Zycus founder and CEO Aatish Dedhia explains how his company built for the agentic AI era long before it became mainstream, and why a deeply embedded, platform-first approach is now reshaping what autonomous procurement really looks like…

There is a familiar narrative that has shaped the enterprise software industry for decades. A company builds something valuable, scales successfully and gradually becomes comfortable. Then a faster, more agile competitor arrives with a new architecture and a more compelling proposition. The incumbent struggles to adapt, weighed down by its own legacy, and the market moves on. It is a cycle that has repeated itself time and again. Zycus does not follow that pattern. The disruption did not come from a new entrant. It came from within.

When artificial intelligence became the defining conversation across enterprise technology in 2023, many procurement vendors reacted quickly, but superficially. Copilots were added, existing tools were reframed and point solutions were positioned as transformation. The urgency was clear, but so too was the lack of depth behind many of those moves. Zycus, by contrast, did not need to reposition. It had already been building toward this moment for years, embedding AI into procurement long before the current wave.

At the centre of that strategy is founder and CEO Aatish Dedhia, whose engineering background continues to shape how the company approaches innovation. His view is that transformation cannot be layered onto legacy systems. It has to be architected from the ground up, with each technological shift building on a strong and connected foundation. “Because we’ve been building AI into procurement since long before it was fashionable and we have the scars, the patents, the production deployments and successful customers to prove it,” he tells us.

That long-term commitment predates the generative AI boom by several years. As he explains, the company’s journey has unfolded in distinct, but connected phases, each building on the last rather than replacing it. “Our AI journey did not start with ChatGPT. We were embedding AI and machine learning into spend classification, contract analytics, AP automation and risk scoring years before generative AI existed. Merlin AI launched in 2018.”

Rapid evolution

This early investment created the conditions for rapid evolution when new technologies emerged. Rather than starting from scratch, Zycus was able to extend an already mature platform into new areas of capability. “When generative AI arrived, we moved to cognitive procurement – using LLMs to understand unstructured data at scale. And when it became clear that copilots and point agents were not going to transform anything, we pivoted hard to agentic AI,” he explains.

The result is a layered progression that places the company several steps ahead of newer entrants. As Dedhia sees it, many of the startups now positioning themselves as AI-native are still grappling with foundational challenges that Zycus addressed years ago. “That’s three distinct waves of AI capability – ML, generative, agentic – each built on top of the previous one,” he tells us. “A startup entering today is building wave one. We are on wave three, with trillions of dollars of spend data analysed and 32 patents behind us.”

This depth, he argues, is what enables genuine transformation. Simply adding AI capabilities onto existing workflows is not enough to fundamentally change how procurement operates. “You cannot take new technology and fit it onto the way you were doing things in the past just by adding apps,” he says. “The transformational effect is going to come through agentic AI.”

However, even that is only part of the equation. The effectiveness of AI depends heavily on the environment in which it operates, and that is where architecture becomes critical. “Here is the critical point: AI alone is not enough. You also need the transactional backbone – the workflows, the policies, the audit trails, the supplier data, the contract repository – to give that intelligence something to act on.”

Without that backbone, the promise of autonomy quickly breaks down. Many organisations, and indeed many vendors, still operate across fragmented systems that limit what AI can realistically achieve. “A startup can build a beautiful intake layer. But what happens after intake?” he posits. “Where does the request go? Who enforces the policy? Where is the contract? Where is the three-way match?”

It is this fragmentation that prevents true autonomy, turning what should be seamless workflows into complex integration challenges. “If those things live in different systems, you’re not autonomous; you are integrated, which is a very different and much harder problem,” he warns.

A single source

Zycus’ answer has been to build everything on a single, unified codebase, allowing AI to operate directly within the system rather than across disconnected tools. “We built everything organically, on a single codebase. Our AI does not talk to our Source-to-Pay system through middleware,” Dedhia explains. “It is the Source-to-Pay system that gives us access to ten times the data that an external bolt-on tool would have. Data is the fuel for AI. More fuel, better outcomes.”

This philosophy of building deep, rather than bolting on, extends into how the company has evolved over time. Reinvention is not a response to market pressure, but more a deliberate and recurring strategy. “Completely deliberate. If you look at our history, we reinvent roughly every three to five years. We started with spend analytics and sourcing optimisation. Then we built the full Source-to-Pay suite. Then we built in Merlin AI for AI & machine learning-driven capabilities across contracts, AP, spend, and risk. Then cognitive procurement with generative AI. And now, the Merlin Agentic Platform.”

Strong foundations

Each phase strengthens the foundation rather than replacing it, creating a compounding effect that becomes increasingly difficult for competitors to replicate. “Each cycle does not replace the previous one. It builds on it. The Source-to-Pay suite is still the backbone,” he says, adding that what changes is the layer of intelligence and user experience built on top.

This compounding advantage is particularly significant in the context of AI, where access to data and system-wide integration directly influences performance. “That’s actually the advantage of being an established player with a proven foundation. A startup has to build the foundation and the AI simultaneously. We already have the foundation. We can focus entirely on what is next,” he explains.

To put the current moment into perspective, Dedhia draws a comparison that goes beyond software cycles and into industrial history. “I’ve been in technology long enough to have seen the personal computer revolution, the internet, mobile, cloud. All of them were significant. But the transformation that AI will drive is on a completely different scale. You have to go back a century, to the industrial revolution, to find something comparable.”

The analogy is not just about scale, but about the nature of the transformation itself. “What the industrial revolution did for muscles, AI is doing for brains. It’s already reshaping industries, economies; the way work is organised.” Within that framework, large language models represent only the starting point.

“The LLM is today’s equivalent,” he explains, before extending the comparison further. “But the steam engine alone did not transform manufacturing. The assembly line did.” In this context, agentic AI becomes the mechanism that turns raw capability into real-world outcomes. “Agentic AI is the assembly line. It harnesses the raw power of LLMs to deliver outcomes. The LLM is 99%. Agentic AI is the 1% on top that organises that power into workflows that actually achieve something.”

Income to outcomes

This current emphasis on outcomes has led Zycus to rethink how it builds AI tools. Early experimentation with a wide range of smaller agents revealed a clear limitation.“Because agent sprawl doesn’t deliver value. We were guilty of it too. At one point we had 50 to 60 small agents,” he says.

While individually useful, these tools failed to deliver meaningful transformation at scale. So, Zycus asked itself a fundamentally different question. Instead of ‘what can AI do?’ they started asking ‘what work can AI replace?’ That shift in thinking reframed the company’s entire approach, focusing on replacing substantial portions of roles rather than automating isolated tasks. “Not a task. A significant chunk of a role; 30, 40, 50% of what a person actually does.”

From that point on, development was guided by a clear principle. “If an agentic flow does not replace a meaningful portion of someone’s workload, it isn’t worth building as a blockbuster flow,” Dedhia explains.

This thinking is reflected in ANA, the Autonomous Negotiation Agent, which moves far beyond suggestion into execution. “ANA is the clearest example of what we mean by a blockbuster flow. It doesn’t summarise a negotiation or suggest a strategy. It executes the negotiation autonomously,” he tells us.

The system handles sourcing events end to end, from supplier discovery to negotiation and final recommendation, with human involvement limited to approval. It is a clear example of how agentic AI shifts the role of technology from support to execution.

Underpinning all of this is the Merlin Agentic Platform, a significant architectural investment that enables this level of capability to be developed and deployed at speed. “This is perhaps the most important investment we have made,” he says.

The full stack

Rather than building isolated solutions, Zycus focused on creating the full stack required for agentic AI to function effectively. “Over the past 18 to 24 months, we have been building the entire architectural stack for agentic AI. Not a single agent – but the platform that agents run on.”

That decision is now paying off. “Now, the platform is proven. And the implication is enormous: every subsequent agentic flow can be built dramatically faster because the stack is already validated,” he enthuses.

Internally, the company has applied the same principles to its own operations, using AI to drive measurable productivity gains. “We believe you can’t sell AI credibly if you are not using it aggressively internally,” he tells us.

The results are clear

The results have been significant, particularly within engineering, where productivity has increased dramatically even as team size has reduced. Looking ahead, Dedhia expects the very structure of procurement software to change. Traditional modules will remain, but largely as invisible infrastructure supporting a new, conversational interface driven by agents.

“Everything we have today in terms of traditional software will slowly move to the back,” he says. What users interact with will instead be a set of agentic flows that operate dynamically across the system. The software modules become invisible. “The agents become the interface,” he explains.

For procurement leaders, the message is clear. The shift to AI is not simply about adopting new tools, but about rethinking how work is structured and executed. “First, experiment – but experiment on deep use cases, not point tasks,” Dedhia advises.

At the same time, according to the Zycus CEO, success depends on organisational commitment and a clear understanding of the underlying technology landscape. “AI will not fix a fragmented technology landscape. It will simply amplify the fragmentation,” he warns.

Ultimately, the companies that succeed in this next phase will not be those that move fastest to adopt AI at the surface level, but those that have built the foundations to support it at scale.

Check out the full edition of CPOstrategy here.