Why Human-Like AI Q&A Still Can't Get Things Done

A common experience is emerging: open a service portal and receive a fluent answer that explains policies, lists required materials, and even reminds you to complete identity verification. Yet when the page redirects, users must rediscover the relevant item; switching to another step forces them to repeat information; and non-standard problems render prior conversation useless.

The core issue is not that intelligent Q&A lacks intelligence, but that "answering a question" and "completing a task" are fundamentally different system capabilities. An answer resolves a momentary explanation — rules, materials, next steps — and ends after one interaction. Task completion, however, spans identity, items, materials, review, corrections, and result delivery. Each step can change what comes next and may produce exceptions no generic answer can resolve.

Users therefore care less about "what else can I ask" and more about concrete questions:

What specific task am I currently processing?

What has the system confirmed, and what is still missing?

Is this prompt generic advice or tied to my current progress?

When a non-standard situation arises, who handles it and how does the process continue?

Without carrying this information through the process, even the most natural dialogue stalls at the entry point — it aids understanding but not arrival.

Many "Intelligent Guidance" Systems Stumble on Service State Understanding

Consider a typical scenario: a user asks what materials an application requires. The system provides a checklist and hints at related items. After the user gathers some materials and returns, the interface restarts with the same generic checklist. When a document cannot be verified online, the system merely says "consult the window."

The model may not have answered incorrectly. The problem is that the system fails to link the two interactions as part of the same service process. Users see a string of individually correct prompts but feel no forward momentum. For service systems, the overlooked challenge is not "can it speak" but "can it recognize where a task has reached."

This explains why some LLM-powered portals feel like they "understand me" while others only "answer well." The difference lies not only in semantic understanding but in whether the system knows which process segment a response serves, whether it carries relevant conditions forward, and whether it can hand off at boundaries to the right person or channel.

"One-Stop Service" Is Not About Packing More Items onto One Page

"Efficiently completing one task" emphasizes optimizing processes, materials, and costs from the citizen and enterprise perspective, and requires forming a normalized promotion mechanism for key items. It reminds us that service integration is not page aggregation but reorganizing originally scattered actions around a single goal.

AI makes this look simpler: ask, search, pre-fill — seemingly close to "intelligent service." But looking deeper, intelligent capabilities only smooth the entrance; they do not automatically fix breakpoints between processes.

Truly continuous services typically keep four categories of information consistent:

Problem : what the user is actually solving, not just the sentence they typed.

Conditions : which qualifications, materials, or prerequisites are satisfied and which remain unconfirmed.

Progress : the current stage and what can be done next.

Exceptions : where to turn when the standard path fails, while preserving prior context.

These four categories together constitute "service state." Without it, intelligent Q&A is a patient explainer; with it, it becomes a reliable guide through a task journey.

Pre-Fill Does Not Decide for Users, Recommendations Should Not Replace Explanations

Public-service intelligence often involves intelligent Q&A, guidance, pre-fill, and assisted handling. All reduce repetitive input and search costs, but "less hassle" must not mean the system bypasses all user judgments.

For example, a recommendation may suggest related items; pre-fill may cut duplicate entry. Yet users still need to know the basis for recommendations, which information is unconfirmed, whether they can modify it, and how to exit or switch to human help when it does not apply. This affects both experience and trust.

For products, what keeps users in the flow is not an all-knowing entrance but a few plain certainties: I understand why the system prompted this; I can correct its understanding; I know what happens next and how exceptions are handled.

This does not mean every service must become a complex "fully autonomous agent." On the contrary, the closer to real handling, the more we must separate automation-friendly repetitive parts from those requiring rule interpretation, human judgment, and accountability.

Evaluating Intelligent Services: Ask Less "Is the Answer Right?"

"Is the answer right" matters, but it measures only a moment. For a continuously advancing service, we should also ask: did this answer reduce uncertainty for the next step? Does the user know their current state? Was an exception routed back to a continuable path?

Evaluation therefore moves beyond clicks, dialogue turns, or single-interaction satisfaction. More valuable signals hide in whether users repeatedly describe the same task, exit at critical nodes, continue smoothly after corrections, and where unfinished items ultimately stall.

As AI enters more complex public-service scenarios, the most needed design may not be "a more human-like responder" but a service relationship that does not easily break. The next step worth watching is how AI can help service systems retain problems, handle exceptions, and make every feedback genuinely improve the process — without replacing necessary human judgment.

Sources and References

State Council guidance on further optimizing government services and promoting "efficiently completing one task," Chinese Government Website, January 2024.

State Council General Office opinion on establishing a normalized promotion mechanism for key "efficiently completing one task" items, public reprint, covering "AI + government services," feedback loops, and data sharing.

State Council opinion on deepening the "AI+" action, China Internet Information Center reprint of Chinese Government Website public text, August 2025.