That’s because day-to-day operations are a lot more complicated than the conditions established in the pilot environment. And, as such, promising ideas remain trapped in a familiar cycle of pilot purgatory.

For Lokesh Jindal, head of products and marketing at Axtria, a driving factor holding companies back from scaling agentic AI is that the technology does not behave in the same way that we have come to expect from traditional enterprise software.

“Ultimately, everything is based on a large language model (LLM),” he says. “And an LLM is by its very nature a probabilistic model. It is not a deterministic model.”

To the average AI user, that distinction may sound academic, but in practice, it has significant consequences for deployment. Whereas traditional enterprise software is designed to operate in a more black-and-white realm of fixed rules and identical results, LLMs work in what Jindal describes as “shades of grey.” Simply put, these models use statistical inference to estimate the most likely outcome. Consequently, the margin for error never completely disappears.

In a pilot, these limitations can be managed, but scale changes that equation.

“A pilot does not give you that exposure,” he says. “You can make it perfect for that controlled environment, but then when you try to put it into production, the environment is not that controlled anymore.”

The issue, he explains, is that many organisations still approach AI as if it should behave like traditional enterprise software. But expecting deterministic behaviour from a probabilistic system creates a gap between what the technology can realistically deliver and what businesses believe they are buying.

“We always asked for commitment of 99.9% accuracy, and suddenly we are saying 94%, 95% is okay,” he explains. “In pilots, fantastic, but the moment you try to scale, and you now get more complexity, the probability kicks in, and it's not acceptable in production.”

It is in that initial mismatch that many pilots are ultimately doomed to fail.

Rajesh Choudhary, who leads Axtria's Agentic AI Centre of Excellence, argues that this is where many organisations begin to run into trouble.

"The primary reason for data challenges is that every organisation is using analytics-ready data, and they are trying to put AI projects on top of that," he says. "But there is a fundamental disconnect, because the way you prepare data for analytics is different."

He likens that difference to the contrast between reheating a prepared meal and cooking from scratch.

Traditional analytics, he explains, resembles a microwave. The calculations have already been performed, and the metrics already exist. The system simply retrieves the answers that have been prepared in advance. Agentic AI, by contrast, is more like preparing a meal in a live kitchen. In order to deliver a high-quality result, you need access to far richer information than a conventional reporting environment was ever designed to provide.

Increasingly, he explains, that missing ingredient is context.

Connecting an AI agent to enterprise data is now fairly straightforward. Teaching it how different datasets relate to one another, how commercial teams use that information, and what business rules should guide its decision-making, remains considerably more challenging.

Choudhary describes this process as building a “context layer”. In essence, he explains, this acts as an organisational guidebook, helping an agent understand both where information resides and how it fits into the company’s bigger structural picture. Without this additional layer, even the most sophisticated model will struggle to move beyond surface-level responses.

Of course, in order to trust the output created by agentic AI, it is important that users are able to understand how an answer was produced. Enterprise users increasingly expect AI to show its reasoning, revealing which sources were used, how information was connected, and why a particular conclusion was reached.

“We are enabling what is called context transparency, which is where you got the answer, but you also are given a complete grounding summary that this is how we were able to give you an answer,” explains Choudhary.

"It's not just what's the answer," he says. "Show your scratchpad. Where did you do this calculation?"

Pharma companies have spent decades organising themselves into systems of specialist functions, each with their own priorities, budgets, and ways of operating. AI did not create these siloes, but it has exposed them.   

"In my opinion, most pilots that fail tend to be isolated experiments," explains Sameer Sardana, product management principal at Axtria. "They prove a point in the little niche that the pilot was designed for, but don't, by design, try to solve the enterprise problem."

The nature of isolated pilots means that scaling can result in a patchwork of different, competing systems that reflect the same fragmented landscape that companies are trying to avoid. The only way to counter this, Sardana argues, is to invert the pyramid: see scale as a starting point, rather than the end goal.  

Of course, this type of mindset shift doesn’t happen overnight. Users need time to understand where AI performs well, where human judgement remains essential, and ultimately how and where the two can work together.

“A large part of building that trust and building that confidence – apart from ensuring that we have done robust testing in the responses and figured out and worked out all different types of edge cases – is about managing change,” he says. “Instead of rolling out large-scale, high-impact decisions in the beginning, we start with rolling out simple decisions.”

Jindal argues that organisations often underestimate how long meaningful change takes. Technology may evolve in months, but people rarely do. Attempts to accelerate organisational change too aggressively risk creating resistance, rather than adoption. "You just have to keep at it," he says. "If you try to drive it too fast, you break things."

Seen through that lens, the challenge facing enterprise AI is no longer primarily technical. To join the likes of Axtria on the other side of pilot purgatory, organisations must be prepared to rethink the way that decisions are made, and ultimately how success is measured. For many companies, that may prove to be the longest journey of all.