AI agents are becoming harder for enterprises to trust. The problem is not just whether an agent can complete a task. It is whether it can access the right data, use it in the right context and take the right action.
For Roshmik Saha, CTO of data privacy company Skyflow, data security has become a fundamental barrier to deploying AI agents and agentic workflows in production.
The challenge has also moved beyond protecting data stored in databases. Enterprises now need to control sensitive information as it moves across applications, models, search systems and agents.
“Initially, enterprises wanted to protect their user and customer data, which was in databases, like Postgres, MongoDB, Databricks and Snowflake,” Saha tells AIM. “With the AI journey, the amount of data that is getting processed or transferred is huge.”
Founded in 2019, Skyflow initially focused on protecting structured enterprise data. It has since expanded to include unstructured data, such as files and documents, as these are increasingly processed by AI systems. The company now describes its platform as a runtime data control layer for AI.
The shift is important because sensitive information can spread across multiple systems once it enters an AI workflow. Traditional controls around databases may not be enough to govern where that data goes or who can access it.
“You can create some guardrails, but they are probabilistic in nature,” Saha says.
That challenge helped pull Skyflow deeper into AI. One of its early customers, healthcare AI company Hippocratic AI, wanted to build a digital nurse while meeting HIPAA requirements.
Skyflow’s approach is to anonymise or tokenise sensitive information before it reaches AI systems. The original data can then be reidentified or retrieved when an authorised user or application needs it.
The company has also expanded into enterprise AI search. Skyflow recently partnered with Glean to help organisations use enterprise data for search and AI applications without exposing sensitive information unnecessarily. Skyflow says its system can sanitise documents before they enter indexes and enforce access controls when results are retrieved.
The Tech Stack
A key part of Skyflow's technology is its ability to identify sensitive information in unstructured data.
Saha says the company uses smaller language models and entity recognition to detect sensitive information. The harder problem is protecting that information without breaking the AI application that needs to process it.
Skyflow replaces sensitive values with tokens or other protected representations. This allows AI systems to work with the data without directly exposing the underlying information.
Its API-based architecture can be integrated at different points in an enterprise data flow. It can sit between an application and a database, work with an AI gateway or integrate with platforms such as Databricks and Snowflake.
Skyflow's documentation says its data privacy vault supports both structured and unstructured data. Its technology stack includes tokenisation, data de-identification, masking and polymorphic encryption.
The system can also be configured around an organisation's definition of sensitive data. Skyflow provides predefined data types, while customers can create their own policies based on their industry and use case.
That could include personally identifiable information, payment data, health information or commercially sensitive data.
Skyflow's tokenisation system supports different formats, including deterministic and format-preserving tokens. This can allow applications to continue using protected data without exposing the original values.
Saha says the technology is less about securing a single database and more about controlling how sensitive information moves through an enterprise.
That distinction becomes increasingly important as AI applications connect more systems and data sources.
The Customers
Skyflow's customer base spans healthcare, fintech, retail and other industries. Its website lists companies such as Flipkart, Visa, and ServiceNow among its customers or partners.
The company has also built a presence in India. Flipkart is among its customers, while Saha says the wider Flipkart group, including Myntra and Cleartrip, has expanded its relationship with Skyflow.
The company also works with financial services organisations in India, including investment banking and asset management firms.
The US remains Skyflow's larger market, which Saha attributes partly to the faster adoption of AI use cases there. He says demand for AI security is also beginning to accelerate in India.
Visa is another example. Skyflow says its technology is used to protect sensitive payment data, while ServiceNow is an OEM partner. The broader use case involves applying AI to workflows such as dispute management without exposing sensitive cardholder information.
When Agents Go Rogue
For enterprises deploying AI agents, Saha sees two main risks.
The first is an agent taking the wrong action.
Consider a DevOps agent with permission to restart servers or fix infrastructure issues. The task itself may be legitimate. But an agent could still perform the action on the wrong system.
This is different from a conventional software failure. Agents operate probabilistically and can make decisions based on changing context.
“These are all probabilistic systems. It will make a mistake sometimes. There is no ML, there is no AI which is correct all the time,” Saha says.
The second risk is data exposure.
An agent may retrieve information that a user should not see, pass sensitive information to another system or expose data through its output. As agents gain access to more enterprise systems, controlling those permissions becomes increasingly important.
This is where Skyflow is positioning its technology. The company says its platform can enforce access policies at runtime and control how data is exposed to agents, models and other applications.
Human oversight, Saha argues, will therefore remain important as enterprises deploy increasingly autonomous systems.
Looking Ahead
AI is now Skyflow's primary growth area. The company has expanded from protecting data in traditional applications to securing data across AI models, search systems and agentic workflows.
Skyflow has raised $100 million in equity funding to date. Its most recent funding round was a $30 million Series B extension announced in March 2024, bringing its total equity capital to $100 million.
Saha says the company's immediate focus is expanding its AI capabilities and integrations.
That includes supporting AI companies that need to protect data used by their models, enterprises that want to connect internal information to external AI search systems, and organisations building agentic workflows.
The underlying problem is likely to become more complex as agents gain access to more systems.
For enterprises, the question is no longer simply whether an AI agent can perform a task. It is whether the organisation can control the data the agent sees, the systems it can access and the actions it is allowed to take.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.