The AI business gates
Capability is only the first test. The commercial case also needs data, authority and buyer economics.
This investigation separates the article's commercial thesis from its evidence anchors. The four gates are an analytical framework; the measured evidence supports narrower claims about output, workflow productivity, governance obligations and incumbent overlap.
- 01Capability
Can a model complete a bounded task reliably?
- 02Data
Can the system lawfully and accurately reach the records that matter?
- 03Authority
Who may read, recommend, approve, pay, change state or reverse an action?
- 04Economics
Did accepted buyer outcomes improve after rework and exceptions?
Output is not the same as accepted work
The strongest software-work anchor reports larger estimated gains for commits than for releases. That distinction is the article's central caution: a model can increase activity without proving end-to-end business completion.
Platforms already own much of the layer
Microsoft and other enterprise platform providers already offer significant agent governance and traceability. Buyers may obtain a large fraction of the proposed value directly from incumbent platforms or internal integrators. Generic agent governance is therefore not a validated startup gap.
Claim audit
| Claim | Evidence class | Primary source | Finding | Limit |
|---|---|---|---|---|
| Coding volume increases more than delivered releases | R2 working-paper analysis | Demirer, Musolff and Yang, NBER w35275, September 2026 revision | Matched design covering more than 500,000 developers: autonomous coding agents are associated with a 240% cumulative commits increase, narrowing to 80% for projects and 30% for releases. | Observational matched study, not randomized proof; developer and repository context, not universal enterprises. |
| AI can boost bounded workflow productivity now | R2 field workplace study | Brynjolfsson, Li and Raymond, NBER w31161 | Study of 5,179 service agents found 14% higher issues resolved per hour on average, with larger measured gains among less experienced workers. | Customer-support context and staggered deployment; not evidence of general autonomy. |
| Autonomous task horizon is progressing, but extrapolation is limited | R2 independent evaluation | METR autonomy research | METR evaluates agent capability by task durations humans normally take to complete at specified success rates and documents progress in controlled software tasks. | Benchmark and scaffolding sensitive; task success does not establish unmonitored enterprise reliability. |
| Oversight, risk mapping and governance are necessary | R1 institutional framework | NIST AI Risk Management Framework Core | The framework calls for mapping AI use context, defining human oversight and managing risks across the system lifecycle. | Voluntary framework; does not establish business viability for a specific product category. |
| Incumbents already provide agent governance capabilities | R1 vendor product documentation | Microsoft Copilot Studio governance documentation | Microsoft documents environments, connector restrictions, security roles, auditing, publishing governance and expanded agent inventory controls. | Vendor self-description; depth of implementation in each customer deployment is not independently audited here. |
| Four gates: capability, data, authority, buyer economics | SOURCE FRAMEWORK | AI_Opportunity_Gates_2026_2035(1).html, 10 September 2026 | The source proposes a coherent selection device across model capability, domain data, authority to act and observable economics. | Author-supplied and unvalidated as a general necessary or sufficient model. |
| Durable cross-model operational state could be a moat | INFERENCE | Original presentation and economic reasoning | A persistent layer of provenance, permissions and outcome memory could remain valuable even if the model is replaced. | Incumbent overlap, proprietary advantage, willingness to pay and retention are not proven. |
Excluded claims and falsification triggers
Excluded from the public narrative
- Worldwide IT expenditure is not treated as TAM or an investable customer market.
- Author probability claims and 2030 readiness estimates are subjective analyst judgments and are not reproduced as facts.
- Claims of unique proprietary state, guaranteed model commoditization and venture returns lack evidence.
- No claim is made that AGI has arrived or that agents can autonomously run all-day business workflows.
Falsification triggers
- Incumbent solution yields at least 80% of customer value without disproportionate integration.
- Workflow end-to-end cycle time or accepted output does not improve after rework.
- Critical data or permission rights are unattainable or too expensive.
- Human intervention and exceptions keep marginal cost high.
- Client declines paid expansion after verified pilot.
Disclosure: No direct interviews, transactions, audited buyer procurement records or proprietary buyer ROI data were available. Market attractiveness remains a hypothesis, not a verified investment conclusion.
Source and status: AU evidence companion, 11 October 2026. Framework adapted from author-supplied AI_Opportunity_Gates_2026_2035(1).html; independent anchors linked in the ledger.