AI agents have stormed the headlines and taken over tech conversations, especially in enterprise circles. But in all the noise, it’s hard to distinguish substance from buzz. That’s why we hosted a no-fluff, straight-talk webinar with Dataiku CEO Florian Douetteau to unpack what’s real, what’s overblown, and what enterprise leaders should actually be doing right now. If you missed the live session, good news: the replay is now available — and it’s worth your time.
Here’s a taste of what we covered and why this session stood out in a sea of overhyped AI content.
The Real Deal on AI Agents: Not Just "New Automation"
Florian kicked off the session by tackling the most fundamental question: is “agentic development” just a fancy way of saying automation?
Spoiler: It’s not. While automation has long existed in software development, AI agents represent a new paradigm with different trade-offs. Traditional software requires deterministic logic, detailed programming, and predictable execution. AI agents, by contrast, operate in a more probabilistic, generative world. They can choose execution pathways, reason over unstructured data, and adapt in ways older automation tools never could.
But that doesn’t mean they’re magic. It means enterprises must grapple with new questions: How do you balance determinism and flexibility? When should you allow agents to “reason,” and when should you lock them into predefined workflows?
Pathways, Planning, and Pragmatism
One of the webinar’s most talked-about points was Florian’s breakdown of “pathways explosion,” or the idea that agents can, in theory, follow infinite sequences of actions. This can be powerful for open-ended tasks like research or discovery. But in the enterprise, unpredictability is rarely a feature — it’s a bug.
Smart organizations are designing semi-deterministic agents: systems that can adapt within guardrails. They might let agents pick between a few preset paths, but not wander off into the digital wilderness. This approach offers a more practical balance between innovation and control, especially in highly regulated or mission-critical environments.
Prompt Engineering: Fad or Foundation?
Another audience-favorite segment? Prompt engineering. Is it a new engineering discipline, or just glorified fiddling with words?
According to Florian, it’s both. Prompt engineering is a skill — especially when used to chain tools, direct workflows, or structure output in precise formats. But much of it is rapidly becoming commoditized as large language models (LLMs) improve and require less hand-holding. What matters more in the long run is the ability to orchestrate agents across tools, APIs, and data sources; essentially, using prompts to build intelligent systems, not just chatbots.
Use Cases That Matter Right Now
There was also plenty of clarity around the question on everyone’s mind: what are AI agents actually good for today?
Florian highlighted several enterprise-grade use cases where agents are already proving value:
- Knowledge assistance and retrieval: Automating research, summarizing documents, and answering questions using internal data.
- Market and RFP intelligence: Gathering competitive intelligence or responding to complex RFPs using structured and unstructured data.
- Customer service augmentation: Automating responses or surfacing relevant information to agents assisting human support.
- DevOps copilots: Accelerating code generation, debugging, and documentation.
Still, he cautioned that we’re in the “messy middle.” Many organizations are launching agent pilots, but scaling them to production remains the bigger hurdle.
From Prototype to Production: Why Change Management Is the Hard Part
Florian emphasized a truth that too many vendor pitches gloss over: building an agent is the easy part. Moving it into production, integrating it into workflows, managing change across teams — that’s where the real work begins.
He stressed the importance of a clear roadmap and measurable impact. ROI doesn’t come from cool demos; it comes from real-world adoption. That means knowing when to use out-of-the-box tools and when to invest in custom agent development tied to your unique data, processes, and goals.
Don't Wait — But Don't Wing It, Either
One audience member asked whether it’s too early to adopt AI agents. Florian’s response? “It’s not too early — but you do need to be thoughtful.”
He noted that every enterprise has a different risk appetite, but no one should be sitting on the sidelines. That doesn’t mean diving into every shiny new tool. It means experimenting with purpose, learning what works, and building internal literacy around agent design, testing, and governance.
The Need for a "Science of Agents"
One of Florian’s most compelling insights was his call for a “science of agents.” Just like data science formalized testing and validation for models, we now need similar frameworks for agents. Without this rigor, agents will remain unpredictable black boxes — hard to trust, hard to scale.
Dataiku’s vision is to help bridge this gap by offering an orchestration layer for analytics, models, and agents, combined with tools for data governance, security, and collaboration. The platform enables IT to manage risk while empowering business users to build and deploy agents in a controlled, scalable way.
What Makes Agents a Competitive Advantage?
Toward the end of the session, Florian tackled a provocative question: Can agents actually be a competitive differentiator?
His answer: Only if you build your own (brings back the old build vs. buy argument in a new way).
Off-the-shelf agents can offer quick wins. But if every company uses the same tools, powered by the same data, their outputs start to look eerily similar. Enterprises that build their own agents — rooted in proprietary data, internal workflows, and business context — will be the ones that maintain differentiation in the age of intelligent automation.
Final Thoughts: Practice > Theory
If you’re feeling overwhelmed by all things agents, you’re not alone. Florian offered one final piece of advice: Just start building. Pick a small, repetitive task you know well. Think through how an agent might replicate it using a few tools or APIs. Learn by doing.
Because ultimately, the path to understanding — and benefiting from — AI agents isn’t paved with theory. It’s paved with experiments, iterations, and real-world applications.