Freshworks CTO: Give AI Agents an Identity, Scoped Permissions and an Audit Trail

Enterprise AI is moving from chat assistants to agents that take real actions. Freshworks CTO Murali Swaminathan talked to The Times of India in Chennai about how the company is handling that shift, and most of what he said was about governance, not models.

The "digital worker"

His starting point: safety guardrails can't live only inside the AI model. An enterprise agent needs an identity, defined permissions and a clear scope of authority, the same way a human employee does.

• Freshworks is introducing a non-human agent it calls a "digital worker", which acts on an employee's behalf.
• Each digital worker is treated as an identity with scoped, time-bound permissions. An employee going on vacation could temporarily delegate authority to theirs.
• System logs have to show clearly which actions the human took and which their digital worker took.
• Oversight should grow with impact. An agent summarising a document is not the same as one changing production infrastructure or modifying access.

Why SaaS still matters when agents do the work

Swaminathan's case for SaaS platforms is that they already have the plumbing agents need: systems of record, APIs, workflows, role-based access and audit trails. Users may increasingly work through agents rather than screens, but the agents still sit on top of that foundation.

He was also frank that AI is disrupting how software gets built. Now that everyone can be a builder, competition is sharper. He sees the upside as well: it lets Freshworks go after large enterprise customers, which brings stricter resiliency demands, urgent support and a much more hands-on sales process.

My take

I think he has the framing right. Most agent failures I see aren't model failures. They're access failures: an agent running on a service account that can do far more than the task needs, and logs that can't tell you who actually did what.

Treating an agent like an employee, with an identity, an owner, permissions that are scoped and expire, and its own audit trail, is something any enterprise can start on now, whatever vendor it uses. Before deploying any agent, I'd ask one question: if it does something wrong at 2 AM, can we tell from the logs that it was the agent, and on whose authority?

On skills, Swaminathan said the bar for entry-level engineers has gone up. They need solid computer science fundamentals to judge whether AI-generated code is correct, and a portfolio of things they've actually built.

Source: The Times of India: 'AI disrupts, but offers avenues to take on large incumbents'