Enterprise AI is moving quickly from individual experimentation to wider business adoption. But for Freshworks CTO Murali Swaminathan, the challenge is not simply turning on new models: organisations need the data, processes and controls to make AI useful at scale.
“AI is not a bolt-on,” Murali said. “AI has to be embedded within your standard business process.” Good operational processes and data, he added, are essential foundations. He recommends beginning with smaller, repeatable use cases, building confidence that they work, and then expanding to larger deployments.
The advice comes as companies confront a growing menu of AI models and tools. Adopting too many can add complexity rather than improve productivity. Instead, businesses should standardise on a manageable set of tools and apply AI deliberately – first to areas where it can assist or make recommendations, and later, where appropriate, automate parts of a process.
“Apply AI in proven tested areas,” he said.
From machine learning to agents
Freshworks’ own AI development has evolved through several phases. Murali said, “Our journey started back in 2018 when we released our AI offerings.” There were machine-learning models for tasks such as classifying service tickets.
After the arrival of generative AI and large language models, Freshworks started applying them to tasks such as summarisation. It has since moved towards agentic workflows: systems that can take multiple steps, select tools and models, and work with other services to complete a task. Murali characterised this next step as a shift “from the LLM era to an agentic era.”
Today, the company offers pre-built AI agents as well as an Agentic Studio for customers who want to create agents “end-to-end” for their own use cases. Those agents can be made available through Freshworks products and other interfaces, including Slack, Microsoft Teams, email, and web chat. The aim is to let agents work through different interfaces and integrations to carry out business tasks.
So, having good operational processes, good data is the underlying foundation for AI to work.
Murali Swaminathan
Murali described password resets as an example of why an apparently simple request can involve complex orchestration. The system may need to identify which identity provider applies, check the user’s context, verify their identity and then call the appropriate service. Some steps involve APIs and integrations rather than AI tokens; language models may be used to interpret the request or ask follow-up questions.
That distinction also shapes Freshworks’ approach to pricing and model selection.
A password reset is a workflow that requires both AI and non-AI costs, rather than one AI operation priced purely by tokens; tokens may be used to interpret requests or ask follow-up questions, while the actual reset can involve APIs, integrations and identity-provider checks. So, token usage could range “from a few tokens to many tokens.” depending on how long it takes to understand and complete the request.
The company’s plans combine seat-based licences with a set amount of tokens or sessions, and option to purchase more if the allowance is exceeded, while an LLM gateway selects models according to the task.
Connecting tools, with guardrails
Murali also touched upon how AI systems connect across enterprise software. Freshworks provides a Model Context Protocol (MCP) interface so third-party AI tools can access information and capabilities in its products. He said the company is also using Agent-to-Agent (A2A) connections, which can allow agents in separate platforms to invoke one another.
For example, a Freshworks agent created in an agentic studio gets an A2A card. It can be registered in Google Vertex Studio, for example, so an agent there can discover and invoke the Freshworks agent directly. Freshworks intends for the reverse to also happen – its own studio orchestrates agents that are hosted elsewhere.
A2A connects agents to other agents while MCP connects AI applications and agents to tools and data exposed by external systems. Ultimately, both approaches are evolving, he noted; A2A’s handling permission to carry the caller’s role-based access through to the agent needs further development, while MCP is working towards finer-grained tool selection and role-based access.
Freshworks uses an MCP gateway to manage how users access backend systems through authorised AI tools “in a secure fashion.” For companies concerned about “shadow AI”- employees using unapproved tools – Murali pointed to IT controls such as single sign-on, allowlists and blocklists.
A simpler layer – or another source of complexity?
Murali expects AI to reduce some of the complexity of enterprise technology by giving users a more intuitive way to work with systems. But achieving that, he said, depends on managing the tools and integrations behind the interface rather than simply adding more products.
Freshworks is also seeking to expand its presence in Southeast Asia. Murali said the company is looking to work with organisations including Singapore’s Economic Development Board (EDB) and AI Singapore (AISG), with the aim of building local connections and engaging with the region’s AI talent pool.
A key challenge for organisations is enabling secure data access for AI-driven automation. For Murali that starts with disciplined implementation to establish the foundations, prove the use cases, and scale from there while managing tool sprawl and secure MCP gateway access.
