A Manager's Dilemma: Guessing Requirements, Gambling on Costs
Anyone who has led a small team knows this scenario well: in a product meeting, everyone argues over whether to build a certain feature, nobody can convince anyone else, and finally the boss decides "let's just build it first and see." Three months later it launches, the numbers are dismal, and only at the retrospective does it become clear—the original debate was, in essence, a debate over each person's imagined user.
The root of the problem isn't a lack of effort—it's that the cost of obtaining real user insights is too high. A proper user interview, from recruiting, scheduling, and conducting to transcribing the recordings, typically takes the team one to two weeks. Independent developers and small teams simply can't afford that time, so requirement decisions degrade into "gut calls + after-the-fact validation." One wrong bet means months of sunk costs.
Outsource the interviews? Prices are subject to the official pricing page, and outsourcing agencies often have little interest in small-scale projects anyway.
That was our team's state six months ago. Since then, we've worked out a path: use AI to simulate user interviews and distill a round of requirement hypotheses before doing real interviews. Note the key word—it's "hypotheses," not "conclusions." AI can't validate the market for you, but it can turn vague arguments into a clear, testable list of hypotheses, doubling the efficiency of the few real interviews that follow.
What AI Can Actually Do for Users Here
From a manager's perspective, I've summarized four core values:
First, translating "I think" into "users might think." In team debates, the rhetoric is "I believe users want this." AI-simulated interviews output structured user personas, scenario descriptions, and pain point statements. When the format changes, the quality of the discussion changes.
Second, compressing validation costs up front. A round of AI-simulated interviews takes just a dozen-plus minutes, and you can iterate through five or six versions of hypotheses in a day. Once the hypotheses converge, use real interviews to validate only the converged ones—real interviews shift from "exploration" to "verification," and sample requirements drop from twenty to five or six.
Third, exposing hidden disagreements within the team. Have different members each describe the target user, then have the AI simulate them separately—you'll often discover you're actually serving two different users. This kind of disagreement usually hides in requirement reviews; AI simulation makes it explicit.
Fourth, a safety net for risk control. Simulated AI users can ask those "too basic, so nobody dares ask" questions on your behalf, like "whose problem does this feature actually solve, and what problem?" Before committing development resources, these cheap questions can block expensive mistakes.
Our Actual Process: Four Steps
Step 1: Align the Team's Description of the Target User
Gather the core members and have each person describe the target user in three sentences. No discussion—just write. Then feed all the descriptions to the AI and have it point out the contradictions. This step usually sparks arguments—a good thing. Arguing early is cheaper than arguing late.
Step 2: Run Simulated Interviews with a Prompt Template
This is the core step. We've polished a template you can copy directly:
你是一位有10年经验的用户研究员,正在主持一场用户访谈。
【产品背景】
产品名称:{产品名}
一句话描述:{解决什么问题}
目标用户画像:{年龄/职业/场景/使用动机}
【访谈目标】
1. 验证核心痛点是否真实存在
2. 探索用户现有替代方案及其不满
3. 识别用户愿意为之付出成本的具体场景
【你的任务】
1. 先以第一人称扮演上述画像的用户,生成这个用户的背景小传(150字以内)
2. 然后模拟一场8-10轮的访谈对话:你扮演访谈者提问,并同步扮演该用户回答。用户的回答要符合画像认知水平,允许出现“不了解”“没想过”“不在乎”等真实反应,不要过度配合
3. 访谈结束后,输出:
- 3条最强烈的痛点陈述(用用户原话语气)
- 2条与团队预设矛盾的意外发现
- 3条待验证的假设,每条标注【高/中/低置信度】和验证方式建议
4. 最后自我审视:本次模拟最可能失真的地方是什么?
【约束】
- 用户回答不得流露出对产品的预先好感
- 禁止生成具体市场数据或百分比The final "self-review" item is something we added—having the AI acknowledge the boundaries of its own simulation effectively prevents the team from treating simulation results as truth.
Step 3: Multiple Rounds, Multiple Personas, Cross-Checking
A single simulation can go off track. We usually run 4-6 different personas, with two rounds per persona (the answers in the first round influence the follow-up directions in the second). Then we have the AI do a cross-persona summary, outputting common pain points and divergent pain points.
Step 4: Generate the Real Interview Outline
Transform the AI's "hypotheses to be validated" into an outline for real interviews, and use that outline to find 5-6 real users. The real interviews have only one task: falsify or confirm these hypotheses.
Before and After Using AI
| Dimension | Before AI | After AI |
|---|---|---|
| Requirement discussion format | Meeting arguments, everyone talking past each other | Aligning point by point around a hypothesis list |
| Hypothesis output | Implicit, undocumented | 10+ explicit hypotheses per round |
| Real interview sample | 15-20 exploratory interviews | 5-6 verification interviews |
| Hypothesis documentation | Basically none | Archived and traceable each iteration |
| Decision risk | Discover the mistake only after betting wrong | Hypotheses converged before launch |
For managers, the most valuable thing is actually beyond that last row—traceability. Behind every decision there's a hypothesis document: who proposed the hypothesis, which simulation round validated it, whether the real interviews confirmed it—the whole trail is on record. No more finger-pointing at retrospectives; just look back at the documents.
Risk Boundaries That Must Be Made Clear
Managers need a clear-eyed understanding—three red lines:
- AI users are synthetic, not real. Their value is converging hypotheses and generating interview outlines, not replacing market validation. Treating simulation results as conclusions is the biggest abuse of this method.
- Beware of the "echo chamber." Persona descriptions carry the team's own biases, and the AI will play along with your biases. So diversify the sources of team descriptions, and keep the AI's "uncooperative answers" setting.
- Do not put AI simulation results into materials for investors or clients. That's synthetic data—it is not, by nature, the voice of users.
Final Thoughts
After this method worked, our team's requirement reviews went from "battlegrounds of opinions" to "hypothesis checklist reviews." AI didn't make decisions for us, but it changed the cost structure of decisions—wrong hypotheses die in the prompt, instead of dying three months into a development cycle.
If your team also wants to build this kind of "AI-assisted research + hypothesis-driven development" workflow, stable model calls are the foundation. Multi-persona, multi-round simulated interviews place demands on API concurrency and stability. You can register and try it at https://api.thistoken.ai/register to get your first round of simulated interviews running.
---
Every example in this post runs with a single API key — get yours at https://api.thistoken.ai/register and start in minutes.
Bạn muốn thử Token.AI?
Tạo API Key cấp dự án, bật kênh trong bảng điều khiển và định cấu hình định tuyến, ngân sách và nhật ký kiểm tra.
注册 ThisToken.AI 并获取 API Key