With deep expertise in machine learning and natural language processing, technologist Laurent Giraid is at the forefront of the enterprise shift from rigid, rules-based automation to intelligent, AI-driven workflows. He joins us to explore this complex transition, discussing how AI agents navigate the ambiguities of modern business processes, the operational changes required to embrace outcome-oriented automation, and the critical need for building trust in non-deterministic systems. Laurent also shares insights on breaking down organizational silos between AI and automation teams and offers a glimpse into the practical lessons learned from deploying these advanced agents at scale.
Modern workflows often involve unstructured data where outcomes can shift based on real-time context. How do AI agents handle this non-deterministic environment compared to rule-based RPA, and what are the first practical steps for a company to begin introducing this new capability?
That’s the fundamental challenge we’re facing. Traditional RPA was built for a predictable world, one with clear rules and structured inputs. But today’s business environment is messy and non-deterministic; inputs from various sources can vary, outcomes can shift, and decisions depend on context in real-time. An AI agent, powered by something like a large language model, doesn’t just extract data points; it reasons. For example, instead of just pulling 30 data fields from a credit agreement, it can find 30 or 40 answers to complex questions, understanding the nuance and context. A practical first step is to identify a process bottlenecked by this kind of unstructured data. Don’t try to boil the ocean. Start by embedding an AI agent into an existing workflow to handle one specific, complex task, allowing your team to get comfortable with the technology at a manageable pace.
You describe a major shift from giving bots step-by-step instructions to giving AI agents a desired outcome, such as “onboard this customer.” Could you walk us through the technical and operational changes a team must make to support this outcome-oriented approach effectively?
It’s a complete paradigm shift, and honestly, a faintly terrifying one for organizations accustomed to the clear guardrails of RPA. We’re moving from saying, ‘Follow step one, two, three, four, five,’ to simply stating, ‘I want this loan reviewed.’ Operationally, this requires a massive change in mindset for your Center of Excellence. They must move from process designers to outcome managers. Technically, it means building new frameworks for oversight and orchestration. You can’t just ‘test’ an outcome like you test a script. You need robust systems to monitor the agent’s reasoning, ensure auditability, and provide a human-in-the-loop for exceptions, especially in the early stages. This isn’t just a technical upgrade; it’s a cultural evolution toward trusting an autonomous system to achieve a goal.
Given known challenges with LLMs like hallucinations and regulatory concerns, what specific guardrails or auditability features are you building to foster trust? How can organizations manage risks while still experimenting with these powerful, non-deterministic tools and realizing their benefits?
This is the single biggest barrier to mass adoption, and for good reason. We know LLMs are prone to hallucinations, we know they drift, and if you change the underlying model, the responses can change unpredictably. Building trust is paramount. The key is to create a controlled environment where the agent operates. This means robust auditability trails that don’t just log actions but capture the reasoning behind decisions. It also involves security and stability checks to ensure consistent performance. For organizations, the path forward is a journey, not a leap. Start with low-risk, high-impact tasks. Let the agents handle complex data analysis or draft responses, but keep a human to approve the final action. This allows you to realize the benefits while the market matures and we all learn how to manage these powerful, but imperfect, tools.
Many companies have established separate AI and process automation teams, sometimes preventing collaboration. What specific friction does this create, and what strategies can an automation Center of Excellence use to blend these capabilities and unlock greater end-to-end process efficiency?
It’s a common thread I see everywhere I go. You’ll have an AI unit established as a separate, almost academic, part of the company, and then you have the process automation team that’s in the trenches running the digital workforce. The friction is palpable; sometimes the automation team isn’t even allowed to use the AI tools the other team is developing. This creates a massive missed opportunity. The strategy for a CoE is to become a bridge. They need to actively work to bring that AI capability into their existing automated workflows. By blending these two worlds, they can tackle the parts of a process that RPA could never touch, unlocking the next 20% or 30% of end-to-end automation and achieving true efficiency gains.
With over 3,500 digital workers and dozens of AI agents in production internally, you have a unique perspective. Could you share a specific example of a complex task one of your own AI agents handles and the key lessons your team learned during its integration?
We are one of the biggest users of this technology in the world, and that experience is invaluable. We have over three and a half thousand digital workers deployed, generating hundreds of millions in run-rate benefit for our business. We’ve attached about 35 AI agents to those digital workers to handle truly complex tasks that were previously out of reach. For instance, an agent might be responsible for interpreting and classifying inbound client communications that are highly variable and filled with industry jargon. The biggest lesson we learned is that this is a journey of continuous refinement. You don’t just deploy an agent and walk away. You monitor it, you see how it handles edge cases, and you fine-tune its instructions. Sharing that journey—the successes and the challenges—is how we help our customers and the industry move forward together.
What is your forecast for agentic automation in the next three to five years?
I believe that within five years, the distinction between “RPA” and “AI” in automation will largely disappear. We’ll simply talk about digital workers with a spectrum of capabilities. Fully autonomous, outcome-driven agents will handle entire business functions in some areas, but the real proliferation will be in hybrid models where AI agents act as incredibly powerful “co-pilots” for both human and digital workers. The market isn’t fully ready for complete autonomy yet, as there’s still a lot of learning to do around trust and regulation. But the journey is well underway, and we’ll see a rapid evolution from task-based bots to goal-oriented agents that are deeply embedded in the core operations of every major enterprise. There will always be another model, another advancement, so the focus will be on building flexible, adaptable automation platforms.
