Many teams are learning the hard way that building and deploying AI agents at scale requires much more than plugging into an LLM and hoping for the best.
In a recent Let’s Talk Agents web series session, “Why Dataiku for AI Agents,” Christian Capdeville, Senior Director of Content and Product Marketing at Dataiku, and Kurt Muehmel, Head of AI Strategy at Dataiku, unpacked the common pitfalls organizations face when implementing AI agents, and laid out a clear, actionable vision for creating, deploying, and managing agents with purpose and control. This blog recaps the main takeaways of the session.
The Agent Hype: A Spectrum of Maturity
Christian opened the session by polling attendees on where they are in their agent journey. Unsurprisingly, responses were mixed — some are brainstorming, others are building, and a growing number already have agents in production.
About two months ago, we saw seven percent of folks with deployed agents. Now, that number is growing. Progress is happening.
— Christian Capdeville, Senior Director of Content and Product Marketing at Dataiku
But with progress comes difficulty. The most cited challenge in deploying agents? Connecting to the right data and systems. Other struggles include handling governance and monitoring and evaluating agents.
Making Sense of the Agent Ecosystem
Rather than obsess over narrow definitions (what makes an AI agent “agentic” enough?), Kurt emphasized the importance of thinking in terms of capability and intent.
He classified AI agents into three primary categories:
- Infrastructure Providers: Think OpenAI, Anthropic, and the cloud giants. They provide power and flexibility — but require significant engineering work and often lack unified monitoring.
- Point Solutions & Startups: Vertical or niche tools that are often user-friendly but raise questions about scalability, governance, and vendor longevity.
- Enterprise Application Agents: Tools built into platforms like HubSpot, Workday, and ServiceNow. These integrate well within single systems, but tend to reinforce silos and resist cross-functional customization.
A New Class of AI-First Enterprise Applications
The future, Kurt argued, lies in cross-application AI agents — systems that can bridge data and functionality across traditionally siloed platforms. He painted vivid examples:
- An IT support agent that traverses ticketing and CRM systems to serve customers holistically.
- A supply chain agent that automatically augments supply chain data with a demand forecast model using the latest data.
- A marketing agent that streamlines content creation and review — all across disparate tools.
This vision aligns closely with Dataiku’s mission: to act as The Universal AI Platform™ where organizations can create these new types of modular, composable, cross-functional applications.
What Went Wrong? A Cautionary Tale
To make things concrete, Kurt shared the fictional — but highly relatable — story of “ReAI Estate,” a real estate firm that rebranded and dove headfirst into building AI agents … and promptly hit a wall.
The team built an agent to support real estate agents — handling scheduling, pricing, document summarization, and more. The agent built by ReAIestate quickly fell short of expectations due to several foundational flaws.
- It was disconnected from the actual needs of its end users — real estate agents — because the technical team lacked sufficient collaboration with the field team to understand their workflows and requirements.
- The agent relied on weak or brittle data pipelines, leading to frequent inaccuracies and confusion, such as mixing up neighborhood data.
- The agent regularly crashed during peak usage periods, particularly when agents were trying to close deals at the end of the month.
Agents are systems of systems… Hallucinations can result in cascades of errors that can be difficult to predict.
— Kurt Muehmel, Head of AI Strategy at Dataiku
The moral? Without centralized control, fragmented tools, data, and responsibilities lead to chaos.
Dataiku’s Vision: Creation + Control
The Dataiku approach rests on two pillars — creation and control — both built to meet the needs of technical experts and business users alike.
Creating AI Agents in Dataiku
Dataiku supports both visual and code-based agent development:
- Visual Agents: Build logic using natural language prompts and point-and-click interfaces. Ideal for analysts and domain experts.
- Code Agents: Use frameworks like LangChain for custom logic and tool creation, giving developers the freedom to fine-tune behavior.
Prebuilt tools give agents the ability to read and write to datasets, get predictions from ML models, send notifications through Slack or email, integrate with ticketing systems like ServiceNow or Jira, and more.Developers can also create custom tools to extend functionality.
All of these tools are deeply connected to Dataiku’s existing capabilities. You don’t need to reinvent the wheel.
— Christian Capdeville, Senior Director of Content and Product Marketing at Dataiku
Controlling AI Agents in Dataiku
Three control pillars underpin scalable, enterprise-grade agent systems:
1. Enterprise Orchestration
- LLM Mesh: Centrally manage access, routing, and security across models like OpenAI, Anthropic, or Hugging Face.
- Safe Guard: Implement rules for agent behavior, especially for mission-critical use.
- Agent Connect: A single conversational interface where users describe tasks in natural language, and multiple agents will work together behind the scenes to get it done.
Agent Connect lets you orchestrate multi-agent workflows in a way that’s simple, traceable, and reliable.
— Christian Capdeville, Senior Director of Content and Product Marketing at Dataiku
2. Continuous Optimization
- Cost Guard: Set usage budgets by project, user, or connection to prevent runaway costs.
- Quality Guard: Automate evaluations to ensure reliability.
- Trace Explorer: Visually audit each step of an agent’s actions — vital for debugging and iteration.
3. Central Governance
- Approval Workflows: Enforce custom sign-offs before deployment.
- Registries: Track agents and LLMs just like models.
- Value & Risk Monitoring: Map ROI and criticality across agents to prioritize oversight.
What About External Agents?
Can Dataiku agents work with third-party agents like those in Salesforce or AgentForce? The answer: Yes. Dataiku is aggressively agnostic. Agents built outside Dataiku can be connected via their APIs as custom tools and integrated into larger agent systems.
This flexibility ensures that teams aren’t locked into walled gardens and can leverage existing investments while maintaining control.
AI Agents in the Wild: Deployment and Data Quality
Two final questions highlighted practical considerations:
- Where do agents run? Agents can run anywhere — on Snowflake, AWS, or even on-prem — while remaining governed in Dataiku.
- How do we ensure data accuracy? “The same way we always have,” Kurt said. Use golden datasets, data quality checks, and prep tools to ensure your agents work with trusted, validated data.
Final Thoughts: Build the Right Way, Right Now
The future of enterprise AI is agentic, but success hinges on doing it right from the start.
We’re still early. You have a window to build agent architecture the right way — before the tech debt piles up.
— Kurt Muehmel, Head of AI Strategy at Dataiku
And with Dataiku, organizations can build agents that are not only powerful, but also governable, reusable, and aligned with real business goals.