In a recent conversation between AI keynote speaker and Trust Insights co-founder and chief data scientist Christopher S. Penn and Dataiku head of AI strategy Kurt Muehmel, the two explored a topic on the minds of many data and digital leaders: What does it actually take to build and scale an AI agent in a modern enterprise?
As the age of GenAI matures, organizations are no longer satisfied with chatbots and simple automation. They’re looking for intelligent, autonomous systems — AI agents — that can reason, adapt, and execute complex tasks with minimal oversight. But building such agents isn’t as simple as connecting to an LLM and letting it run wild. It’s a journey and one that mirrors how companies have approached automation for decades. Be sure to check out part one and part two of this blog series as well.
3 Stages of AI Agent Evolution: Done by You, With You, and for You
Penn outlines the progression to full AI autonomy in three distinct stages:
- Done by You: This is the foundational workflow level. Think of it as a recipe — you’re manually entering prompts, moving data between systems, and reviewing output yourself. It’s labor-intensive, but it allows for deep understanding and experimentation. “You are the one copying and pasting prompts … It is a workflow. You can document it,” said Penn.
- Done With You: At this intermediate stage, parts of the process are automated. You may build a custom GPT model with pre-loaded context or system instructions, so you don’t need to reinvent the wheel every time. It still requires human input and iteration, but less effort and fewer repetitive tasks.
- Done for You: This is the holy grail — an AI agent that runs entirely on its own. “You don't do anything except check your email,” Penn explains, describing a use case where an agent monitors the AP News front page, analyzes the impact on a company’s operations, and delivers a daily report, fully autonomously.
A Simple but Practical Use Case: The News Alert Agent
One compelling example discussed is the News Alert Agent, a system that ingests daily news, contextualizes it for a specific business, and provides actionable insights. Imagine you’re a multinational company. An AI agent checks the front page of the Associated Press each morning. It then:
- Pulls in relevant news articles
- References your company’s value proposition, customer base, and marketing strategy
- Evaluates whether any current events might affect your business
- Generates a summary report and sends it directly to your inbox
For example, if the AI detects new tariffs or geopolitical tensions that affect your target markets, it can flag potential revenue risks or operational impacts — all without human intervention. This kind of capability is no longer science fiction. It’s feasible with today’s LLMs, given the right scaffolding. But the key lies in building it step-by-step — starting with a manual prototype, then automating parts, and finally deploying it as a fully autonomous agent.
Infrastructure, Security, and Enterprise Realities
So what’s the catch?
As with most things in enterprise AI, the biggest barriers aren’t the algorithms: They’re infrastructure and security. “For a lot of enterprises, [building agents] means having that infrastructure, having the ability to generate and run code, and dealing with all the joys of security that come with that,” said Penn.
Enterprise-grade AI agents need to integrate safely with internal systems, databases, APIs, and existing workflows. They must be auditable, consistent, and secure. Dataiku enables this kind of evolution — from workflow building blocks, to scheduled automation, to full agents.
The tooling exists. But enterprises need clear governance, modular design, and a strong experimentation culture to navigate the transition.
How Broad Should an Agent’s Capabilities Be?
Another important design decision is how narrowly or broadly to scope an AI agent.
Should you create one agent per task — like a news analyst or customer complaint summarizer? Or should you build a mega-agent that can handle a wide array of tasks?
Penn suggests taking a page from the Unix philosophy: Build small, modular tools that do one thing well. “The narrower you scope it, the less randomness there is,” he notes. This is critical, because language models are probabilistic by nature. The more you ask of them in a single agent, the more unpredictable the outcomes can be. You can, though, always build a “meta-agent” that directs a request to the correct specialized agent, like with Dataiku Agent Connect.
This isn’t just a technical consideration — it’s a strategic one. In highly regulated or risk-averse industries, minimizing uncertainty is essential. Scoping agents narrowly allows for more consistent results, easier troubleshooting, and better integration with enterprise quality controls.
Building AI Agents Is an Evolution, Not a Leap
For companies looking to adopt AI agents, the advice is clear: Don’t try to build the self-driving car from day one. Start with a bicycle. Then add an engine. Eventually, you’ll get to autonomy — but only after mastering the basics.
The journey starts with documenting a repeatable workflow. From there, use automation to reduce effort and increase speed. Once the system is proven, you can hand over the keys to an AI agent that does the job for you, end to end.
As Penn and Muehmel emphasize, this kind of transformation requires patience, structure, and a healthy respect for complexity. But for companies willing to invest the time and rigor, AI agents can unlock a new frontier of intelligent automation.