Collibra Wants to Be the Air-Traffic Control for Your Company's AI

At a large company, “customer” can mean a buyer, an account, a household, a patient or the legal entity whose name appears on an invoice. Every department may own a plausible answer. Then an AI agent arrives, asks the database for customer revenue and returns one crisp number with the confidence of a game-show host. The number can be wrong before the model has made a single mistake.

That semantic mess is Collibra’s home turf. The company sells enterprise software that identifies data, records what it means, assigns an owner, traces where it came from, measures whether it is healthy and applies rules about who may use it. The result is part catalog, part map, part policy book and part workplace for the people charged with keeping data useful. In 2026, Collibra is stretching that machinery over AI models and autonomous agents and calling the whole thing an “enterprise AI control plane.”

The phrase sounds freshly minted. The underlying bet is old: shared meaning becomes infrastructure when enough people and machines depend on it.

A university spin-off goes hunting for nouns

Collibra began in 2008 at Vrije Universiteit Brussel’s STARLab, the Laboratory for Semantic Technology and Applications Research. Founders Felix Van de Maele, Stijn “Stan” Christiaens, Pieter De Leenheer and Damien Trog had been working on ontology - a formal way to describe concepts and the relationships between them. The academic language was dry. The commercial problem was not. Businesses were accumulating databases faster than they could reconcile what was inside them.

Their research method carried the wonderfully severe acronym DOGMA, short for Developing Ontology-Guided Mediation for Agents. The team joked about having a dogma at a free-thinking university. It also foreshadowed the company’s current pitch: machines need governed meaning if they are going to act on an organization’s behalf.

The company was a slow burn by venture standards. It raised about €1 million in 2012 and explored the American market. A $23 million Series B arrived in 2015, followed by $50 million in 2017 and $58 million in 2018. Then the curve steepened: $100 million in 2019 made Collibra Belgium’s first technology unicorn; $112.5 million in 2020 valued it at $2.35 billion; and a $250 million Series G in November 2021 put the post-money valuation at $5.25 billion. Altogether, public databases place total funding around $600 million.

What changed investors’ minds was not one demo. Privacy rules such as GDPR made undocumented data risky, cloud migration scattered it across more systems, and analytics moved from a specialist activity into ordinary work. Each trend made metadata - data about data - less like filing and more like operations.

The value is not the list of tables. It is the agreement around what those tables mean, who owns them and what is allowed to happen next.The Collibra thesis, stripped of product names

Five questions, one metadata graph

A user searching Collibra might find a sales dataset, see a plain-English definition, identify its steward, inspect quality scores, follow lineage back through a pipeline and request access. An analyst can discover a certified data product instead of opening another Slack thread. A privacy officer can locate personally identifiable information and attach policy. An engineer can see which dashboards will break before changing a column. An AI team can register a use case, document its model and training data, assess risk and monitor it after launch.

The current platform spans Data Catalog and Marketplace, Data Governance, Data Lineage, Data Quality & Observability, privacy and access governance, plus AI Governance and the newer AI Command Center. More than 100 integrations pull metadata from systems including Databricks, Snowflake, Google Cloud, Azure, SAP and Tableau. The platform does not replace those systems. It sits above them, joining their technical facts to business context.

That cross-platform view is the main distinction from a cloud-native catalog. Databricks Unity Catalog or Microsoft Purview can be compelling inside their own ecosystems. Alation, Informatica and Atlan compete more directly across environments; point products such as Monte Carlo attack quality and observability. Collibra’s wager is that a large company wants one graph connecting definitions, policies, lineage, quality and AI controls even when its data lives everywhere.

From documenting data to steering agents

Collibra has bought its way into the rough edges of that graph. OwlDQ, acquired in 2021, strengthened data quality. SQL notebook maker Husprey came aboard in 2023. Raito added data-access governance in 2025, while Deasy Labs brought automated discovery and enrichment for contracts, transcripts, emails and other unstructured material. That last category matters because enterprise AI reads far more than tidy database rows.

In May 2026, Collibra launched AI Command Center. Its ambition is to move governance from a form completed before launch to a live system that inventories models and agents, shows ownership and risk, traces dependencies, watches behavior and supports intervention. Partnerships with Databricks, Snowflake and Google Cloud push Collibra’s business context into the platforms where customers build and run AI. The company’s open-source MCP server, CHIP, lets agents query governed assets and glossary definitions directly.

This is the boldest part of the strategy and the least settled. “Control plane” is becoming crowded territory. Cloud providers, model platforms, identity vendors and observability companies all have reasons to claim the layer above AI. Collibra brings a mature metadata graph and deep relationships with data offices. It still has to prove that governance teams can influence runtime behavior quickly enough for agentic software.

The first thing to fail is usually not software

Collibra says it serves more than 700 customers, including dozens of Fortune 500 companies. Its public stories cover Heineken, Toyota Motor Europe, TELUS, Western Governors University, UC Davis Health and Northern Trust. The appeal is clearest in regulated or complicated organizations: banks documenting critical data for regulators, hospitals protecting health information, manufacturers tracing definitions across regions and telecom companies trying to make datasets discoverable.

The outcomes can be concrete. TELUS reported reducing the time spent searching for data from five hours to 50 minutes. Western Governors University says data curation and governance work fell from weeks to seconds. An IDC study commissioned by Collibra calculated a 484 percent three-year return across interviewed customers. These are reported results, not a promise attached to every contract.

The more revealing case is the one that failed first. Froedtert & the Medical College of Wisconsin launched a governance effort in 2017 without a dedicated governance team; ownership sat largely with IT. It restarted in 2018 after securing executive support and creating a cross-functional Data Citizenship and Governance Consortium. Collibra became a home for the resulting assets, but the change in sponsorship and operating model came first.

Without leadership buy-in and user engagement, governance programs may go live but fail to deliver visible value and often collapse.Collibra, on why adoption takes time

That is also the answer to “what did it cost?” Core platform prices are negotiated, not posted, and certain AI features consume usage credits called Collibra Units. The bigger bill is organizational: implementation, integration, training, data owners’ time and the continuing work of keeping definitions current. A half-populated catalog is not half valuable. It can be actively misleading.

Buy the operating model before the platform

The lesson readers can copy does not require a Collibra contract. Start with one expensive confusion: a regulatory report nobody can reproduce, a metric with three definitions, a dataset analysts cannot find or an AI use case whose inputs nobody owns. Name the executive accountable for fixing it. Assign stewards. Connect the business term to the actual tables and pipelines. Measure search time, error work, access turnaround or audit effort. Only then widen the domain.

The approach will not work when leadership treats governance as an IT cleanup, when employees receive no time or incentive to steward data, or when the organization is too small to need a formal layer. In those cases, a disciplined wiki, clear owners and a cloud platform’s native catalog may be enough. Collibra earns its complexity only when the company already has complexity worth governing.

Its 18-year journey is a neat inversion. Collibra started by making abstract semantics practical for people. AI now makes those semantics practical for machines. Whether the company owns the control plane is undecided. But the problem it began with has become harder to ignore: if a business cannot explain what its data means, its agents will explain it for them.