Building an AI-ready data mesh
In this article...
- Why analytics-ready data platforms fall short when AI and agents enter the picture
- How Amadeus evolved its data mesh to support real-time, context-rich, and machine-ready data
- What it takes to move from human-driven analytics to AI-ready data that enables agents to interpret, reason, and act
Many companies have spent the last few years investing in modern data platforms. Most of that effort initially focused on supporting analytics and helping people make better decisions.
With the rise of artificial intelligence, and as Amadeus continues transforming travel with AI, that foundation is being tested. Data is now consumed by machine learning models, LLMs, and autonomous agents, which interact with data much differently than humans, pushing platforms to evolve.
At Amadeus, this shift was particularly visible. Our technology powers a global travel ecosystem where data isn’t just a byproduct, it’s a core asset.
So, the question wasn’t whether this would change the way we approached data. The real question was whether we needed to rebuild or adapt our data mesh to serve the new agentic user.
How Amadeus’ data mesh is structured
To fully harness data, we originally built a mesh grounded in three pillars:
- First, domain ownership. Each domain, such as airlines, airports, and hospitality, owns its data and is responsible for governing and sharing it.
- Second, data as a product. Data is designed for consumption and exposed in different ways, from datasets to APIs, following key principles: completeness, trustworthiness, discoverability, accessibility, and security.
- Third, Data 360. Data is transformed into knowledge and fed back into products, improving personalization and decision-making over time.
This setup is supported by a medallion architecture:
Bronze layer: stored as ingested, preserving full history for audit and lineage
Silver layer: clean, reliable, and structured format, with relational models with cross-domain correlations
Gold layer: pre-computed KPIs and data marts for direct usage
This model has worked well for human consumers. However, the rise of AI introduces new expectations.
Why data must be rethought for AI
For humans, good data is curated, aggregated, trustworthy, and explainable. It aligns with how people interpret information.
Data for AI, however, must be:
raw and atomic, since noise can help detect edge cases.
semantically explicit, because AI cannot reliably infer meaning or context on its own.
exhaustive, as it can’t compensate for missing information.
contextual, because missing context leads to incorrect conclusions and hallucinations.
Here is a direct comparison of traditional analytics and agent-based data usage:
| Data dimension | Traditional BI | With agents |
|---|---|---|
| Context | Users know why they query data | Context must be persisted in short and long-term memory |
| Data semantic | Users rely on business knowledge | Meaning must be explicitly defined and linked across domains |
| Data aggregation | Aggregated data is preferred | Access to both aggregated and raw atomic data is required |
| Data freshness | Daily refresh is usually enough | Real-time data is needed for accurate decisions |
| Taking actions | Actions are manual | Agents call APIs to act directly |
Despite these differences, one thing remains constant: data governance. For humans and agents alike, access rights and policies apply equally.
Ultimately, this shift required us to evolve our data mesh to make it AI-ready.
How we adapted our data mesh for AI
Context
Humans querying data already understand their intent. Agents don’t.
We introduced mechanisms to store short and long-term context within the data mesh, allowing agents to retrieve relevant history. This requires reliably identifying the agent to ensure the right context is applied.
Data semantics
Acronyms like RASK (Revenue Available Seat Kilometer), ASK (Available Seat Kilometer), or LF (Load Factor) are standard in the airline industry, but agents don’t recognize them.
We strengthened our semantic layer to describe each data field in detail. This layer is part of the Amadeus Travel Data Model, our central repository of data definitions. Making relationships between datasets explicit, including across domains, allows AI to reason more accurately.
Data freshness
Daily batch updates are enough for analytics, and they help manage compute costs. For AI, this is often too slow. Agents need up-to-date data to make timely decisions.
We introduced a real-time data layer that complements the batch pipeline. It combines large-scale storage with fast processing, enabling a shift from reactive decision-making to more proactive monitoring and action.
Data aggregation level
Humans rely on aggregated views to make data easier to consume. Agents work differently.
They can start with KPIs, but they must be able to drill down into raw and atomic data to explain changes, detect anomalies, or identify edge cases that may not appear in aggregated views. Also, through data semantics and context, AI agents can switch from aggregated to atomic data by understanding the business logic behind each KPI.
Ultimately, providing access to both layers of data is essential.
Taking actions
Decisions based on data typically involve humans coordinating. However, agents allow actions to be executed directly through APIs, changing how they must be designed.
Traditional APIs assume the client knows the correct sequence of steps. AI assistants do not. They must decide what to do next at runtime, based on system state, user intent, and incomplete information. As a result, APIs need to make workflows explicit, machine-readable, and recoverable.
In practice, this means an assistant must be able to understand what exists, what state it is in, what actions are possible, what is missing, and how to recover safely from failure. Instead of relying on hard-coded logic, agents can navigate systems by interpreting their current state.
Of course, this opens possibilities but also adds risk, as action‑enabled APIs can lead to undesired outcomes. Therefore, access granted to agents, particularly when they can update systems, needs to be tightly governed and properly controlled.
From insight to action, powered by the same data
Analytics-ready data helps humans understand what has happened and why. It powers dashboards, reports, and KPIs.
AI-ready data builds on that foundation, yet it enables machines to interpret data, learn from it, and take action.
At Amadeus, moving from analytics-ready to AI-ready wasn’t a complete redesign. It was a natural evolution of the data mesh we had already built.
The result is a platform that supports both humans and agents, each using the same data in different ways. And that balance is what will shape the next phase of data platforms.
TO TOP