Gartner made it official. Announcing its top strategic technology trends for 2027 in Orlando, the firm named AI context platforms a top-four trend and warned that without a dedicated context layer, organizations face “overprivileged agents, compliance failures, low trust, and excessive token consumption.” The category now has a name and a budget line. What it does not yet have is an operator’s account of what breaks on the way there.

In CDO Magazine’s recent Generative + Agentic AI Trend Report, Murali Kallem explored the context layer: the meaning infrastructure that lets agents operate on governed definitions rather than raw data. It is the right architecture. This article is about the other side of it: what I have watched happen across two decades of enterprise data work when the context layer is missing.

The failures are not random. They follow patterns.

For chief data officers (CDOs), building a reliable context layer starts with five essential components, each one a build, not a buy.

1. Governed definitions: One meaning per concept in a semantic layer an agent can query
2. Machine-readable lineage: Source, freshness, and transformations carried as metadata on every dataset so an agent can judge its trustworthiness
3. Named ownership: A human accountable for every definition and dataset, with authority to correct both
4. Decision rights: Explicit boundaries on what an agent may decide, recommend, or escalate
5. Documented knowledge: The rules, exceptions, and judgment that currently live in people’s heads, captured as data a model can use

The components interlock. Definitions give an agent meaning, lineage vouches for the data behind it, ownership keeps both true, decision rights bound action, and documentation preserves the judgment no dataset holds. Remove any one, and the others degrade.

According to McKinsey, 40% of large organizations are now scaling AI agents. According to VentureBeat Intelligence, 64% of enterprises have already traced a confidently wrong agent answer to problems in their own company data: wrong metric definitions, stale data, missing context. The researchers call these context failures. From what I have seen, they are the dominant failure mode of the agentic era.

Failure Pattern 1: The Confident Disagreement

A merchandising report says “revenue.” Finance says “revenue.” Marketing says “revenue.” Three definitions, one word. A human analyst pauses, asks a colleague, checks the glossary, and resolves it. An agent does not pause. It approximates confidently at machine speed across thousands of decisions.

I have sat in reconciliation meetings where executives spent an hour debating two different revenue figures from separate dashboards, both of which were defensible. Rather than focusing on business decisions, they were left trying to reconcile definitions that had never been aligned in the first place.

The lesson: Semantic consistency is not a data-quality nicety. It is the precondition for agents that can work together. A semantic layer provides the foundation by ensuring each concept has a governed definition that is consistently mapped to physical data. That doesn’t mean every department must use the same definition. Finance and merchandising, for example, may have valid reasons to measure revenue differently. The problem is not that definitions vary but that those differences often go unrecognized. Rather than eliminating them, organizations should ensure each variation has a name, an owner, a stated purpose, and a machine-readable scope. Each workflow should also specify which definition applies. When context is ambiguous, agents should escalate rather than make assumptions. The goal is to establish one governed definition per concept within each context.

Anthropic’s data team has measured what the backbone is worth. Without curated reference material encoding its metric definitions, trusted sources, and known gotchas, Claude’s accuracy on analytics questions topped out at 21%. With that material in place, it ran consistently above 95%.

Definitions are not documentation. They are infrastructure.

Failure Pattern 2: The Silent Corruption

An upstream system changes a field, or a data feed arrives six hours late or stops arriving altogether. The agent downstream notices none of it because its lineage lives in a wiki diagram that no machine can read. It treats outdated information as equally reliable as current data and acts on it.

At one organization I worked with, an upstream team renamed a field over a weekend without triggering an alert. Downstream reports continued producing plausible but inaccurate numbers for weeks before anyone traced the discrepancy to its source. The problem was not the change itself but the lack of visibility into how it affected downstream systems.

The lesson: Data lineage must be maintained as current, queryable metadata associated with every dataset an agent uses, rather than static diagrams. When agents can trace metrics to their sources, they can assess the reliability of the information. Without that visibility, errors can go undetected.

Lineage is not compliance overhead. It is an agent’s sense of direction.

Failure Pattern 3: The Orphaned Error

An agent produces an incorrect output, and someone in the business notices. And then: nothing. No one owns the definition the agent used, so no one owns the correction.

The lesson: Every domain, metric, and agent-facing dataset needs a named owner with SLAs, accountability, and the authority to say no. I have seen citizen-stewardship programs put ownership in the hands of more than 100 domain SMEs and lift stewardship efficiency by 30%. The problem it solved was a queue: a small central team was the bottleneck for thousands of datasets, and requests waited weeks. The program shifted stewardship responsibilities to trained subject-matter experts within each business domain, allowing them to work according to common standards while the central team provided guidance and support. This experience taught me that effective data ownership does not scale by expanding centralized teams. Instead, it requires empowering the people closest to the data with the resources and support they need. This approach strengthened accountability, built trust, and encouraged adoption.

Ownership is not a RACI chart. It is an operating commitment.

Failure Pattern 4: The Boundary Breach

An agent is told to optimize a workflow, and while it performs the task effectively, it continues beyond the point where a human would have stopped. The problem is not malicious intent but a lack of context. Without clearly defined decision rights, the agent cannot distinguish between decisions it can make independently, those that require human input, and those that must be escalated.

The lesson: AI must be tested for what it should not do, not just what it can do. Decision rights put that principle into practice. For every agent workflow, the three questions (decide, recommend, and escalate) need explicit answers before deployment, not after the first breach.

Boundaries are not constraints on AI. They are the conditions that make autonomy safe.

Failure Pattern 5: The Tribal Knowledge Cliff

Every organization relies on institutional knowledge that isn’t always documented, whether it’s an operator who knows why a forecast breaks every third quarter or the rationale behind a policy buried in a years-old email thread. I have seen teams discover, only after an experienced employee leaves, that critical information was never documented. Employees can often fill these gaps by consulting colleagues, but AI agents cannot rely on the same approach. I have called this gap the missing documentation layer.

Even when documentation exists, it can quickly become outdated. Anthropic found that its analytics accuracy declined from roughly 95% to 65% within a month as its reference documentation became outdated.

I call the accumulation of outdated or missing information context debt. In an earlier piece, “AI Exposes More Than You Think,” I described it as the emerging enterprise liability of the agentic era. Data debt slows an organization down. Context debt breaks its AI.

The lesson: Documentation is data. Organizations should capture the meaning, rules, and assumptions behind critical datasets and decisions in a format AI models can find and use.

Documentation is not paperwork. It is the transfer of judgment.

The Practical Steps: A Monday-Morning Audit

You do not need a three-year program to start. Pick one agent pilot and audit its context with five questions, one per failure pattern:

1. Are its key definitions governed, with each variation named, owned, and scoped to a context?
2. Can it query lineage for every dataset it touches?
3. Is every one of those datasets owned by a named human?
4. Are its decision rights (decide, recommend, and escalate) written down?
5. Is the tribal knowledge it depends on documented as data?

Once you’ve identified the gaps, work through them in the following order: 

1. Start with ownership (Question 3): Every other fix stalls without a named human accountable for it
2. Address definitions and lineage (Questions 1 and 2): Focus only on the datasets the pilot touches. This should take days to weeks, not require a lengthy program
3. Capture tribal knowledge (Question 5): Document critical information while the experts are still in the room
4. Establish decision rights (Question 4): Treat these as a go-live gate, documenting them before the agent’s responsibilities expand
5. Address remaining gaps: Treat every remaining “no” as a scope decision, not a stop sign. Narrow what the agent may do to what its context supports, then expand its responsibilities as the context matures

Every “no” represents a missing load-bearing beam in your skyscraper. I argued in “One Cannot Build Skyscrapers in Venice” that AI magnifies patterns. Give an agent clarity, and it amplifies clarity. Give it ambiguity, and it amplifies ambiguity faster than any governance committee can convene.

The context layer is not built in a day, but readiness compounds. Organizations should build the context layer before the agents arrive, especially as adoption continues to increase.