From hype to hard truths - UiPath's Ed Challis on what's holding agentic AI back

More than 90% of enterprises say they've deployed generative AI. But over 80% report it's made no measurable difference to revenue or margin.

That contradiction isn't just a growing pain – it's a pretty glaring sign about enterprise AI adoption today.

For all the enthusiasm around large language models (LLMs) and agentic automation, most organizations remain trapped in what Dr Ed Challis, Head of AI Strategy at UiPath, calls the "unrealized zone" – a gap between impressive technical capabilities and actual business impact where promising pilots fail to scale into measurable returns.

Speaking at UiPath on Tour London, Challis laid out both the remarkable progress in AI capabilities and the persistent barriers preventing enterprise value creation.

The exponential intelligence growth

Challis began by establishing the unprecedented scale of AI progress. "We've achieved something really remarkable, which is a 400% annual growth rate in the size and the complexity of our AI model," he explains, noting this represents a dramatic acceleration beyond the 40% annual growth that followed Moore's Law for the first 40 years of computing.

The numbers are staggering, with Challis noting that the largest models now have 2 trillion parameters, with training requirements so intensive that using the best supercomputer from the year 2000 to train today's models would take around 36 billion years.

A critical breakthrough has been synthetic training data generation, as Challis explains:

AI models that can create their own training data work by giving models reasoning challenges, like math challenges, coding challenges, and they can take many attempts to solve these problems, hundreds of attempts, and then we can keep only those solutions where it gets it right.

The performance gains are measurable. Current models now outperform humans on GPQA (graduate-level Google Proof Q&A benchmark) challenges, leading to what professors describe as intelligence that "feels like a very good graduate student."

Goldman Sachs economists estimate this technology could add seven percent global GDP over a decade. Yet the deployment reality tells a different story.

Agents vs. traditional automation

To understand the deployment gap, Challis distinguished between traditional rules-based software and agentic systems. He explains:

We have had for decades now, traditional rules-based software... defined by a sequence of conditional logical statements. And now we have this new type of software which is goal-based.

An agent, he defined, requires three components. First, "the large language model at the heart of an agent can do the reasoning and deduction." Second, "tools - because without tools, the agent is nothing more than a kind of next word chat interface." Tools allow the LLM to interact with other systems, access data and retrieve new information. Finally, a goal or an objective that the LLM is seeking to solve.

Despite these capabilities, UiPath's survey of over 250 executives revealed consistent implementation challenges. Security concerns topped the list, but extend beyond basic data hosting questions. Challis noted: 

Security concerns covered a much broader range of issues around access control, monitoring, logging, really understanding those reasoning chains about how certain decisions are made.

Development complexity ranked second. Challis continued:

If this technology can only be developed by extremely expensive, extremely rare AI engineers that are aware of all of the different state of the art packages. This just isn't going to be scalable.

Integration challenges proved equally persistent. These findings align with third-party research highlighting the need for autonomy, integration, orchestration, scalability, and reliability in enterprise deployments.

Beyond the presentation - the shallow deployment problem

Following his presentation, I sat down with Challis to explore these challenges in greater detail. He expanded on why current implementations fail to deliver measurable value, even when individual productivity improvements are evident.

"I don't think what businesses have done is bad," he emphasized, noting his personal ChatGPT usage:

It's just that it's very, very hard to quantify the value that ChatGPT delivers.

Even rigorous studies showing 20% faster development speeds don't necessarily translate to business outcomes: "Does that actually mean that like features get out the door 20% faster?" The challenge extends to daily productivity, as it may help to write a better email but that doesn't necessarily equal more productivity.

Even where rigorous studies demonstrate benefits—coding copilots showing 20% faster development speeds—translating those gains into business outcomes proves challenging. Individual productivity improvements don't automatically translate into faster feature delivery or increased revenue. The implication challenges current deployment strategies. Chat-based tools may serve as useful introductions to AI capabilities, but they aren't the ultimate goal for enterprise value creation. This requires process-aware, role-specific agents that automate complete flows rather than providing assistive nudges which means transformation at scale, as Challis notes:

For real values, we realized we need a whole workflow, a whole process, the whole business critical thing to be deeply transformed. And that requires integration, the people integrated with the right systems, or the right contextual data, and maybe a redesign of how we do that work.

UiPath has announced forward deployed engineers as one response to the implementation gap, which Challis explains are:

people that are deeply technically capable but highly embedded with the client and their problems that will ultimately help us deliver higher value solutions to the customer, but also provide much higher quality feedback to our product engineering teams.

Challis explained how agents could improve through operational experience, using an employee analogy that holds true from the last time we spoke: 

You build an agent for a specific task and put it into production. It gets 60% of them right and 40% of them wrong, but for those 60% it got right, you confirm it by giving it a positive signal, so those can go back into its memory.

When asked whether agents will eventually compose workflows rather than just perform them, Challis sees a hybrid evolution: 

I think businesses want certainty. Businesses want design authority, particularly given that human experts... have known about the claims adjudication process they've worked on for 50 years, and an AI model just doesn't have that contextual knowledge.

However, AI's role in continuous improvement appears more promising. Challis sees orchestration and process mining coming closer together, with AI generating hypotheses about process improvements rather than requiring humans to make the imaginative deductive leaps about what could be changed.

UiPath is exploring self-healing capabilities, which he explains as "an agent that looks at the error logs, thinks about what could be done to fix it, and either comes back with a proposal or actually fixes it, depending on how significant that change is." This combination of observability and adaptability suggests a future where agents become more resilient through operational experience rather than requiring constant human intervention.

The integration challenge that Challis highlighted in his presentation remains central. UiPath's approach involves integrating new buyers with a strategy that he says is "ultimately agnostic. So if you have an agent from Salesforce or an agent from some other company, and you want to orchestrate that. We want to be the best platform for helping to achieve that."

While UiPath has invested in training programs access alone proves insufficient. Scaling agentic capabilities requires addressing human as well as technical challenges. UiPath has invested in free training programs (I had an excellent conversation at the UiPath Community booth on the expo floor), access to training alone proves insufficient.

Cultural reinforcement through reward structures becomes essential for widespread adoption. Organizations must signal through compensation, promotion, and recognition that developing AI skills represents a strategic priority rather than optional professional development.As Challis emphasized:

It's one thing to say the tools there, the Learning Academy is there, but it's another thing to say we value you by actually prioritizing that workforce to learn. That's a cultural challenge.

Looking forward, Challis believes organizations underestimate both the technology and the speed of change. He predicts:

This technology will cause huge disruption to incumbents, and I think you'll see very large significant businesses diminish if they don't adopt the technology, just like we saw with the Internet."

My take

The conversation between presentation insights and interview details reveals both the remarkable potential and persistent challenges of agentic AI. While model capabilities have reached new levels of performance and synthetic training approaches promise continuous improvement, the gap between technical sophistication and business impact remains substantial.

The limiting factors center on integration complexity, cultural adaptation, and the challenge of moving from shallow, isolated deployments to deep, process-transforming implementations. Success requires addressing security frameworks, development scalability, legacy system integration, and organizational change management.

The intelligence exists. The question now is whether enterprises can orchestrate the surrounding elements needed to unlock its transformative potential. The willingness of organizations to embrace change, and the authenticity of vendors to actively work with their customers through those challenges will determine whether the 90%/80% contradiction resolves into measurable business transformation or remains a cautionary tale about the gap between technological capability and operational reality.