Collibra’s agent strategy starts with inventories and contracts

Counting agents is the first control

At a Collibra summit, Van de Meule asked attendees whether they knew how many agents were running in their companies and what those agents were doing. Most could answer neither question. Companies encourage employees to experiment with agents, he said, making an accurate inventory difficult.

“If you have one, two or five agents, you can manage them. When you have 100 or 1,000, everything grows exponentially,” he said.

That gap matters beyond bookkeeping. A Harris Poll commissioned by Collibra found that nearly 30% of companies had seen agents access or expose data without permission. A company cannot secure agents it does not know exist.

The next step is an approval process that defines what an agent must pass before deployment. But controls that take too long invite teams to bypass them. Checks need to be quick enough that following the process is easier than skipping it.

The data is there; the context may not be

The case for stronger controls is not simply that models make mistakes. For companies, the harder problem is that there is often one correct answer: a chatbot may give several acceptable responses to a prompt, but a question about company revenue has only one right answer.

In a Collibra poll, 90% of respondents said the lack of a consistent data foundation stood in the way of reliable results. Companies have spent years building data platforms and generally have information available. Yet agents may query the wrong tables, misunderstand fields or use poor-quality records.

Van de Meule’s diagnosis is that the problem lies more in data governance and context than in the data itself. Collibra’s proposed answer is an ontology: a description of company data, what it means in business terms and the rules for using it. The company was founded by Van de Meule to capture business context; now, he argues, that same layer can give agents knowledge of how a business works.

Keeping the ontology separate from models and platforms is central to the idea. Models will change, and many companies will change their data and AI platforms in the coming years. An ontology describes the business, so it can move with the company rather than being tied to a particular system.

Context decays, too

Context can also reduce token use. An agent that already knows the relevant definitions does not have to repeatedly work them out. But context is not a one-time fix: it has to stay current.

One advanced-model lab recently reported that its analytical agent answered about 30% of questions correctly without context and about 90% with it. Two months later, accuracy had fallen to 60%. The result makes context look less like a static reference library and more like a maintained operational asset.

Over the next decade, Van de Meule expects companies to face problems with fragmented, isolated and low-quality context, much as they have spent the past 20 years tackling similar problems with data. Context will sit across Anthropic’s Cowork, ChatGPT, Microsoft Copilot, internal agent systems and data platforms. Those systems and departments may end up using different definitions for the same business concepts.

Collibra’s proposed controls are agent contracts: machine-readable instructions built from existing governance rules and context, which can be checked and enforced while an agent operates. Shared contracts set common boundaries; specialized ones adapt those boundaries for particular categories of agents.

The company is also developing supervisory agents to check a proposed action before it happens and block actions prohibited by the contract. Collibra says those checks should be deterministic rather than handled by large language models, which would add latency and cost while making decisions that could be as unreliable as the agents they oversee.

I think the more important test is not whether an agent can pass a contract check, but whether the contract reflects current business reality. An outdated ontology can make enforcement consistent without making it correct. Collibra’s argument is that companies can gain an edge by collecting their own knowledge and delivering it effectively to agents, since everyone has access to the same models. That advantage depends on treating context as something to maintain, not something to write once and forget.