Coverage analysis
Most of it, but not all. Our working estimate puts about four-fifths of decisive AI outputs within reach of a rule, a symbol or a drift bound. Meaning needs an experiment first, and some open judgement should stay with people.
The estimate
Triodian working estimate, judgement-weighted, as a share of modelled decisive AI outputs across the economy. Not yet independently reviewed; ranges and sector weights to be published.
The strata
| Kind of constraint | Share | Regulated-entity example | Operator example | How it is constrained |
|---|---|---|---|---|
| Rules | 34% | Payment limits, claims and credit cut-offs, mandate bands, required fields in a decision | Speed and exclusion-zone limits for a haul truck or a drill | Rules enforcement In build · December 2026 |
| Symbols | 23% | A closed set of actions an agent may take: approve, refer, decline | A closed set of commands: proceed, slow, hold, hand to operator | Grammar and command sets In build · December 2026 |
| Hybrid and drift | 24% | A credit, claims or investment book drifting toward concentration while each decision passes | A fleet drifting toward more near-misses while every move passes | Aggregate drift bounds In build · mid-2027 |
| Meaning | 10% | Whether a message amounts to advice, or a hardship or claims call is fair | Whether a recommendation respects a rule of engagement | Semantic check After the experiment |
| Hard to constrain | 9% | Open judgement with no policy that can be declared in advance | Open judgement with no policy that can be declared in advance | Stays with people |
The revision that matters
A nine-digit string is or is not a tax file number depending on what it refers to. "I'll waive the fee" binds the principal only in context. These properties are more than syntax but less than full meaning. Once a span is classified, comparators, state machines and bounds over those classes are deterministic.
The one probabilistic element, the classifier, sits outside the verdict path, in the same proposer and enforcer separation the architecture already uses. A session ledger that fires on a prohibited combination of individually permissible spans closes the multi-turn gap that per-decision checks cannot see.
By sector
| Sector | Where enforcement binds | Why |
|---|---|---|
| Finance and insurance | Rules and drift, with a symbolic layer for commitments and disclosures | Transaction processing, reporting and surveillance are rule- or drift-governable; underwriting and advice have a semantic core. |
| Professional and business services | Drift and meaning, with a symbolic layer for commitments | Legal, consulting and design are semantic; accounting, payroll and administration are rule-governable. |
| Health, education and social assistance | Drift and meaning, with a symbolic layer for doses and disclosures | Billing, coding and clinical protocols are rule- or drift-governable; diagnosis, teaching and care are semantic. |
| Information | Drift, with a symbolic layer for typed content | Content platforms and AI services are drift-governable at the boundary; editorial and creative work is semantic. |
| Manufacturing, utilities and transport | Rules and drift | Process control, quality control, routing and safety envelopes are rule- or drift-governable; design and last-mile judgement stay semantic. |
| Government | Mixed | Benefits, tax and procurement rules are rule-governable; policy, adjudication and management are semantic. |
The hard ceiling
Constrained decoding provably guarantees a schema is satisfied. A session ledger provably blocks a prohibited combination. Neither guarantees the content is true, appropriate or intended. That boundary is not a limitation being worked around; it defines what the layer is for.
This measures where enforcement can attach, not whether AI can do the work or whether it should.
The boundaries are contested, and the figures are order-of-magnitude estimates, not survey results.
The meaning tier loads only if its pre-registered experiment passes. The rules tier does not depend on it.
Get in touch
We will walk through the strata, the assumptions and where we think the estimate is weakest.
Book a 30-minute conversation →