Enterprises that run more than a handful of AI agents in production face risk, cost and oversight burdens that grow exponentially with every agent they add, says Felix Van de Maele, founder and CEO of data governance company Collibra. In a conversation with VentureBeat founder and CEO Matt Marshall at VentureBeat's AI Week event in New York, Van de Maele traced most agent failures to the business context those agents receive — and said companies need deterministic controls on agent behavior before their agent count grows.
"If you have one, two, five agents, that's manageable," Van de Maele said. "If you have 100, 1,000, that just explodes exponentially."
More than 75% of enterprises hit a production wall when they move from a single copilot to 10 or 15 autonomous agents, and VentureBeat's June survey of 157 enterprise technology leaders shows how little margin those companies have. Thirty four percent let agents push code or system changes into production on the strength of automated evals alone, with no human review, while only 5% of all respondents fully trust their evals.
The hallucination gap in enterprise AI
The distance between what vendors promised and what enterprises receive is the hallucination gap: Agents are in fact not always right, humans remain necessary, and AI costs are mounting. A team running five agents can absorb these challenges by assigning people to validate outputs, recertifying agents and updating the context those agents depend on, but that's brute-force management. The headcount it requires grows with the number of agents, so a company with a thousand agents cannot staff its way to safety.
Compute spending follows the same pattern. Current models can usually reason their way to a correct answer, but there are costs, Van de Maele said.
"Models are getting really good, but they're going to loop and loop and loop and loop to get there, and every pass consumes tokens,” he explained. "We've all learned, unfortunately, over the last year, that AI is not free."
Governing AI agents in production
Collibra commissioned a Harris Poll which found that nearly 30% of companies reported catching agents accessing or surfacing data the agents had no permission to see. So how can a company enforce governance once agents touch live production systems?
Inventory is the first step. At Collibra's own AI summit earlier in the week, Van de Maele asked the audience who knew how many agents their company had in production, and then who knew what those agents were doing. Most of the room could not answer either question. Companies encourage employees to build and experiment with agents, and that openness makes an accurate count hard to maintain, he said.
"How do you control what you don't even know you have?" he added.
The second step is a deployment lifecycle that defines what an agent must clear before it reaches production. But a slow process defeats itself, because teams route around any gate that stalls their work. The lifecycle has to move an agent through review quickly enough that following the process usually costs a team less time than skipping it does. The third step accounts for agents that clear the first two hurdles.
Why AI agents get enterprise data wrong
A common argument in Silicon Valley, voiced by Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman, among others, is that steadily improving models will eventually stop making these errors. That might be the case for consumer use, but not for enterprises. A consumer chatbot can give many acceptable answers to the same prompt, while a request for a company's revenue number has exactly one correct answer, and large language models struggle to return the same answer every time.
Collibra's survey found that 90% of respondents blame the lack of an aligned data foundation for keeping them from reliable results. A decade of investment in data platforms has given most enterprises enough data and enough access to it, and agents still fail because they query the wrong tables, misinterpret the fields they find, or consume records full of quality defects, Van de Maele said.
"It's more of a data governance, or I would argue a context problem, than a data problem," he explained.
His answer is an ontology that captures a company's data, the business meaning of that data, and the processes that govern its use. Van de Maele co-founded Collibra years ago to capture business context, and that same work now supplies the grounding agents need to reflect how the business operates.
That ontology should be held in a context layer independent of any model or platform. Companies will swap models, and many will change data and AI platforms over the next few years. The ontology describes the business itself, so it changes when the business changes and survives each of those platform migrations intact.
Context drift and fragmentation reduce AI agent accuracy
Better context also cuts token spending because an agent that already holds the definitions it needs skips the loops it would otherwise run to infer them. The harder problem is keeping that context accurate after deployment. One frontier model lab published recent research on its in-house analytics agent which found that the agent answered about 30% of questions correctly without context and about 90% with it. Two months later its accuracy had dropped to 60%, because context needs the same ongoing maintenance enterprises already give their data pipelines.
According to Van de Maele, enterprises will spend the next decade tackling context fragmentation, context silos and context quality — the same problems they spent the last 20 years solving for data. Context will live in Anthropic’s Cowork, ChatGPT, Microsoft Copilot, custom agent harnesses and data platforms at the same time, creating the risk that different systems and business units develop conflicting definitions of the same business terms.
How Collibra's guardian agents enforce agent contracts
Collibra's answer to the governance and context question is an agent contract, which turn existing governance context into machine-readable rules that can be read and enforced at runtime, and be consistently applied across an agent fleet. Global contracts establish common boundaries, while specialized contracts adapt those requirements to different agent categories. As new agent types emerge, governance can expand with them.
A guardian checks an agent's intended action before the agent executes it and blocks any action the contract forbids. Guardians cannot run on large language models, because an LLM check adds too much latency and cost to every agent action and its verdicts carry the same unreliability as the agent it supervises. Collibra builds its guardians to evaluate contract rules deterministically, so the same action produces the same verdict every time.
A single control won't close every gap, and the industry, Collibra included, is still researching how to manage those failures. In the meantime, AI investment should be directed at the context layer, Van de Maele said.
"We all have access to the same models, so the differentiation comes from your ability to capture your specific proprietary knowledge and make that available to agents in the most efficient way possible," he explained. "That's where your competitive advantage lies."