Research
The Constraint Premium
Why the machines will do the work, and someone else will collect the rent.
Abstract
Capability is getting cheap. Permission is not.
Language models that can draft, reconcile and approve are now inputs priced in fractions of a cent. What remains scarce inside a regulated enterprise is the right to act: identity, jurisdiction, an audit trail, revocation, and a cost that can be forecast. The thesis names the value that accrues there the constraint premium — earned not by the most capable system, but by the most governed one.
Insurers, banks and asset managers answer to auditors and regulators. They will not rebuild their systems of record around a technology whose behavior cannot yet be guaranteed; they will connect plain-language AI as a new channel into the records they already keep. Whichever model sits behind that channel can be swapped. The accountability for what enters the record cannot.
The contribution is threefold: a mechanism explaining why the governed layer collects the rent rather than merely surviving; two metrics for the context cost agentic systems leave unmeasured; and an architecture — federation rather than migration — that removes the switching cost which stalls most modernization.
01 — The scissor
Two curves, one bill
The price of a fixed level of capability has collapsed while the cost of frontier capability keeps climbing. Most commentary merges the two into a single claim that AI is getting cheaper. The enterprise bill sits between them, set by which tier of intelligence is pointed at which task.
- Fall in inference price at fixed capability, Nov 2022 – Oct 2024
- 280×
- Stanford AI Index 2025
- Annual rise in the cost of running frontier models on hard benchmarks
- 3–18×
- Gundlach et al., arXiv:2511.23455
- Forecast agent-software spend, 2026
- $206.5bn
- Gartner, May 2026
- Firms abandoning most AI initiatives in 2025, up from 17%
- 42%
- S&P Global Market Intelligence
- Commodity intelligence — price at fixed capability. $20.00 → $0.07 per million tokens, Nov 2022 – Oct 2024 (Stanford AI Index 2025)
- Frontier intelligence — cost of the best available on hard benchmarks, rising 3–18× a year (Gundlach et al., arXiv:2511.23455)
02 — The mechanism
Meaning is re-derived — now at machine speed
A single commercial fact — a renewal, a claim, a disputed charge — crosses ERP, CRM, compliance, billing and reporting. At every hop something reconstructs what it means: which customer, which entitlement, which regulator, who may see it. Agents do not change that logic. They change its frequency, and every reconstruction is billed.
- 01ERPOriginationMeaning re-derived
- 02CRMOpportunityMeaning re-derived
- 03ComplianceJurisdictionMeaning re-derived
- 04BillingSettlementMeaning re-derived
- 05ReportingDisclosure
Meaning defined once is an asset. Meaning re-derived on every execution is a bill.
The claim the programme sets out to measure: the dominant cost term in an agentic enterprise is not the model but context — assembled, retrieved and reconstructed on every execution, and spread across so many components that it never surfaces as a line item. A de-identified, pre-agentic case from a large regulated financial-services deployment shows the shape: a servicing view that took roughly 35 seconds, more than 40% of it spent re-assembling data into display form, with up to 97% of state re-sent from one call to the next.
03 — The hinge
Where the control point belongs
The correct control point is a governed semantic layer: an ontology that binds permissions, lineage, business rules and definitions to enterprise concepts rather than to the applications that touch them. Industry standards — FIBO in financial services, ACORD in insurance — supply the schema each institution customizes.
Governed execution
- 07Governance & trustPermissions, jurisdiction, audit, identity, predictable costRent accrues
- 06Action & orchestrationWorkflows, integrations, decisioning — the governed execution layerRent accrues
- 05Reasoning & intelligenceModels and agents — commoditizing; intelligence is not the moat
The hinge
- 04Semantics & ontologyMeaning defined once and referenced; permissions bound to conceptsRent accrues
Commodity substrate
- 03Compute & processingPipelines, transformation, APIs
- 02Storage & managementLake, warehouse, catalog — federate, don’t migrate
- 01Data originationSource systems, transactions, documents, events
Control-point placements compared
- FailsPermissions at the service accountOver-broad, effectively non-revocable, invisible in audit
- FailsMeaning in the promptUnversioned, untestable, silently divergent across agents
- FailsGovernance at the application boundaryCannot express constraints that span systems
- FailsSpend caps as governanceLimits the cost of a mistake, not the action itself
- HoldsSemantic layer as control pointAuthority bound to the business concept itself
An ordinal judgement of blast radius when the control point is misplaced, not a measurement.
04 — Routing
The cheapest tier that works
If commodity and frontier intelligence diverge this sharply, pointing a frontier model at every task is waste, not sophistication. The discipline is tier selection — optimizing for outcomes rather than tokens — under one permission and business-rule layer that stays constant while models are swapped beneath it.
- Deterministic rulesRepetitive tasks — no model invoked$0.00per task
- Lightweight MLClassification and anomaly detection~$0.01per task
- Small modelsDocument processing and extraction~$0.05per task
- Frontier reasoningGenuinely hard, open-ended problems$2–5per task
05 — Measurement
The metrics nobody keeps
Teams meter tokens consumed; almost none meter meaning re-assembled from source. The programme proposes two first-class measures and a staged, cost-capped plan to produce them. No values are reported yet: Tier 0 runs first.
Context re-derivation frequency
RDF = Σ re-transmitted input tokens ÷ Σ total input tokens
The share of an agent’s input already transmitted earlier in the same task.
Context-assembly cost share
CACS = $ attributable to re-transmitted input ÷ total task $
Priced at what the provider actually charged, cache discounts included.
Reconciled against provider usage APIs to ±2%, or the harness halts. Reported as distributions — median and p90 — because same-task variance of up to 30× is documented.
| Stage | Budget cap | Scope |
|---|---|---|
| Tier 0 | $0 | Public agent trajectories and free-tier dry runs — proof that the metric discriminates. |
| Tier 1 | $360 | A low-cost model under both cache conditions — the minimum publishable result. |
| Tier 2 | $1,360 | A 100-task subset with a frontier validation slice — the strongest version. |
Cumulative caps. Nothing is spent until the previous tier earns it.
06 — Boundaries
What would prove it wrong
A thesis worth holding states its failure condition. This one names its counter-reading and the evidence that currently cuts against it.
The failure condition
If regulated firms begin retiring systems of record for model-native ones — or business comes to be recorded and answered for inside the model’s interface — the rent follows the interface, as aggregation theory predicts, and the thesis is wrong.
The live test: the 2026 partnership placing a frontier model inside a CRM vendor’s permission and business-rule layer, alongside the vendor’s own CRM reasoning model. Vendor-announced; outcomes not yet observable.
What cuts against it
- 9% production adoption of Apache Iceberg against 68% with no current plans — zero-copy federation is a forward-looking claim, not current practice. dbt Labs 2026, n=363, vendor survey.
- 12× more AI projects in production among teams with formal AI governance — correlational; governance may proxy for organizational maturity. Databricks 2026 telemetry.
- Prior claims to the conclusion — Christensen, Spolsky, Gartner, Deloitte, Thompson and API-led connectivity reached the governance layer by other roads. None supplies the mechanism or a way to measure it.
07 — Publication pipeline
One worldview, three disjoint pieces
Each piece asks a different question of a different reader, and no prose is reused across them. Titles are working titles; venues are named once a piece is accepted. Status moves from preparation to submission, acceptance and publication.
| Working title | Question | Core claim | Status |
|---|---|---|---|
| The Constraint PremiumBusiness book proposal · International prize for emerging business authors | Why does value move? | The constraint premium: rent accrues to governed execution, not to the most capable model. | In preparation2026 |
| The semantic layer as control pointFeature article · Peer-reviewed software-engineering publication | Where is the control point — and how do you instrument what it saves? | The semantic layer as control point; context re-derivation frequency and context-assembly cost share proposed as first-class metrics. | In preparationQ4 2026 |
| Agents do the work; enterprise software keeps the rentViewpoint · Computing practitioners’ journal | Where does the business land when machines do the work — and what would show it wrong? | A 2026 partnership between a CRM vendor and a model provider examined as a live test of coexistence rather than replacement. | In preparationQ4 2026 |
| What dominates the bill of an agentic systemPractice article · Practitioner journal | What dominates the bill of an agentic system, measured? | Measured re-derivation and context-cost distributions from the staged measurement programme. | PlannedAfter Tier 1 measurements |
Disclosure
The author was employed by Salesforce (2021–2024) and by Mphasis Silverline, a Salesforce consulting partner (2024–2025). He holds no shares in, and has no current financial relationship with, Salesforce, Anthropic, NVIDIA or Amazon. Measured cases are de-identified; no internal figure is attributed to a named company. Vendor-published and correlational data are flagged as such.
Research
Editorial, speaking and peer-review inquiries
For editors, programme committees and practitioners working on AI governance, agentic cost measurement, or enterprise architecture in regulated industries.