Don't Ask Users What They Want—Observe Work to Find Real AI Opportunities

Observe How Work Happens, Don't Start by Asking for Features

Many AI project kickoffs begin with a direct question: "What features do you want AI to build for you?" This question seems straightforward, but the answers are usually patch suggestions for existing systems: a Q&A bot, auto-generated reports, an alert, or data integration. These answers can go into a backlog, but they don't necessarily represent the AI opportunities worth investing in.

Real AI opportunities hide not in feature lists, but in the waits, repeated checks, information gaps, experience-based judgments, and exception handling within actual work processes.

First, See How Work Actually Happens

Business people often struggle to describe their real work in a meeting room. They describe the official process but miss the ad-hoc communications, cross-system lookups, manual compilation, experience judgments, and remedial actions that happen in practice.

Therefore, research must return to the workplace and follow a concrete task end-to-end. This method, called shadowing (or job shadowing), is not just attending a process walkthrough; it observes how tasks are triggered, how information flows, how judgments are made, and what ad-hoc communications and fixes exist outside the formal process.

Key observation points:

Where the task originates and who confirms completion;

Which systems, documents, messages, or human conversations provide information;

Which steps consume the most time and which repeat;

Which judgments rely on experience and where evidence is missing or contradictory;

Who intervenes when exceptions occur, what basis they use, and how the outcome is recorded.

The goal is not a pretty flowchart but to discover where work truly gets stuck.

Real AI Opportunities Often Appear in Five Types of Friction

First, repeated searching and compiling information. For example, staff must jump between multiple systems, documents, and chat logs to manually assemble a complete context for a case, order, device, or customer.

Second, repeated comparison and verification. For example, repeatedly checking whether materials are complete, data is consistent, rules are satisfied, or different people's statements or records conflict.

Third, experience-based judgments that are hard to pass on. For example, senior staff know which signal combinations indicate risk, while juniors can only learn by repeated consultation and post-mortems.

Fourth, exceptions discovered too late. For example, problems are often found only after approval, inspection, archiving, or complaints, by which time earlier processing cannot be corrected at low cost.

Fifth, processing results that never become feedback. For example, manual corrections, disposal conclusions, and failure reasons stay in personal experience and never feed back into knowledge, rules, or datasets.

These frictions don't all need AI, but they are entry points worth deeper analysis.

Not Every Pain Point Is Suitable for AI

Whether a problem fits AI depends not on how difficult it sounds, but on whether basic conditions hold:

Relatively stable input materials, data, or evidence exist;

The understanding, retrieval, verification, generation, recommendation, or invocation actions AI should assist can be clearly defined;

Necessary human confirmation and responsibility boundaries can be retained;

Observable processing results exist to verify whether AI actually improves the work;

A single wrong suggestion would not cause unacceptable consequences.

If rules and responsibilities are unclear and data cannot be traced, introducing AI often only amplifies the existing chaos.

Turn Field Findings into AI Opportunity Cards

After business research, don't just deliver interview notes and requirement lists. A more useful artifact is an AI Opportunity Card that captures at least:

Target Role: Who is handling the problem;
Trigger Scenario: At what node the problem occurs;
Current Friction: Where time is spent and where risk appears;
Available Evidence: Where systems, documents, logs, rules, and human confirmations come from;
AI-Assisted Actions: Understanding, retrieval, comparison, verification, generation, alerting, or invocation;
Human Boundary: Who confirms and under what circumstances they must take over;
Validation Metrics: How to prove efficiency, quality, risk, or experience improves.

Traditional Requirement:

I want an intelligent device Q&A.

AI Opportunity Card:

Help device maintenance staff quickly aggregate recent device alarms, repairs, and process records before inspections, highlight test points that need focused verification, and let staff confirm before acting.

The former is just a feature name; the latter defines the served role, business node, available evidence, AI-assisted actions, and human boundaries—making it an evaluable AI opportunity.

The Endpoint of Research Is Opportunity Prioritization, Not Requirement Collection

Frontline staff can raise many valuable problems, but the first project cannot tackle them all. After research, the team should place opportunity cards on a single priority list, comparing business impact, urgency, evidence foundation, process boundaries, risk controllability, and validation difficulty.

Then the next discussion shifts from "which feature is cooler" to "which opportunity deserves to enter a pilot first."

Summary

Business research is not about asking users for a feature list; it's about understanding how real work happens, where friction exists, which judgments lack support, and within what boundaries AI can help.

Turning these findings into verifiable AI Opportunity Cards and then into pilot selection avoids translating a vague pain point directly into a functionally complex, value-unclear AI system.