Perceptional swaps static surveys for AI-led interviews

The situation

Amer Abu Khajil spent years at Amazon, one of the most customer-obsessed companies in the world, and still couldn't get real feedback from his own users. Not because the tools didn't exist. Because every tool gave him the same shallow answers.

A static survey asks the same questions in the same order to every respondent. Surveys collect answers. They don't collect context. If someone says they find a feature confusing, the survey moves on. There is no follow-up. No "what specifically was confusing?" No "how did you try to work around it?" The hardest skill in research has never been asking questions. It is the follow-up: a weak moderator writes down "confusing," and a good one spends the next four questions finding out what confusing actually meant. Product managers end up with data that confirms what they already suspected but never surfaces what they actually needed to know.

Amer came to us with a prototype idea. We built Perceptional: an AI platform that interviews users the way a researcher would. It adapts its questions in real time based on what each user says, maintains full context across the entire conversation, and generates a structured summary within 48 hours of the interview completing.

What we had to solve

• Keeping a multi-turn conversation coherent from first question to last

  A real interview has a thread. The researcher remembers what was said in question two when they ask question eight. Standard AI implementations lose this context as conversations grow longer, treating each exchange as near-isolated rather than as part of a continuous session. The known failure mode is deterministic probing: the AI re-asks the same question until the participant gets irritated, instead of laddering deeper into what they just said.

  We needed Perceptional to hold the entire conversation in working memory, reference earlier answers when deciding what to ask next, and never repeat a question the user already answered. Claude's extended context window made this possible. Smaller models would have broken interview coherence mid-session.

• Turning unstructured conversations into structured summaries a PM can act on

  A conversation with 15 exchanges produces a lot of text. Product managers need the signal, not the transcript. We had to design a summarisation layer that could read the full conversation, identify the themes and patterns that appeared across responses, and format them into a structured summary a PM could use in a sprint planning meeting. The summary had to be accurate enough to reference specific user language but concise enough to read in five minutes. Getting this right required iteration on the prompt engineering and the output format with Amer's team until the summaries were genuinely useful.

1. Every interview follows the user's actual answers, not a fixed script

   The AI reads each response and decides what to ask next based on what the user actually said, not a fixed script. If a user mentions confusion with a specific feature, the platform follows up on that confusion. If they express enthusiasm, it probes what drove it. Each interview takes a different path because each user's answers are different. This is what separates an AI interviewer from a survey with a voice layer.

   

2. Product managers get answers before the next sprint, no transcript reading required

   After each interview session closes, the platform analyses the full conversation thread and generates a structured summary with themes, specific user language, and concrete findings. No manual reading of transcripts. Research stops arriving one sprint behind the decisions it was meant to inform.

   

3. Campaign launches without an infrastructure team watching it

   Built on AWS Lambda, Perceptional runs thousands of simultaneous interviews without infrastructure management. A product manager can launch a campaign to thousands of users and the platform handles the load automatically. The beta and enterprise workloads run on the same architecture.

   

4. Themes that repeat across interviews surface differently than one-offs

   The platform analyses entire conversation threads for context, not just individual answers, and surfaces patterns that repeat across multiple users. A theme that appears across many interviews is surfaced differently than one that appeared once. Product managers see where the signal is strong and where it is noise.

   

How it runs

AWS Lambda

Amer was building alone. Lambda means Perceptional handles thousands of simultaneous interviews without an infrastructure team. No servers to provision, no idle capacity to pay for between interview campaigns.

Claude (Anthropic)

Claude's extended context window keeps the full conversation in memory. The AI remembers what was said in the first exchange when asking the tenth question, which is what makes follow-ups feel like a real interview rather than a disconnected form.

AWS Bedrock

Bedrock gives a single control plane for model management and versioning. When Anthropic ships a better model, we update Perceptional without touching application code. The platform stays current without a re-architecture.