The entrance has shifted: users no longer open an app; they tell an AI what they need

English concept: Appless World · Intent Routing

Introduction: AI is moving from a “tool” toward “an extension of the person,” and the entrance is shifting

For the past twenty years, our relationship with digital services has been called “a person looking for an app”: tap the icon, walk through the pages, find the button, finish the task. AI, too, has mostly been an efficiency plugin attached to a single application — it knows how to do the thing, but it does not know who you are. Now two lines of evolution are accelerating at once: one extends outward to new terminals — smart glasses, a car system, voice earphones, a home screen, and wearables — and lets the understanding of intent and the execution of an action happen close to the body; the other converges inward into an “AI-native” task hub, aggregating actions that used to be scattered across many apps into a work chain, and letting an agent act as a proxy for the tap, the check, and the delivery. The app has not disappeared, but it has quietly receded into a “library of capabilities.” The default interaction at the front is being rewritten as “say the purpose clearly → receive the result.” The five sections below each say clearly “what is happening,” “why we have arrived here,” and “what it truly means for consumers and for the organization.”

NO° 01 | A diversification of ends: new devices step “opening an app” back into the background

As multimodal sensing and on-device reasoning gradually mature, many actions that used to require unlocking a screen and switching apps are beginning to be finished naturally in the environment. In the car you say “help me expense this invoice”; the glasses camera first does OCR and extracts the key fields; the earphone microphone captures your spoken supplement; a small on-device AI model assembles the two into a structured form, and completes a first round of checking locally; the system then, according to permissions, sends the ticket into the company’s expense flow. You ask the way facing a shop sign at the entrance of an alley; the glasses’ vision model understands the scene and the landmark, and then casts the best walking route into a corner of the field of view. In a meeting room, a low-latency network together with a local cache lets a voice summary, a list of decisions, and a change of schedule “land as you speak,” and the fingers no longer have to walk through a multi-level menu. This kind of experience has become feasible because on-device compute can handle a basic understanding of voice, image, and context; sensors turn intent into a parsable signal; a short-range connection reduces waiting; and, most important, people’s opportunity cost of attention is becoming more and more severe. Whoever can deliver in zero steps wins the home ground.

For consumers, interaction moves from “finding a function” to “stating a purpose,” especially in scenes that are “not suitable for staring at a screen,” such as driving, walking, and housework. But the question of trust also rises at the same time: the device seems “always listening, always watching” — whether it can be switched off, when it goes to the cloud, what is computed locally, and whether refusal and deletion can be traced, will directly affect purchase and retention. For enterprises, the rules of distribution are rewritten: the marginal value of a first-screen position and of install volume falls; whether you can be “called by default” by an on-device AI is the new exposure. Research-and-development also has to break the product into atomic capabilities (recognize, parse, query, submit, sign, receipt), register and be scored with the device OS in the form of a plugin or an API, and move from listing in an app store to admission and optimization in a “catalog of capabilities.”

NO° 02 | AI-native apps become a super console: one entrance, and the work is done across apps

In a world where the phone is still the home ground, the new entrance is not yet another app. It is “an app that will help you use other apps.” You only need to describe, in natural language, the purpose and the constraints — time, budget, preference, compliance, and permission — and an agent can automatically invoke the browser, the calendar, payment, mail, a cloud drive, and a mini-program, and compose the fragmentary actions into a task graph: after a ticket is booked, expense it automatically; after an approval, archive it automatically; after research, automatically produce a draft and a media plan. This can be done because the function-calling and tool orchestration of large models have become reliable, and because the major mobile ecologies have also provided bridges that a machine can dispatch, such as deep links, intent invocation, and automatic form-filling; and because users have a real need for “express it once, have the whole thing handled.” For consumers, the upgrade of experience is not only a few fewer taps. It is also that historical preferences, account assets, and authorized data can be kept ready, and repeated input is avoided. But transparency must also rise at the same time, or it is hard to be at ease handing the whole matter to the system — you need to see the progress, be able to interrupt or revoke at any time, and know who did what at every step. For enterprises, the master control of service aggregation is outsourced; the process memory that used to be held by their own app will be diluted by an AI-native master console. To respond, you have to actively list key actions on the mainstream “action marketplace / plugin system,” and, in a way that is observable, rollback-able, and auditable, compete to become a segment of the default path.

NO° 03 | An Agent API replaces the user’s tap: the interaction of person and service is “agentized”

Without a contract, an agent is only a “proxy mouse.” With a contract, AI can get things done like a reliable colleague. An enterprise needs to register executable actions as a capability contract an agent can call, and clearly define the name of the action, the required parameters, the prior validation, the boundary of permission, rollback on failure, the callback format, and the observable metrics. When an agent receives an intent, it no longer blindly tries the UI. It matches a capability, fills in the parameters, executes, and reports status. The user’s task moves from tapping a string of screens to approving an operation once. The value of this pattern runs both ways: the AI end obtains a stable, controllable, rollback-able path of execution; the enterprise end can write risk, responsibility, and compliance into the boundary of the interface, and close the loop of logs and audit to ensure traceability. For consumers, the burden of operation falls markedly, but explainability and traceability must be in place — what I authorized, which steps were taken, what can be revoked, and where an error can be retried. For the organization, the form of the product will move from “pages + a process map” to “an Intent / Action Schema + a state machine,” and drive an update of roles and of measurement: a capability product manager is responsible for breaking the business into a contract of behavior; semantic operations are responsible for maintaining prompts and task templates; the platform team commits an SLA outward on “the success rate of being called, first-pass rate, latency, and the rate at which errors can be explained.”

NO° 04 | From “publishing a page” to “planting a semantic seed”: cite-ability matters more than PV

The first principle of on-device AI and of a mobile agent is to lower uncertainty. This makes content that “can be retrieved, can be verified, can be sliced” the preferred source. An enterprise needs to recast FAQs, prices and terms, SOPs, cases, and manuals into knowledge chunks usable by RAG, add a source, a timestamp, and a version to each chunk, and allow a canary update, a rollback, and a differential comparison. In this way, an answer card can be “read at a glance,” and can also, when cited, mark responsibility and freshness; and when AI assembles a plan, it can also, in a consistent standard, extract first-hand “official” material. For consumers, the number of jumps falls, an answer is received faster, and they can see the source, the date of update, and the party responsible, so a sense of trust rises with it. But for brand recognition, exposure at the front will be diluted; consumers care more about “whether the answer is reliable” than about “whose page is designed beautifully.” For enterprises, content will truly be “turned into an asset”: it is not writing a few more blog posts, but turning content written for the human eye into knowledge fed to AI. The logic of brand exposure is also rewritten; “being able to be trusted” becomes a new place in the mind. The operations of content have to be versioned like an application, and have testing, rollback, and audit.

NO° 05 | A rearrangement of KPIs and of the organization: from “install volume” to “the rate of being dispatched”

When the entrance migrates from an icon to semantic routing, the system of measurement naturally migrates with it. The marginal value of “install volume, first-screen position, PV, CTR” falls; “the number of calls by an on-device / agent, first-success rate, average time to completion, the number of follow-up turns, the rate at which errors can be explained, and the share of citations” come to the front. Monetization also moves from “installs / ad slots / clicks” to capability billing and task-pack subscriptions of “per call / per success / per throttle.” For consumers, loyalty will be built on “whether it was done, whether it was done in one go, and whether the handling of exceptions is clear.” For enterprises, OKRs and KPIs must bring in “the rate of being dispatched, consistency, explainability, and SLA”; the inventory of talent must add capability engineering, semantic operations, and trust compliance; and the financial standard must also move from a “spend—conversion” accounting to a servitized accounting of “the supply of capability—the return of being called.”

A manager’s action guide | Change “page thinking” into “capability thinking”

The first step is “capability-ization.” Break high-frequency processes into atomic capabilities, build for every action a complete contract description and observable metrics, and form a company-level Intent / Action catalog; at the same time, complete adaptation tests for the on-device OS capability catalog and for the mainstream AI-native plugin marketplaces, and make sure your contract of behavior can be called quickly and can respond steadily. The second step is “semantic-ization.” Build a knowledge base that is truly RAG-ready, present FAQs, terms, prices, SOPs, and cases in a structure that can be sliced, and provide, for typical tasks, a clear template of goal, constraint, and boundary, so that an agent takes fewer detours when assembling a plan; at the same time, write explainability into the specification of the output, so that when there is a failure or an uncertainty it can state the reason, propose an alternative, and retry. The third step is “operations-ization.” Switch operating metrics to “the rate of being dispatched, consistency, first-success rate, follow-up turns, average time to completion, the rate at which errors can be explained, and the share of citations”; bring in graded authorization, multi-factor confirmation, and a trail on the full chain; establish standing teams for semantic operations and for capability PMs; and continuously improve intent recognition, the filling-in of parameters, and the routing of errors. Finally, “ecology-ization.” Jointly build high-frequency task templates with device makers, OS, and model platforms, and first open up scenes such as travel expenses, contract approval, after-sales claims, and repurchase subscriptions; take an active part in industry standards of “cite-ability” and in compliance labels, and settle “being trusted” as a brand asset; run a dual-track pilot, let on-device calls and mobile-agent calls run in parallel, iterate weekly, review monthly, and let the organization grow muscle memory.

Closing: Two main lines, one change of generation at the base

Line one is “the end extends outward”: new devices move the understanding of intent and its execution to one’s side, and the app recedes into a library of capabilities. Line two is “the middle platform is agentized”: an AI-native task hub orchestrates scattered capabilities into a chain, and an agent taps for you, confirms for you. For an enterprise, the true upgrade is not in “making another app,” but in turning the service into a set of capabilities that an end and an agent can dispatch at any time. For a brand, the true place in the mind is not in first-screen exposure, but in being trusted by the system and cited first. The entrance has not disappeared. It has only moved from the screen to semantic routing — whoever first completes the switch from “page thinking” to “capability thinking” first holds the next generation of entrance and the engine of growth.

More deep-dive pieces

The judgment, restated

My judgment is that the entrance has already shifted — users tell an AI what they need, and the app recedes into a library of capabilities; people still fighting for a first-screen icon will miss the chance of being dispatched.

Three things this week

  • Inventory the capability APIs that can be dispatched by an agent (input / output / SLA).
  • Write one page of a minimum agent contract (authorization, confirmation, downgrade on failure).
  • Change one KPI into “the rate of being dispatched / the rate of being cited.”

If you are switching from page thinking to capability thinking, you are welcome to subscribe to Future Insights Lab, or to reply and discuss the sticking points of a shift of entrance.

訂閱 Future Insights Lab

不定期寄送 AI × Business 短篇筆記,協助你掌握實務重點。隨時可取消訂閱。