Jay Limburn is Chief Product Officer at Ataccama, building AI-powered data products that deliver trust and business impact.getty​Semantic layers have become the architecture of choice for enterprises serious about AI, giving agents a shared foundation of business definitions, governed metrics and consistent logic to reason from. That investment is sound. But as AI moves from generating recommendations to executing decisions, semantic architecture alone can no longer answer the more important question: Has the data behind those decisions actually earned the right to inform them?The push toward the semantic layer is not incidental. AI agents need shared business definitions to operate coherently across systems. Without them, the cracks show quickly. When a model reasons about revenue, that concept needs to carry the same meaning whether the agent pulls from a data warehouse, queries a CRM or interprets a downstream report. ​Because of this, the major data platforms are now building that semantic capability natively rather than leaving it to each tool to interpret. The natural next step is to certify the data populating those definitions to the same standard as the definitions themselves, because that determines whether organizations can trust the systems built on top of them to act.Agents cannot see everything from inside a semantic view.​A well-constructed semantic model describes relationships, definitions and intended usage with precision, yet says nothing about whether the underlying data is current, complete or fit for the decision it is about to inform. Consider a credit risk agent inside a major financial institution that sits on a well-constructed semantic layer where exposure means the same thing across every system it touches, with metrics governed end to end. The agent can query a semantic view, calculate a risk position and act. What it cannot see is that teams last reconciled the counterparty records three days ago, that a batch job failed silently over the weekend and that two accounts it treated as separate entities have sat on a remediation list as duplicates for six weeks.The semantic layer did its job perfectly. The data beneath it had not earned the trust the agent placed in it. In a human-reviewed workflow, someone will likely spot the discrepancy before the report goes out. Automated execution removes that feedback loop, and the agent acts at a speed that makes post-hoc correction expensive. Whether data has been certified for the purpose it now serves must be addressed at the same architectural level as the semantic layer itself.Alerts fire after agents act, and that sequence is the problem.Monitoring pipelines, tracking freshness and alerting on anomalies all matter. In agentic environments, though, the interval between anomaly and action compresses to near zero. By the time an alert fires, an agent may have already acted on the affected data. Observability surfaces deviation without preventing the wrong action from occurring, and those become meaningfully different things once AI systems move from informing decisions to making them.Organizations need a quantifiable signal that travels with data the way lineage metadata does and reaches agents before they act. That means enforcing quality standards at ingestion rather than detecting problems in transit, connecting anomalies directly to the business processes they disrupt and giving agents governance provenance that shows who owns a dataset, when teams last certified it and whether the business cleared it for the purpose at hand.There are two layers but one architectural requirement.Semantic layers and trust layers are not competing architectures. They address different problems at different levels, and assuming one substitutes for the other leaves a gap that surfaces only once agents act. The semantic layer resolves meaning through shared metrics, governed definitions and consistent business logic. The trust layer resolves fitness through quality certification, lineage traceability, freshness signals and provenance. Both need to be readable by the systems consuming them, because AI agents do not pause for manual validation and need those signals embedded in the environment rather than surfaced in a dashboard that someone checks the following morning.​Building the trust layer is an operating decision before it is a technology one.​In my work building data trust platforms, I've found that treating trust as infrastructure starts with ownership rather than tooling. By assigning clear accountability for certification, data owners can define what fit for purpose means in their domain, platform engineering can embed checks where data enters the environment and governance can set the standard that makes a trust signal mean the same thing everywhere. When no one owns that decision, certification decays quietly, and the signal stops being reliable long before anyone notices.The process comes with challenges, though, as moving quality enforcement upstream takes engineering effort that competes with delivery deadlines. Stricter gates can slow pipelines that teams have tuned for speed.​Fitness is also contextual, since the same dataset can be trustworthy for one decision and unfit for another, which makes a single universal score harder to design than it first appears. Most organizations underestimate this and treat trust as a one-time cleanup rather than a continuous discipline. Organizations will also struggle if they build trust signals that nothing downstream actually reads. A certification state only has value when the agents and applications consuming the data check it before acting. The practical starting point, therefore, is to work backward from a high-stakes agent use case, identify the data it depends on and make trust machine-readable and queryable at that point of consumption. Data contracts between producers and consumers, paired with a regular certification cadence, turn that from a project into an operating rhythm.Semantic consistency makes AI reliable, and fitness for action makes that reliability operational. Treating both as infrastructure, with a certification state agents can query before they act, rather than a monitoring alert they receive afterward, is the discipline worth building.​​​​Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?