Trust and quality notes
- Last updated
- September 27, 2026
The most visible AI products are conversational and polished. Tina He argues that some of the most valuable businesses after automation may look much less exciting. They will be the systems that move data, route tasks, enforce rules, and connect machine decisions to real-world outcomes.
In her original statement, He describes a future where agents can behave as customers, make rapid economic decisions, and communicate directly with software.[1] In that environment, attractive screens matter less to the machine. Reliable infrastructure, specific workflows, security, and compliance matter more.
Her thesis is about which businesses may win. It also shows how human priorities change when machines become active participants in commerce.
Agents behave differently from human customers
He opens with a hypothetical customer relationship management contract canceled at 2 a.m. The reason is an agent acting for a sales representative, comparing the annual cost with the value received and changing the service without a meeting or negotiation.[1]
The example supports her claim that agents are not loyal in the human sense. They can evaluate numbers quickly and make a change whenever it appears rational for the business. Personal relationships, habit, or the inconvenience of scheduling a conversation may carry less weight.
This is a forecast, not a reported event. He uses it to illustrate how machine customers could alter familiar commercial behavior. If software can continually evaluate value and switch services, a vendor may have less protection from inertia.
My interpretation is that companies serving agents will need to make value and performance more legible to machines. A relationship cannot rely only on a pleasant interface or a renewal conversation. The underlying service has to deliver a result that can survive repeated evaluation.
Software may have no human users
He distinguishes software made for people from headless architecture built for machine-to-machine communication. Some companies, she argues, may have agents as their only users.[1]
My interpretation is that a human user needs navigation, explanation, and a visual path through a product, while an agent can interact through structured interfaces without seeing a screen. This does not make design irrelevant, but it changes what is being designed. The focus moves toward clear contracts between systems, dependable access to data, predictable actions, and error handling.
The source specifically notes that an agent does not care whether software looks good or feels easy for a person to use. That claim applies to machine interaction, not to every human affected by the system. People may still choose the service, define policies, investigate failures, and remain accountable for outcomes. This distinction is my interpretation of the human role around the headless systems He describes.
The practical implication is to avoid assuming that every product needs a conventional application interface. If the direct user is another system, reliable communication can be more important than visual polish.
Boring systems carry consequential work
He also expects strength in domains where mistakes are costly and moving quickly is not the only priority. She names regulatory approval, banking, and compliance systems, including the secure delivery of wire transfers.[1]
These businesses do not win by being the most sophisticated in appearance. They compete on efficient, dependable execution of constrained work. They know the rules, enforce them, and connect an instruction to an outcome that must occur correctly.
As models converge in capability and make similar decisions, He expects competition to shift away from possessing the best model and toward the infrastructure that turns decisions into action. She names task routing, data access, workflow orchestration, and rule enforcement as essential systems.[1]
These functions can sound ordinary, which is central to her point. They provide the control layer between a general model and a specific business process. Without them, an intelligent recommendation remains separate from the permissions, records, sequence, and rules required to do real work.
Specific use cases beat broad claims
He argues that companies serving specific customers with specific use cases will win over broad agentic use cases.[1] The source does not say general models will be unimportant. It says value can collect in the systems that understand a particular context and reliably complete its required work.
Specificity makes the constraints visible. A task-routing system needs to know which work goes to whom. A data-access layer needs rules about which information may be read or changed. Workflow orchestration needs a sequence and a response when a step fails. Rule enforcement needs a source of authority and a record of what happened.
My interpretation is that a narrow workflow can be more defensible than a broad promise because it accumulates operational knowledge about exceptions, permissions, and outcomes. This is not a claim that every narrow product succeeds. It is an explanation of why He assigns value to infrastructure tied to concrete use cases.
For a team considering automation, this suggests starting with the connection between a decision and an outcome. Identify what must happen after the model proposes an action. Then identify the systems, rules, and approvals required to carry it out safely.
Infrastructure becomes a toll road
He compares these businesses to toll roads. A customer can avoid paying by building a bridge, but in regulated or compliance-heavy work that alternative could take years and cost hundreds of millions.[1]
The metaphor explains the source of durable value. Infrastructure sits on a necessary path. It is difficult to replace not because it is glamorous, but because recreating its accumulated compliance and operational capability is expensive.
This remains He’s forward-looking business thesis. The source does not provide evidence that all such infrastructure will become dominant or that every model will converge. It proposes that less visible layers can capture value as AI capabilities become more common.
My interpretation is that the human lesson is equally practical. Automation still needs institutions that define permission, safety, and accountability. Agents may act quickly and communicate directly, but real-world systems cannot rely on intelligence alone. They need rails that determine what may happen, preserve records, and protect consequential processes.
That is why boring infrastructure can win. It does not compete for attention. It makes automated action dependable enough to matter.
If you want to identify the rules and infrastructure one repeated process needs before it is automated, review it with Agentic Workers.