Murali Swaminathan, Chief Technology Officer of Freshworks, believes organisations fumble their AI deployments through a lack of purpose and preparation. He argues that AI itself isn’t the challenge for enterprises, but rather the complexity of the systems beneath it.
In this interview with Frontier Enterprise, Swaminathan explains why successful AI deployment depends on a clearly defined business problem and systems that allow the technology to act reliably.
What problem is agentic AI solving that existing automation can’t?
One of the key limitations of traditional automation is that it can execute only predefined rules and workflows. This works well for repetitive, structured tasks, but is less effective when a customer or employee request requires an understanding of intent and context, decisions, and actions across multiple systems. Agentic AI addresses this gap. Rather than simply following a predefined workflow, an AI agent can understand what the user is trying to accomplish, determine the appropriate next steps, use the relevant tools and systems, and adapt to what it discovers.
For example, an employee might simply input the request, “I need access to the applications required for my new role.” An AI agent can use the employee’s role and context to identify the appropriate applications and permissions, check existing access, initiate any required approvals, provision access through connected systems, and confirm completion. The employee wouldn’t have to navigate multiple forms, workflows, or teams. Similarly, an AI agent handling a customer request can assess the issue, retrieve the relevant customer and account context, diagnose the likely problem, take action across systems, and resolve the issue instead of simply categorising the ticket and routing it to another queue.
The main problem agentic AI addresses, therefore, is the shift from automating predefined steps to independently achieving the requested outcome. This involves understanding intent, reasoning through context, and taking action across systems to address customer and employee needs from beginning to end.
What tends to break when AI is layered onto existing enterprise systems?
The problem is often not AI itself, but the complexity of the underlying systems. Many organisations still operate fragmented legacy systems, with data spread across disconnected platforms. This makes it difficult for AI agents to access the context required to produce accurate and consistent results.
Freshworks’ 2026 “Global Cost of Complexity Report” links technology complexity to lost productivity and delayed returns, citing disconnected applications, fragmented data sources, and overcomplicated workflows. The research, which included Singapore among six markets, found that 27% of mid-market IT decision-makers identified system integration complexity as one of the leading reasons AI pilots fail to become full programs.
Adding AI to fragmented processes and siloed data can compound the existing complexity, leading to inconsistent results, lower trust, and limited ROI. A capable model alone isn’t enough. AI also needs an integrated, adaptable platform and an accessible data layer that provide the context and ability to act reliably across an organisation. AI can’t fully compensate for a fragmented foundation; the resulting experience depends on the platforms, data, and systems supporting it.
How should enterprises rethink software architecture in an AI-first environment?
As organisations increase their use of AI, they need to rethink architecture so AI doesn’t become another layer of complexity on an already fragmented technology stack. This doesn’t require them to replace the entire stack. Instead, the architecture should allow AI agents to work across existing systems through interoperability, accessible data, well-defined APIs, and governance built into the foundation.
Architecture has direct business consequences because integration challenges, governance overhead, data management issues, and rework can consume resources before AI produces meaningful value. In practical terms, enterprises need an accessible data layer that gives AI the necessary context, pre-built connectors and APIs that allow agents to act across systems, and governance and permissions that define what an agent can access and do.
This means building a platform on which AI can understand, reason, and act without requiring organisations to rebuild their technology environments. The goal shouldn’t be to create a separate AI stack alongside the existing enterprise stack, but to use AI to manage activity across it. Architecture therefore needs to move beyond connecting applications and allow AI to act across them.
How does AI change the way enterprise software should be evaluated and bought?
AI is changing what organisations expect enterprise software to do. Historically, organisations bought software to help people perform tasks more efficiently. AI can increasingly complete those tasks, or entire workflows, on their behalf. Buyers therefore need to look beyond feature lists and seat counts and ask: What work can this platform complete for me? How quickly can I realise that value?
This shifts the focus from automating steps within a workflow to completing the intended task. For example, the value of a service management platform no longer depends solely on how efficiently it routes an employee’s request to the appropriate team. Buyers must also consider whether AI can understand the request, resolve it using relevant organisational information, take the necessary actions across systems, and complete the work without a human handoff.
Enterprises should therefore place greater emphasis on outcomes, integration, ease of use, governance, and time-to-value. AI should operate within the platform and its workflows rather than as a separate layer for users to manage. Enterprise software will increasingly be assessed not by how much of it people use, but by how much work it can complete for them safely and reliably.
Where do AI projects succeed or fail after deployment, and what makes the difference?
The model itself is rarely the reason an AI project fails. Success or failure depends on what happens after deployment: whether the AI addresses a meaningful business problem, operates within everyday workflows, and produces measurable results when expanded. Successful initiatives begin with clearly defined use cases and business outcomes rather than with the technology. They incorporate AI into the systems employees and customers already use and provide access to the data, knowledge, and actions required to complete work from start to finish.
Data quality is particularly important: Duplicate records, outdated information, fragmented knowledge bases, and inaccessible data can reduce the quality of AI’s results. This becomes more consequential as AI grows more agentic because the system may make decisions and take action based on that information, rather than simply generate an answer.
Technology alone isn’t enough. AI also requires strong governance, workforce readiness, and a clear implementation plan. Employees need to understand where AI can help, when to trust it, and where human judgement remains important.
In customer and employee service, AI should operate within existing workflows and be measured against outcomes such as faster resolution, greater self-service, improved productivity, and better customer and employee experiences. Deployment alone isn’t a measure of success. What matters is whether the AI is used, trusted, and produces measurable business results.
