The Enterprise AI Bottleneck Is No Longer Just the Model

Enterprise artificial intelligence is entering a new phase. The first challenge was getting AI to work. The next is making it work reliably across an entire organization without allowing costs, security risks and operational complexity to spiral out of control.

That distinction matters because an AI agent is not simply a chatbot that answers questions. Depending on its permissions, an agent can retrieve sensitive information, execute workflows, modify systems or push code into production. As companies deploy more of these systems, the consequences of a mistake become harder to contain.

At VentureBeat’s AI Week event in New York, Felix Van de Maele, founder and CEO of data governance company Collibra, argued that businesses need three foundations to scale AI agents responsibly: a complete inventory of their agents, enforceable rules governing their behavior, and business context that remains accurate after deployment.

His argument points to a broader shift in enterprise AI. The competitive advantage may not belong to the company with the most powerful model or the largest number of agents. It may belong to the company that can make those agents understand its business, act within defined boundaries and produce consistent results.

The Enterprise AI Scaling Problem Is Already Visible

For many organizations, the first AI deployment looks manageable. A department introduces a copilot, a development team experiments with an automated coding assistant, or an operations group builds an agent to handle repetitive work. Employees review the outputs, correct mistakes and gradually improve the system.

The difficulty emerges when successful experiments spread across the company.

Van de Maele described the difference between managing a handful of agents and operating hundreds or thousands as a fundamental change in complexity. Five agents can be monitored through direct human oversight. A fleet of 100 or 1,000 agents creates a different governance problem, particularly when those systems interact with shared data and production infrastructure.

VentureBeat cited a June survey of 157 enterprise technology leaders. More than 75% of enterprises reportedly encounter a production barrier when moving from a single copilot to approximately 10 or 15 autonomous agents.

The survey also revealed a significant gap between automation and confidence. Some 34% of respondents allowed agents to push code or system changes into production based solely on automated evaluations, without human review. Yet only 5% of all respondents said they fully trusted their evaluations.

These findings illustrate a difficult contradiction. Organizations are granting AI systems greater operational authority while expressing limited confidence in the mechanisms intended to validate their decisions.

The problem is not necessarily a lack of ambition or technical expertise. It is that enterprise deployment introduces requirements that demonstrations and isolated pilots can conceal: accountability, consistent data interpretation, access control, auditability and predictable behavior.

Why Better Models Cannot Solve Everything

The AI industry often presents model improvement as the primary route toward reliability. More capable models should reason better, follow instructions more accurately and make fewer mistakes.

Progress in those areas is important. But enterprise applications introduce a special requirement: correctness must often be defined against a specific business reality.

A consumer chatbot might provide several acceptable answers to a general question. A corporate financial system does not have that flexibility when asked for the company's quarterly revenue. There is a correct figure, a defined reporting period, an approved accounting interpretation and a source of record.

An agent that retrieves the wrong database table, misinterprets a field or combines inconsistent records can produce an answer that sounds convincing but is operationally wrong. Improving the model does not automatically repair those underlying problems.

According to Collibra's survey, 90% of respondents identified the absence of an aligned data foundation as an obstacle to reliable AI results.

That statistic highlights an important distinction. Many enterprises have already invested heavily in data platforms, analytics infrastructure and access to corporate information. The remaining challenge is ensuring that AI systems understand what the information means, which definitions apply and how the organization uses it.

Van de Maele characterizes this as a context and data-governance problem rather than simply a data problem.

In practice, an organization may have a customer table, a revenue dashboard and a reporting system that each use similar terminology differently. Employees may understand those differences through experience, documentation and established business processes. An agent may not.

Without that context, the system can retrieve technically valid information and still reach the wrong conclusion.

The First Requirement Is an Inventory of Every Agent

The first step toward control is knowing what exists.

At a Collibra AI summit, Van de Maele asked attendees whether they knew how many AI agents were operating in production at their companies and what those agents were doing. Most people in the room could not answer either question.

That uncertainty is understandable. Enterprises encourage experimentation because it can reveal useful applications quickly. Employees build agents, teams connect tools and departments develop specialized workflows. Over time, these efforts can spread beyond the visibility of a central technology or governance team.

The result is a growing population of AI systems with different owners, permissions, data sources and operational responsibilities.

An inventory establishes the foundation for managing that population. Businesses need visibility into which agents exist, who is responsible for them, what systems they can access, what actions they can perform and whether they are still required.

This is not merely an administrative exercise. An agent that can summarize public information presents a different risk from one that can access confidential customer records or alter production infrastructure.

Without a reliable inventory, an organization cannot consistently assess those differences, identify duplicated work or determine where oversight is most urgently needed.

Inventory also supports cost management. If businesses cannot identify all the agents running in their environments, they will struggle to understand where model calls, token consumption and overlapping workflows are generating expenses.

The principle is straightforward: companies cannot govern systems they do not know they have.

Deployment Contracts Must Make Governance Enforceable

Knowing which agents exist is necessary, but it does not determine what they should be allowed to do.

The second requirement is a deployment lifecycle that establishes what an agent must satisfy before entering production. Such a process can define permissions, validation requirements, operational limits and the conditions under which a system is approved for use.

However, governance can fail if it becomes an obstacle to delivery.

When reviews are slow, complicated or disconnected from engineering workflows, teams have incentives to bypass them. A process that is theoretically comprehensive but routinely ignored provides little practical protection.

The challenge is therefore to make the approved route efficient enough that teams have a reason to follow it. Governance must operate at the speed of development without abandoning meaningful controls.

The third requirement addresses what happens after an agent passes review. Even a properly registered and approved system can make an unauthorized decision or act on outdated information during execution.

Collibra's proposed answer is an agent contract: a machine-readable set of rules derived from existing governance policies and business context.

These contracts translate organizational requirements into conditions that can be checked when an agent attempts an action. Global contracts establish shared boundaries across the enterprise, while specialized contracts adapt those requirements to particular categories of agents.

As organizations introduce new types of agents, they can extend the governance framework rather than inventing an entirely separate control system for each deployment.

The underlying idea is to move governance from documentation into operational enforcement.

How Guardian Agents Could Control Autonomous Actions

Collibra describes guardian agents as a mechanism for checking an agent's intended action before execution.

If the proposed action violates an applicable contract, the guardian blocks it. If the action satisfies the defined rules, the decision can proceed under those controls.

A critical design choice is that these guardians are not intended to rely on another large language model to decide whether an action is acceptable.

Using one AI model to supervise another may seem attractive, but it introduces additional latency and token costs. It also creates a reliability problem: a model-based supervisor can make inconsistent judgments, including mistakes similar to those made by the agent it is supposed to control.

Collibra instead builds its guardians to evaluate contract rules deterministically. Given the same action and the same relevant conditions, the control is designed to produce the same verdict.

This distinction is important. Language models are useful for interpreting requests, generating plans and working with unstructured information. Deterministic controls are better suited to enforcing explicit requirements that should not change unpredictably from one execution to another.

Consider an agent that prepares a payment, updates a customer record or modifies a production system. The organization may permit the agent to perform certain operations while prohibiting others. A contract-based control can check the proposed operation against defined boundaries before the change occurs.

Such a mechanism does not eliminate every form of AI failure. A guardian can enforce only the rules it has been given, and those rules may be incomplete or poorly designed. It can also be affected by incorrect information about the action or its surrounding context.

Nevertheless, separating flexible AI reasoning from deterministic enforcement offers a practical way to reduce the risks of autonomous execution.

Context Is the Missing Layer Between Data and Decisions

The most consequential part of Collibra's argument concerns context.

An enterprise ontology captures the meaning of business data, the relationships between concepts and the processes governing their use. It helps an AI agent distinguish not only what a field contains, but also what that field represents in the company's operating environment.

This is different from simply connecting a model to more databases or documents. More information does not necessarily produce better decisions if the system cannot distinguish authoritative records from outdated ones, reconcile competing definitions or identify the business process relevant to a question.

An independent context layer can preserve these definitions across changes in models and platforms. Enterprises may adopt new language models, switch data platforms or deploy agents through several different interfaces. The underlying meaning of the business should not need to be recreated from scratch each time.

This is especially relevant as agents spread across Anthropic's Cowork, ChatGPT, Microsoft Copilot, custom agent frameworks and enterprise data platforms. When each environment maintains its own interpretation of business terms, conflicting definitions can emerge across departments.

A context layer offers a way to establish shared meaning across those environments. But it must be maintained. Business definitions change, products evolve, reporting requirements shift and data sources are updated. Context that was accurate at launch can become misleading later.

Research from a frontier-model laboratory, as described in the source report, illustrates the potential importance of this factor. Its in-house analytics agent reportedly answered about 30% of questions correctly without context and about 90% with context. Two months later, its accuracy had fallen to 60%.

Those figures should be understood as results from the specific example reported, not as a universal benchmark for every enterprise agent. Even so, the decline illustrates a critical operational lesson: context is not a one-time configuration task. It requires ongoing maintenance, just as data pipelines do.

The Hidden Economics of AI Agent Fleets

The business case for better context extends beyond accuracy.

Van de Maele noted that advanced models can repeatedly work through a problem before reaching an answer. Every additional pass consumes tokens, adding to the cost of running the system.

When an agent lacks the definitions or background information needed to interpret a request, it may spend additional cycles inferring what the business means. If it starts with reliable context, it can avoid some of that unnecessary reasoning.

Across a large fleet, these inefficiencies can become meaningful. An individual agent may consume only a modest amount of additional compute during one task, but repeated across thousands of workflows, the cost can accumulate.

Human oversight has its own economics. Employees can validate outputs, update agent instructions and recertify systems, but the workload grows as the number of agents increases. Businesses cannot assume that adding people indefinitely is a sustainable answer to autonomous-system risk.

The objective should be to reduce preventable errors and unnecessary work through shared context, reusable governance rules and consistent enforcement.

The source report does not provide quantified savings, pricing details or a return-on-investment calculation for Collibra's approach. Any financial benefit therefore remains dependent on implementation costs, the frequency of agent use, the cost of mistakes and the effectiveness of the controls.

Still, the economic logic is clear: reliable context can improve both the quality of decisions and the efficiency with which those decisions are produced.

Platform or Product. Where the Competitive Advantage Lies

Enterprise AI is increasingly becoming a platform question rather than a contest over individual applications.

A single agent can solve a narrow problem. A platform-level governance and context layer can potentially support many agents across departments, models and operational systems.

That distinction changes how companies should evaluate their investments. A team selecting an agent should consider not only whether it performs a task well, but also whether its business definitions, permissions and controls can be reused elsewhere.

Collibra's position is built around data governance and business context. Its proposed agent contracts and deterministic guardians extend that foundation into the operational environment where agents act.

The broader opportunity is to make enterprise knowledge usable across multiple AI systems without allowing each system to invent its own interpretation of the business.

However, the source does not establish that Collibra's approach has already become an industry standard or that it outperforms every competing governance solution. Its long-term position will depend on implementation quality, integration breadth, ease of use and its ability to keep context accurate as organizations change.

The strategic argument goes beyond one vendor. If access to capable models becomes widely available, model access alone may provide less differentiation than it once did. A company's proprietary knowledge, business processes and institutional understanding can become more important.

The organization that translates those assets into reliable, accessible context may be better positioned to build useful agents than a competitor with similar models but fragmented internal knowledge.

The Risks That Governance Cannot Eliminate

No single control can make autonomous AI completely safe.

An inventory can be incomplete. A deployment review can miss a weakness. A contract can encode the wrong business rule. A guardian can correctly enforce a rule that no longer reflects how the organization operates.

Context can also fragment across teams or drift out of date. If two departments define a critical metric differently, a shared AI system may still produce conflicting answers unless the disagreement is resolved at the organizational level.

There is also a balance between control and flexibility. Rules that are too restrictive can prevent useful automation, while rules that are too permissive may offer little protection. Businesses need clear ownership for exceptions, policy updates and decisions that require human judgment.

Security deserves particular attention. A Harris Poll commissioned by Collibra found that nearly 30% of companies reported discovering agents accessing or surfacing data they did not have permission to see. This finding signals that access governance is not a hypothetical future concern.

Enterprises should therefore treat agent permissions, sensitive-data handling, monitoring and incident response as core deployment requirements. They should also recognize that governance mechanisms are complementary. Deterministic checks can constrain defined actions, but they cannot replace sound security architecture, accurate data or accountable business owners.

What This Means for the Global Enterprise

The implications extend across industries and markets. Companies in Europe and the United States face the same basic operational question: how can they expand AI use without losing control over information, decisions and production systems?

Organizations with complex operations may find the challenge especially visible because their information is distributed across departments, applications and business units. Yet the problem is not limited to large corporations. Any company that allows multiple agents to access shared systems can encounter questions about permissions, consistency and accountability.

The next stage of enterprise AI is likely to reward organizations that connect experimentation to operating discipline. They will need to understand what is deployed, define what each system is permitted to do and maintain the context that makes its outputs meaningful.

That work is less visible than launching a new model or announcing an autonomous workflow. It may also be more durable. Models and applications can be replaced, but the definitions, relationships and governance practices that explain how a business operates are long-term organizational assets.

Practical Strategy for Businesses Scaling AI

Companies should begin with a current inventory of production agents and the systems each can access. They should identify owners, intended purposes, sensitive-data exposure and the consequences of unauthorized actions.

Next, they should establish a deployment process that is proportionate to risk and fast enough for teams to use consistently. High-impact agents should face stronger testing, clearer permissions and appropriate human approval before they can change critical systems.

Third, organizations should invest in a maintained context layer. Business terms, authoritative data sources, reporting definitions and workflow rules should be documented in forms that can be shared across the agent environments the company uses.

Finally, they should evaluate controls continuously. Relevant measures include answer accuracy, unauthorized-action attempts, policy violations, token consumption, review time and the frequency with which context must be corrected. These measurements help businesses determine whether a governance program is improving performance rather than simply adding process.

The objective is not to eliminate human involvement everywhere. It is to direct human attention toward decisions that require judgment while allowing well-defined, lower-risk activities to proceed within enforceable boundaries.

Enterprise AI has reached a point where adding more agents is no longer the only meaningful measure of progress. The real test is whether those agents can operate consistently, securely and economically in the environment where the business actually functions.

The organizations that succeed will not necessarily be those that automate the fastest. They will be those that combine capable models with reliable context, clear contracts and controls that hold up in production. In a market where many companies can access similar AI technology, the ability to turn proprietary knowledge into dependable action may become one of the most important sources of competitive advantage.

The Enterprise AI Bottleneck Is No Longer Just the Model

enterprise AI agents, AI agent governance, AI context layer, data governance, AI agent contracts, deterministic AI controls, enterprise AI scaling, AI hallucination, AI production risks, agent inventory, AI token costs, business data context, autonomous AI systems, AI competitive advantage