Semantic debt is real, and it compounds: YAML captures a definition once. It has no mechanism for detecting drift, resolving conflicting interpretations, or telling your AI agents which version to apply in context. And it never will.
version: 2
metrics:
- name: revenue
label: Revenue
description: "Total revenue" ← which revenue?
type: simple
type_params:
measure: total_revenue
# TODO: align with Finance on ARR definition
# see Slack thread from March, unresolved
# Sales uses contracted value
# Finance uses IFRS recognised
# BI shows ARR +24%, ERP shows +12%
# both technically correct
- name: churn_rate
# conflicts with definition in
# analytics/churn_v2.yml (deprecated?)
This file was accurate 4 months ago. Two unresolved conflicts. Zero drift detection. Your AI agents are reading this right now.
// The analogy that explains it
Yahoo built a directory of the web. Humans curated it. It worked, until the web grew faster than any editorial team could maintain. Google didn't hire more editors. They changed the model.
Enterprise semantic layers are at exactly the same inflection point.
editors, replaced by PageRank
Humans write definitions. Humans review PRs. Humans update YAML files. The business changes. The files don't know.
signals from real usage, continuously
Definitions evolve from how the business actually uses data. Conflicts surface automatically. A lightweight approval process (not a PR queue) confirms changes.
// The real cost
These are the numbers data leaders give us when we ask what the Git/YAML setup actually costs. Not the dbt Cloud licence. The hidden tax.
Resolving "which number is right" across teams. Manual reconciliation that doesn't compound, doesn't scale, and disappears the moment the person who knows leaves.
Business user spots an issue → Slack → ticket → engineer writes YAML → PR → review → merge → deploy. The business moves in hours. Your semantic layer moves in days.
Tooling (~$80k–$150k) + engineering time on semantic upkeep (~$110k–$180k) + business logic stuck in Git queue (~$20k–$50k) + LLM token costs (~$120k+ and rising). And 60–80% of those token calls never needed to hit the model at all.
// How LazyFox works
LazyFox connects read-only to your existing systems (Redshift, Databricks, MongoDB, Salesforce) on top of your current stack. No ETL. No migration. No new interface for your team.
LazyFox connects to your existing data sources without moving a single byte. For a known connector type, you're live within hours. Your current dbt models and LookML definitions are imported as the starting point, not replaced.
The LLM runs once at setup to enrich your schema, identifying active columns, mapping relationships across systems, flagging garbage fields. Token cost is front-loaded here. Not per query. Ever.
A lightweight semantic programme team reviews and approves definitions in plain language, no YAML, no PRs. Conflicts are surfaced automatically before any definition is saved. Multiple contextual versions co-exist: Finance's ARR and Sales' ARR, both governed.
Every AI tool you already use (Claude, Gemini, Genie, Dot, Slack AI) queries through LazyFox. They return consistent, governed answers. Your people never log in. They just get the right number.
// Where Git/YAML ends and LazyFox begins
Detect when a definition has drifted
Git today: file unchanged since March. Business changed in April.YAML has no awareness of the gap between what it says and how the business is actually using the data.
LazyFox continuously monitors query patterns and usage signals across all connected tools. When a definition diverges from real usage, it surfaces a conflict, automatically, before it propagates into AI outputs.
Hold two legitimate truths simultaneously
Git today: one file, one winner. The other team works around it.PR-based governance forces a single canonical definition. Finance's ARR and Sales' ARR cannot co-exist in a YAML file.
LazyFox governs multiple contextual versions of the same metric simultaneously. Finance gets their definition. Sales gets theirs. Every AI tool applies the right one based on who is asking and why.
Let business users participate in governance
Git today: can't open a PR → files a Slack message → hopes an engineer acts.If you don't know YAML, you don't exist in the governance workflow. The people who understand the business have no path to change the definitions that govern it.
LazyFox's governance interface accepts natural language. A Category Manager, Finance BP, or Sales Ops lead can flag and propose definition changes directly, confirmed by a named programme owner, not an engineering backlog.
Govern meaning outside the warehouse
Git today: only covers data that's been migrated. Salesforce, MongoDB, operational sources: ungoverned.dbt Semantic Layer and Databricks Metric Views only apply to data inside the warehouse. The 80%+ that hasn't migrated is invisible to the governance layer.
LazyFox connects read-only to every source, regardless of migration status. A definition of ARR governs queries against Salesforce, Redshift, and MongoDB simultaneously. Migration is not a prerequisite.
Survive a warehouse migration
Git today: definitions live in dbt models tied to Redshift/Snowflake configs. Migration = rebuild from scratch.Every warehouse migration forces a semantic layer rebuild. The definitions move with the warehouse, and usually get left behind.
LazyFox stores definitions independently of any warehouse. Redshift → Databricks carries zero semantic rework. You arrive at the new warehouse with governance already in place.
Feed every AI tool from one governed source
Git today: definitions copy-pasted into Genie, Claude, Slack AI separately, diverging immediately.Without a headless governance layer, every AI tool you add requires manually copying the current definition into it, creating n separate copies that drift independently.
Every AI tool queries through one LazyFox layer via MCP or API. One change propagates everywhere, instantly. No copy-paste. No version skew between tools.
"Claude might be capable of the task, but do you have the right data infrastructure?"
Peter McCrory, Head of Economics · Anthropic · February 2026
If your semantic layer lives in a Git repo and you're about to scale AI agents on top of it, we'd like to show you what the next layer looks like.
Book a Demo →No pitch deck. A 30-minute conversation. We'll tell you if you need us.