is a way of building AI applications using large language models. Each agent, essentially, gets triggered. They complete one task, they evaluate if they’ve done the task, and they sign off and pass it on to the next agent. Agentic AI is when you also have an orchestration layer that first plans out which agent to call. It’s not performing a task; it’s achieving a goal. Agentic AI has been possible only in the last six or seven months because we needed reasoning models, we needed models that could plan.
There are AI agents and there is agentic AI. I know it sounds like a weird semantic trick. An AI agent does not think. There is no complex reasoning. There’s no memory.
“You need to experiment as widely as possible. Avoid getting locked into a single vendor. Do not give all your data to one company. Do not use just one framework from one cloud.”
For a CFO who’s involved in purchasing software or SaaS, what expertise related to gen AI do they need at this point? Or what do they need to learn about?
Everyone needs to immerse themselves and use AI in everything they do. Only then will they understand where it should be used, where it should not be used; where it works, where it doesn’t work. Because this is so fast-moving, it’s better to have meta knowledge—knowledge that some broad patterns are true—rather than just knowledge itself.
And with disruptive technology like AI, you can standardize [on systems] too soon. It’s still changing. You need to experiment as widely as possible. Avoid getting locked into a single vendor. Do not give all your data to one company. Do not use just one framework from one cloud. Every week these models will compete and get better, and it’s impossible to predict who will outdo the other.
An organization shouldn’t get locked in. But when choosing among ERP systems, for example, what AI capabilities should be on the requirements list?
If I’m buying an ERP, I want my AI stack to be heterogeneous. I don’t want my CRM app to be my only source of AI. I would like them to bring their own AI agents, which would make it easier for me to navigate and extract information from their systems.
The entire history of ERP and other systems is essentially building unusable, complicated proprietary stacks that companies are dependent on. AI has created an opportunity to disrupt that, which is why all of these companies are in a rush to roll out AI agents of their own. [As a customer,] I just need access to [the system’s] data. I don’t need [the ERP system’s] workflows. I don’t need its screens, because I can assemble that using AI on the fly. And AI will autonomously figure out that I don’t need a complicated workflow screen if I have AI agents orchestrating behind the scenes.
Therefore, I want to make a choice based on the openness of the architecture, the openness of the data, the quality of the overall data architecture. Is your AI stack interoperable with agents that [someone else] will build outside your stack? I do not want [the vendor] to prevent my bespoke agent from accessing the data in my ERP system.
Will it be challenging to connect enterprise services with AI tools?
Everyone got excited doing rapid [proofs of concept], developing a document search chatbot in under 15 days, for example. But it would take six months to scale it to production, because you have to integrate it with authentication and roles and access controls. And maybe you’ll find out that your company has in the past had four mergers and acquisitions, [that it has] eight active directories and multiple ERP systems, and each of them uses a different schema. The real world is so phenomenally messy that 80% to 90% of the work of scaling AI is not AI work. It’s just plain IT integration work. I think that will hold back many companies from truly getting value from this.
This interview first appeared in the August 8 issue of CFO Leadership’s Finance & Accounting Technology Briefing.
