The Bounded
AI Stack
A Bounded AI Stack is a set of AI systems deployed inside an explicit control boundary: every AI output is constrained to verified, provenance-labelled source material, every claim the system can make is traceable to a dated source, and anything outside the boundary is refused rather than guessed.
Origin of the term
The term was coined by Jason T Wade, founder of BackTier, in the 2026 field study AI in Deathcare: The Bounded AI Stack. That study mapped where AI can safely operate in cemetery and funeral-home recordkeeping — an industry where a wrong date or a misattributed record causes real harm — and named the architecture that keeps AI inside verifiable limits.
The idea generalizes beyond deathcare: any organization deploying AI on records that matter — legal, medical, financial, historical — needs the boundary designed before the tools are deployed, not discovered after an error.
The six layers
Each layer gets its own boundary.
- 1. Records and data entry. AI may transcribe and normalize records, but a human verifies against the source document before anything is committed.
- 2. Retrieval and search. AI may find and rank records, but it returns the record — it does not summarize away the source.
- 3. Drafting and communication. AI may draft from verified records only, with every factual claim traceable to one.
- 4. Public-facing answers. AI answers only from the provenance-labelled corpus and refuses questions outside it instead of guessing.
- 5. Analysis and reporting. AI may identify patterns, but causal claims are labelled as inference, never presented as fact.
- 6. Autonomous action. The tightest boundary: AI does not act on records — sending, changing, publishing — without human approval.
The rule of thumb: the further a layer is from verified source records, the tighter the boundary must be.
Why provenance is the core
An AI system that cannot show where a claim came from cannot be audited. In a Bounded AI Stack every claim carries a provenance label — a dated, inspectable source — so a wrong answer can be traced to the exact record that produced it and corrected at the source.
This is the same standard BackTier applies to its own public claims: every field study and experiment on this site labels each statement FACT or INFERENCE and keeps the verbatim evidence in a dated file. See BackTier Experiments for the running record.
Questions
Frequently asked questions
Related
- AI in Deathcare: The Bounded AI Stack — the field study that introduced the term
- Entity Engineering — the discipline of making AI systems recognize and cite a brand correctly
- BackTier Experiments — live AI-visibility tests with FACT/INFERENCE labels