Data governance: 6 keys to a successful agentic AI project

As an information systems manager and leader, Guillaume Pommier supports his clients in defining, deploying, and securing their technical infrastructures, as well as steering digital transformation projects.

Key takeaways:

• 40% of agentic AI projects will fail due to poor data quality: data preparation is your best insurance policy.
• Multi-agent orchestration is the cornerstone of high-performing systems.
• We are on the verge of a transition from a human web to an agent-driven web: prepare your infrastructure for the Agent Experience.
• A gradual and secure approach is essential for the successful deployment of agentic AI.

From ChatGPT to virtual employee: how agentic AI supercharges your ROI

Generative AI has created a lot of buzz with the emergence of sophisticated chatbots. However, the real revolution lies elsewhere: moving from a simple chatbot that answers your questions to an autonomous virtual employee that generates ROI 24/7. A tool like ChatGPT can give you an answer; an AI agent goes far beyond that: it acts, decides, and executes.

Agentic AI is built on three fundamental pillars:

• Planning
• Action
• Reasoning

Imagine an agent capable of planning a marketing campaign from A to Z, executing it, and then continuously optimizing it, all autonomously. This business impact translates into a 60% reduction in processing time for repetitive tasks, freeing up your teams for higher-value missions.

The evolution towards coordinated multi-agent systems is the next crucial step. Think of a team of specialized agents – one for data, another for creative, a third for media, and an analyst – all working in concert. It is this synergy that will multiply efficiency.

According to predictions from AWS, the window of opportunity is short before this technology becomes widespread. By the end of 2026, 70% of Fortune 500 companies will have already deployed agents.

Architecture at the center of project success

L’Oréal is automating its media campaigns with agents. What’s your use case? This is the first question to ask. The architecture of your agent systems is a determining factor. Choosing between cascading agent chains or parallel agents is not a trivial decision: a poor architecture can lead to a project that costs 3x more and delivers only 50% of the expected value.

Take the example of cascading agents, which are perfect for a sequential process like lead qualification, content creation, and then distribution. Conversely, parallel agents will excel at simultaneous multi-channel A/B testing, thereby maximizing the effectiveness of your campaigns.

These cascading and parallel agent models are two key implementations of multi-agent orchestration. The architecture goes beyond this, including, for example, hierarchical systems (where senior agents delegate and supervise the tasks of subordinate agents) or Blackboard architectures, which are ideal for complex problems requiring contributions from multiple specialized agents to a centralized workspace.

The choice of architecture is therefore a strategic decision that directly impacts the project’s performance and cost. Whatever approach is chosen, one element of these systems proves indispensable for risk management and continuous optimization: control.

Monitoring and quality control agents are another revolution. They promise an 80% reduction in costly errors and automatic continuous improvement. Imagine a supervisor agent that detects budget drifts in real-time, thus preventing massive financial losses.

Concrete use cases are already numerous and their impact is measurable:

• Marketing automation: +35% conversion.
• Media buying: -40% customer acquisition cost.
• Customer experience: Increased personalization in retail, fraud detection in finance, diagnostic assistance in healthcare.

The transition from UX to AX (Agent Experience) is a major disruption. Your current interfaces are at risk of becoming obsolete within two years. This is a strategic opportunity: early adopters could capture 60% of the market.

6 keys to avoiding failure in your agentic AI project

1. Data quality

The quality of your data is the Achilles’ heel of any agentic AI project. Outdated data can lead your agents to make large-scale erroneous decisions. For example, an e-commerce agent could cause your company to lose €200,000 in one week due to outdated product data. Freshness, quality, and consistency of data are essential criteria for a successful agentic project.

2. Dataset structuring and preparation

Structuring and preparing datasets is essential. A poorly structured product taxonomy could, for instance, lead your agent to recommend competing products, thereby sabotaging your sales efforts. This is a cost that must be viewed in terms of its ROI: for every €1 invested in data preparation, you save €5 in subsequent maintenance and corrections.

3. Data protection

Without security, there is no trust: protect your data to protect your value. Your agents are ready to achieve unprecedented performance thanks to high-quality data. But this entire structure collapses if its weakest link fails: security. A data breach doesn’t just expose your company to GDPR fines; it destroys the customer trust needed to achieve promised conversion gains and ruins your reputation. That’s why securing data flows (encryption, access management, strong authentication) is not a secondary technical issue, but the foundation upon which your entire strategy of controlled autonomy rests.

4. Technology choices and integrations

Technology choices and integrations are major strategic decisions. Being tied to a single provider (vendor lock-in) versus prioritizing interoperability can mean being 3 years ahead of or behind your competitors. Carefully evaluate Microsoft, OpenAI, or open-source stacks based on your specific needs.

5. Performance metrics and monitoring

Performance metrics and monitoring are essential. Companies that closely measure their agents show 40% higher performance. Implementing a real-time dashboard to track success rates, latency, and cost per transaction is crucial for continuous optimization.

6. Human supervision

Finally, human supervision is the key to scalable and secure agentic AI. Testing and validation in a controlled environment are imperative. It is recommended to use a test environment, or “sandbox,” to ensure a controlled deployment. Starting by allocating a small portion of the budget, for example 1%, to a test agent before generalizing its use is a pertinent strategy.

Furthermore, the traceability of agent decisions helps ensure regulatory compliance and continuous process improvement. Implementing a complete audit trail, including the justification for each decision made by the agent, is recommended to facilitate monitoring and system optimization.

Agentic AI is a (r)evolution that is redefining operational efficiency and competitiveness. The companies that seize this opportunity now will position themselves as undisputed leaders. Are you ready to transform your business with agentic AI?