The most valuable AI business may not be the one with the smartest model. It may be the one that knows what happened before a model arrived, what it is allowed to do next, and whether its work produced a result.
That sounds less revolutionary than artificial general intelligence. It is also a problem companies can test today. A model can draft an invoice dispute, write code or propose a purchase order. A business still has to know which invoice is disputed, whose approval counts, what changed in the ledger and whether the order was fulfilled. The difference between generating an answer and completing accountable work is where a large part of the next AI market may be contested.
The output trap
Consider the closest thing to a natural experiment in software work. In a September 2026 revision of an NBER working paper, Mert Demirer, Leon Musolff and Liyuan Yang study more than 500,000 GitHub developers and their AI-tool usage. In their matched event-study design, autonomous coding agents are associated with a 240% cumulative increase in commits. The effect narrows to 80% for projects and 30% for releases. The magnitudes are study estimates, not universal predictions. Their important distinction is between activity and a delivered outcome.
A commit is not a release. Nor is a generated compliance memo an accepted audit, a customer-support suggestion a resolved case, or a completed insurance form a paid claim. The distance between those things contains identity, approvals, exceptions, outside systems and human responsibility. AI can speed up a step while leaving the larger process constrained.
Earlier workplace evidence makes the complementary point. A study of 5,179 customer-support agents found that access to a generative-AI assistant raised issues resolved per hour by 14% on average, with larger measured gains among less experienced workers. It shows that AI can produce concrete operational value before broad autonomy. It does not show that every enterprise workflow will enjoy comparable gains.
Four gates, not one
An author-supplied research presentation, AI Opportunity Gates 2026-2035 (10 September 2026), proposes that a useful commercial AI system must clear four gates: model capability, domain data, authority to act and observable buyer economics. That is a framework, not a measured law of innovation. Its attraction is that each gate can be challenged independently.
Capability asks whether a model can perform a bounded task reliably. Access asks whether the system can legally and accurately obtain the records that make the task meaningful. Authority asks who may read, recommend, approve, pay, change state or reverse an action. Economics asks whether the buyer receives faster releases, fewer errors, lower costs or some other accepted outcome—not merely more AI-generated activity.
None of this requires waiting for a machine that can do every cognitive task. It requires choosing a repeatable job and proving an improvement across the entire workflow. For a procurement process, that might mean fewer days from approved requisition to a correctly received order, including exceptions. If the apparent productivity gain disappears when approvals and rework are counted, the opportunity is smaller than the demonstration suggested.
Why the agent itself may be replaceable
The research presentation argues for a persistent layer of operational memory, source provenance, permissions and result tracking that can survive a model change. It calls the opportunity a “governed continuity layer.” The label is less important than the test: if the language model were replaced tomorrow, would the customer still possess a useful, trustworthy record of decisions, rights, integrations and outcomes?
Independent evaluations underline both the opportunity and its limits. METR measures agent capability partly by the duration of tasks that humans normally take to complete at a specified success rate. Its time-horizon work documents rapid progress in bounded software tasks, but cautions against directly translating benchmark trends into dependable real-world work. A longer task horizon is not an operating license. Reliability, integration and accountability remain separate questions.
The US National Institute of Standards and Technology's AI Risk Management Framework likewise calls for mapping the context of AI use, defining human oversight and managing risks through the system lifecycle. It is not proof that a standalone governance product will win. It is evidence that organizations using AI face governance obligations even when the underlying model is impressive.
The strongest objection: the platforms are already building this
There is an obvious threat to the thesis. The governance layer is not an empty market. Microsoft's Copilot Studio already documents environments, connector restrictions, security roles, auditing and approval processes. In 2026 it expanded agent inventory and identity-related controls. Microsoft and other enterprise platform providers can bundle many of the supposed missing components with software customers already use.
That changes the commercial proposition. A new company selling generic “memory plus permissions for agents” may simply be repackaging features of an incumbent. The source presentation itself proposes a kill rule: reject opportunities when most of the promised value can be reproduced using an existing platform at modest integration cost. There is no independent evidence in the presentation that a new horizontal supplier has overcome that test.
The narrower opening is where an organization needs durable, auditable state across systems and vendors that do not share one authority model—and where a specific buyer can prove that solving the mismatch is worth paying for. Even there, contractual data rights, access to incumbent APIs, switching costs and integration overhead may destroy the business case. Owning a diagram of the workflow is not owning its economics.
What should be built before AGI?
Not a universal AI employee. Start with one expensive, frequent, bounded workflow. Measure its present cycle time, error rate, intervention burden and total accepted output. Give AI a limited role first: prepare a recommendation, attach its sources and ask the authorized human to approve it. Expand toward reversible actions only where monitoring and recovery work. Preserve every decision and outcome in a format that is useful independently of the current model.
Three results would strengthen the commercial case: a repeated improvement in the buyer's accepted deliverables; a lawful, stable route to the necessary information and action rights; and an advantage that remains after comparing the product with existing enterprise platforms. Failure on any one should change the investment conclusion.
The most important milestone may therefore arrive before general intelligence: the point at which software can complete a narrow process reliably, prove what it did and remain accountable to the people and institutions authorizing it. That is a smaller promise than AGI. It is also one a buyer can inspect on Monday morning.
Source and status, 11 October 2026 — publication version 1. The four-gate model and governed-continuity thesis are adapted from the author-supplied AI_Opportunity_Gates_2026_2035(1).html, dated 10 September 2026. Its probabilities, opportunity rankings and projections are analyst judgments, not measured forecasts, and are deliberately not reproduced here. Independent anchors include Demirer, Musolff and Yang, Writing Code vs. Shipping Code (NBER, revised September 2026), Brynjolfsson, Li and Raymond, Generative AI at Work (NBER), METR's autonomy research, NIST AI RMF Core and Microsoft's Copilot Studio governance documentation. The commercial thesis is an inference; no new-market revenue, founder success probability or payback estimate is established. Companion evidence review is available in the public investigation. No corrections recorded.