AI Agents Need More Than Data — They Need Metadata
In February, Klarna — the Swedish fintech best known for its "buy now, pay later" services — revealed that its new OpenAI-powered assistant had handled two-thirds of all customer service chats in its first month, adding around $40 million to the bottom line.
The real breakthrough, however, wasn't the AI model itself — it was the metadata behind it.
Metadata is data about data: it describes where information comes from, what it means, and how it should be interpreted. Klarna's team didn't just connect GPT to their helpdesk; they rewrote their manuals and workflows to be clear, structured, and machine-readable. By improving the quality of their documentation — effectively their internal metadata — they gave AI a foundation of meaning and context.
Metadata in the Real World
In the digital world, metadata distinguishes a document's contents from its authorship, history, and reliability. In the operational world — factories, grids, and utilities — it makes the difference between a value and a decision that can be trusted.
For example:
That's valid data, but meaningless in isolation. Is it a temperature from a heat exchanger or a storage tank? Was it recorded two minutes ago, or two hours ago? Is another device — perhaps from another vendor, operating in the same network — showing a similar anomaly? What about a third?
Without metadata, even the most capable AI is guessing. With it, AI can reason, verify, and act safely in the real world.
The State of Operational Data
Across modern infrastructure, billions of sensors, controllers, and machines generate data every second. Yet most operational technology (OT) data remains inaccessible — locked within non-IP field networks and proprietary systems never designed for sharing.
Of the fraction that is available, much arrives with little or no context. APIs and middleware were built to transmit data, not to explain it. They deliver numbers and timestamps, but omit the relationships, reliability indicators, and operating states that define meaning.
The result is automation that appears intelligent but behaves inconsistently. It's the same pattern seen in coding agents that lose track of variables mid-task: once context disappears, logic collapses.
Understanding the Challenge
In OT environments, the issue is rarely poor data quality — it is poor data understanding. A temperature reading from a smart meter, a valve position from SCADA, and a cost entry from ERP may all describe the same process, yet they exist in separate silos with different schemas and assumptions.
For AI to reason safely, the documentation must become dynamic — a live metadata layer that explains:
what each value represents, where it came from, how reliable it is, and how it connects to other data sources.
This is precisely what Altior was built to provide.
How Altior Adds Metadata
Altior creates this layer of meaning as data moves, not after the fact. It acts as a semantic overlay that enriches data in real time, working alongside existing infrastructure rather than replacing it.
It does this through two complementary routes.
1. Creating Digital Instances of Real Devices
Altior can create a lightweight digital instance — a digital twin — for each physical device or data source. Every message from that device is automatically tagged with metadata such as unit, timestamp, device type, quality, and relationships, making the data self-describing.
Before enrichment:
After enrichment:
Now each message carries its own context — interpretable by both humans and AI systems.
In large-scale environments such as smart water networks, this enables AI to reason across thousands of devices, detect anomalies early, and prioritise genuine faults over noise — saving time, cost, and carbon.
2. Adding Metadata Above APIs
Many organisations rely on existing APIs — from SCADA, IoT, or ERP systems — that exchange data without meaning.
Altior also works above these APIs, reading their structure, identifying relationships, and enriching data dynamically. It determines:
the source and schema of each feed, the relationships between data points, and the links between systems — for instance, a pump in SCADA that corresponds to a maintenance record in ERP.
By triangulating data across APIs, Altior aligns timestamps, validates consistency, and builds a coherent context graph.
Even when APIs lack meaning, Altior can infer and attach it, turning fragmented feeds into coherent, machine-readable documentation.
The commercial benefits are immediate: if a pressure drop or voltage fluctuation can be confirmed remotely through correlated data, operators can avoid unnecessary truck rolls — cutting cost, downtime, and emissions.
AI That Understands What It Sees
By enriching data at the source or in transit, Altior ensures that every message carries consistent meaning from device to data lake to AI model.
AI agents can then operate on context-rich information, able to reason across systems, test their logic, and improve over time — much as an engineer validates a result against physical reality.
Metadata as the Next Advantage
Every successful AI deployment — from Klarna's customer assistant to industrial control systems — depends on how clearly we describe the world AI is meant to reason about.
Data without metadata is like conversation without memory.
Altior brings that missing layer of meaning to the physical world. It transforms disconnected APIs and legacy systems into a shared operational language, where every signal can be trusted, traced, and understood.
As Sequoia Capital observed in its analysis of modern AI, the next era of progress will come not from larger models, but from better data. And the measure of that quality is metadata — the structure that gives information purpose, provenance, and trust.