Deploying AI capabilities without AI governance is risky- a topic that we discussed at last week’s Coffee With Digital Trailblazers on the hidden cost of confident AI. Felix Van de Maele, founder and CEO of Collibra, unpacked the issues driving the hallucination tax and shared how enterprises should implement AI governance beyond documenting and communicating policies. 

What is the hallucination tax

I’ve written about AI debt, AI cost debt, and data debt. So I asked Felix to explain the hallucination tax.

“The hallucination tax is the cost of the agents getting the task, the decision, or the outcome wrong,” says Felix. “And if we don’t trust what the agents are doing, we will have to babysit them, continuously validate, review, and double-check. The more independent and autonomous AI becomes, the more we have to trust what those agents do. They’re only valuable if we can trust the outcomes they’re going to drive.”

The hallucination tax captures several different AI risks, including data, context, and costs, under one term and quantifies its impact.

Derrick A. Butts, enterprise CXO AI cyber resiliency advisor, CISO, and founder of Continuums Strategy, added. “The danger isn’t the AI getting things wrong — humans get things wrong too. The danger is AI getting things wrong with absolute confidence. It creates a false sense of certainty.”

Examples of what increases the hallucination tax

Let’s unpack the hallucination tax through three examples:

1. Answering a question using the wrong data

Every ERP, CRM, and HRMS is filled with dates, dollars, and different views of prospects, accounts, and customers. Without data dictionaries or a semantic layer, agents may pull out the most accessible data instead of the valid data given the user’s question.

Felix shares, “If we ask our agent how many customers did we add last quarter, there are lots of different ways to count what makes a customer. If we don’t tell the agent that, the number it comes back with will be completely meaningless.”

2. Inflating compute costs

Larger document stores without standardized metadata make it harder to govern unstructured data. Making AI agents scan all these documents without any structured guidelines can also contribute to the hallucination tax through wasted tokens, higher costs, and longer response times. 

 “Many organizations have tens, if not hundreds of thousands of documents in SharePoint. An agent can parse all those documents to answer a question — but imagine how expensive that question would be,” says Felix. “A more cost-effective way is to use a different agent to parse and structure them first, so the agent looks at 100 documents, not 100,000.”

3. Drifting models and decaying data

Felix’s article on why agents fail shares several other examples of the hallucination tax. Data is dynamic: documents lose value over time, data gets updated, and models drift. Organizations that aren’t monitoring their AI agents or capturing feedback will miss the early signs that they’re adrift and contributing to the hallucination tax – or worse.

“Once we’ve defined the guardrails, the policies applicable to these agents, how do we then enforce that in production?” asks Felix. “How do we make sure the agent actually behaves the way we’ve defined it can behave?”

Implementing context and AI governance

During the episode, we reviewed details on implementing guardrails, semantic layers, and context layers.

Felix recommends developing a unified AI registry of AI use cases – “the combination of the model, the data going to be used in that model, and what to solve for, the intent, which plays a huge role in what risk the AI use case brings and how it applies to policies.”

The registry, paired with an AI command center, provides several ways to reduce the hallucination tax by implementing AI governance based on defined policies. Implementing AI governance includes

1. Grounding AI agents in shared semantics and terminology to improve accuracy, reduce costs, and ensure there’s consistency in how data is used in decision-making.
2. Managing the context layer’s lifecycle by monitoring for decay and drift.
3. Governing access and policies across multiple platforms, including AI models, data management platforms, SaaS, and wherever AI agents are deployed.
4. Establishing a governed deployment cycle for AI agents to ensure stakeholder alignment, ownership accountability, and policy compliance.
5. Enforcing AI agent runtime policies, governing inputs and outputs, and building feedback loops that improve accuracy.

“Governance often gets a bad rep because it feels like it slows things down and it’s bureaucratic,” says Felix. “Governance done well can be a force function to accelerate — to create clarity, predictability, and clear guardrails for everyone to operate within.”

If you missed it, watch the full episode on the hidden cost of confident AI, and attend the upcoming webinar I’ll be speaking at on governing unstructured data for AI agents.

This article is brought to you by Collibra.

The views and opinions expressed herein are those of the author and do not necessarily represent the views and opinions of Collibra.