Mehdi Goodarzi, Global Head of AI Business Unit at Hexaware spoke to IoT Insider about pivoting to become an “AI-first company”, using AI to disrupt the industries it serves and pushing the boundaries of the technology, in a conversation that took place at the AI Summit London.
Tracking the launch of ChatGPT as a moment that crystallised what the technology could do for people as a Generative AI (GenAI) chatbot, AI in 2022 was about “experimentation and proof of concepts”, according to Goodarzi.
“Hexaware was very much involved in helping those proof of concepts scale and bringing it into ecosystems,” Goodarzi explained. “In the last six [to] seven months, the whole conversation is around agentic AI and physical AI.”
While Physical AI describes AI interacting with the physical world — like robots or devices — it is often seen as one of the domains where Agentic AI operates. Agentic AI refers to autonomous AI agents that can make decisions and act across both digital and physical environments, including areas such as customer service, healthcare, and workflow management.
“For example, your email automatically puts emails into a spam folder. That’s the AI agent,” said Goodarzi. “More advanced use cases would be around how AI can help the underwriter to qualify and approve a loan, end-to-end. That’s the complicated [part] of the workflow.”
Model Context Protocol (MCP), an open-source framework standardising how LLMs integrate and share data, is being adopted and integrated by Hexaware as a protocol layer to support AI agents interacting with one another.
“MCP is like a USB C on your computer,” shared Goodarzi. “You can connect different drivers … you can connect different USB ports, you can connect your file systems into that.
“It’s a standard protocol that enables the large language model and agents not only to access their own data, but also external data,” he continued. “For example, you want to book a trip to Manchester … imagine you just talk to an agent and say, ‘I’m travelling with my husband to Manchester. I have two kids. Could you please book my itinerary?’ That agent knows about your travel patterns in the past, it knows what … loyalty programmes you have in your place.”
Drawing on this information from different sources and using MCP for easy data access, the AI agent can identify the best flights, accommodation and schedule trips for you – a completely autonomous, intelligent way of planning and taking over some of the manual work from the human.
This is on the end consumer side. On the enterprise side, agentic AI supported by MCP can help organisations with managing different data sources, often in silos. Taking the example of the recruitment and hiring process of a candidate, this would typically involve a manager posting a job profile, reviewing CVs, scheduling interviews off the back of that, and then interviewing and analysing candidates in terms of competence.
However, Goodarzi said, this can all be done with an AI agent, and its capabilities can even go beyond hiring someone.
“It can orchestrate the whole onboarding process in terms of email access … for procurement, for supply chain, for sales, every aspect of the business,” he said, demonstrating that the possibilities are seemingly endless.
Challenges of MCP
I asked Goodarzi about data security and privacy. Presumably this has to be all important with using agentic AI supported by MCP?
“That’s one of the limitations at the moment [with] MCP,” he acknowledged, “There are so many people contributing to that, there is a challenge [that] there are no regulations or a kind of embedded privacy framework, similar to GDPR.
“So the early adopters of MCP [still] have the chance to influence the roadmap of MCP.”
As a result, businesses should experiment with adopting MCP into one function – for instance, HR or procurement – come up with a hypothesis, validate it, and bring in a few data sources to confirm it works. Notably, they should exercise caution when it comes to what data they use.
There is an open-source community all contributing to the development of MCP. Google introduced its own agent-to-agent communication earlier this year, in April, although this is slightly different to MCP in that MCP facilitates agent to LLMs communication.
“One of the challenges that we have seen in the market [has been] many organisations … stop at the proof of concept and experimentation phase, they’ve never managed to scale it and bring it into production,” said Goodarzi.
Another challenge facing MCP was the availability of data, because LLM data is frozen at a training set.
“You need to teach the LLM what context you are caring about. That’s very important if you want to adopt it at the enterprise level and for the customers, because the whole LLM is based on probabilistic [outcomes] not deterministic,” said Goodarzi. “To reduce that probabilistic behaviour, you have to provide more context into LLMs so they can make more informed decisions.
“At the end of the day we would like to move [into an] autonomous kind of agency.”
There’s plenty of other editorial on our sister site, Electronic Specifier! Or you can always join in the conversation by visiting our LinkedIn page.