The semantic layer is a bottleneck wearing a solution’s badge

For a decade, the prescription for inconsistent metrics was the same: build a semantic layer. Define everything once, centrally, correctly, and downstream chaos ends. It was good advice for a slower world where definitions changed little and humans were the only ones asking questions.

That world is gone. And agents have joined the query path: consuming definitions at machine speed, and generating new tables, models, and metrics faster than any committee can canonize them.

Here’s a quick breakdown of why the semantic layer as we know it falls short for AI, and what to do differently.

Act 1 — An honest history

The semantic layer solved a real problem, and there is a lot to praise here: single source of definitions, metric consistency, the end of "whose revenue number is this."

But the rise of the semantic layer buried three assumptions:

• Definitions change slower than humans encode them;
• Queries come from people who can be trained;
• The set of things needing definition is finite and enumerable.

Each was true in 2015, wobbly by 2022, and false the day agents joined the query path. And that’s before you consider how LLMs made semantic-layer syntax largely irrelevant.

Act II — Best practice becomes a bottleneck

A semantic layer maintained by human hands has become the slowest component in an accelerating system. When the business changes logic weekly and the layer updates monthly: the layer isn't the source of truth but a source of confident staleness.

And gents make it worse twice over:

• First as consumers: an agent applies a stale definition ten thousand times before a human would have applied it once.
• Second as producers: agent-generated models and metrics multiply the surface needing definition faster than any manual process can cover it.

Imagine two certified definitions of the same metric. Both blessed, both stale in different ways. The semantic layer didn't prevent the divergence; it laminated it.

Semantic layers aren’t wrong. But maintenance by hand is now load-bearing. Hand speed is now the system's ceiling.

If your path from AI POC to production depends on a hand-maintained semantic layer, you’re building a human-speed dependency into a machine-speed system. The more you invest in manually curating that layer, the further you may be from AI at production scale.

Act III — Meaning without the bottleneck

The goal of the semantic layer was never "a layer humans maintain." It was "people reasoning from correct, current logic." The layer was the mechanism. It's just that the mechanism aged out.

This certainly does not mean abandoning meaning. Instead, semantics should be captured and maintained through automation as the business changes, rather than through the heroics of manually legislating them.

Semantics are created all the time, everywhere, not in one curated place. And the correct logic can be derived from actual lineage, actual usage, actual ownership, continuously, with humans enriching and certifying rather than typing definitions into a committee's backlog.

There’s still room for humans in the loop. Just 2–3 orders of magnitude less. Humans own judgment calls, institutional knowledge, certification of what's AI-ready, guardrails management.

Automation is what carries the cataloging, people carry the meaning.

Epilogue

The semantic layer asked the business to slow down to human documentation speed. But in an AI-first enterprise, manually curated semantics are a scaling ceiling, not a foundation. The winners will be the companies that can automate, scope, and govern context and meaning as fast as they change.

Don't you think?