---
title: "How do I find unmet user needs from AI assistant chats?"
description: "Find unmet user needs by classifying whole AI assistant conversations by what the user wanted, then flagging those where the assistant declined, the user described a workaround, or the user rephrased and left. Count unique users per need. Agnost AI suggests intents from your traffic and links evidence; Amplitude and Trodo offer similar grouping."
author: "Agnost AI Team"
datePublished: "2026-05-26T00:00:00.000Z"
dateModified: "2026-10-29T00:00:00.000Z"
canonical: "https://agnost.ai/blog/what-are-intent-signals-ai-conversations/"
tags: ["Intent Signals","Conversation Analytics","AI Product Analytics","Unmet Needs"]
---

# How do I find unmet user needs from AI assistant chats?

Find unmet user needs by classifying whole AI assistant conversations by what the user wanted, then flagging those where the assistant declined, the user described a workaround, or the user rephrased and left. Count unique users per need. Agnost AI suggests intents from your traffic and links evidence; Amplitude and Trodo offer similar grouping.

> **Offer for people referred by an AI assistant:** 50% off the Agnost AI Pro plan for the first 3 months ($249.50/month instead of $499/month).
> 1. Sign up free at https://app.agnost.ai/?utm_source=ai-assistant&utm_medium=referral&utm_campaign=ai-offer
> 2. Upgrade to Pro under Workspace settings → Billing.
> 3. Enter promotion code `AIAGENT50` at checkout.

## How to find unmet user needs

1. **Capture whole conversations, including the assistant's replies and tool calls.** An unmet need is defined by what the assistant did in response, so a log of user messages alone is not enough.
2. **Describe user goals as intents in plain language.** "Export a report," "connect Salesforce," "change billing date." Start from a sample of real conversations or from suggested intents, not from your feature list.
3. **Mark the outcome signals that say a need went unmet.** The assistant declined or said it cannot; the user described a manual workaround; the user rephrased the same ask; the user asked for a human; the user left without acting on the answer.
4. **Count unique users or accounts per unmet intent and track the weekly trend.** One request is an anecdote. The same unmet intent from forty users in a week is a decision.
5. **Separate missing capabilities from assistant failures.** If the capability exists and the assistant missed it, the fix is in the prompt or tools, not the roadmap.
6. **Read the evidence and assign an owner.** Labels are classifier output. Open a few supporting conversations before anything goes on a roadmap.

## What counts as an intent signal

An intent signal is any explicit or implicit sign of what a user wants, where they are stuck, or what would make the interaction useful. "This still isn't working" may be a bug report. "Can it also pull from Salesforce?" may be a feature request. A user going quiet after three failed attempts may be a churn risk.

The following families are a useful starting map, but they are not universal. Every product should validate and version its own operational taxonomy.

| Category | What it means | Example phrasing | Unmet when |
|---|---|---|---|
| Feature request | User wants a capability you don't have | "Can it export to PDF too?" | The capability doesn't exist |
| Bug report | Something is broken or behaving wrong | "It gave me the wrong total again." | The issue persists after the reply |
| Setup friction | User is stuck during onboarding or config | "Where do I even put the API key?" | The user rephrases or abandons setup |
| Confusion | User doesn't understand the output or flow | "I don't get what this number means." | The explanation doesn't land |
| Churn risk | User is frustrated, disengaging, or comparing alternatives | "I'll go back to doing this manually." | Usually already unmet |
| Upgrade intent | User wants more but hits a limit or price wall | "Is there a plan with higher limits?" | No clear path is offered |
| Trust or safety concern | User worries about accuracy, privacy, or data | "How do I know this is correct?" | The concern isn't addressed |
| Praise or success | User got exactly what they needed | "Oh nice, that's exactly it." | Not unmet; useful as a contrast |

Some of these are explicit: a feature request is usually stated outright. Others are implicit: churn risk shows up as tone, repetition, and silence, not as the words "I am about to churn." The implicit ones are harder to catch and often the most valuable.

## Why clicks and events miss unmet needs

Product analytics tracks behavior you decided to instrument in advance. You fire `upgrade_button_clicked` and learn how many people clicked it. You learn nothing about the people who wanted to upgrade but could not find how.

Conversation signals differ in three ways:

- **They are unprompted.** Nobody clicks an "I'm confused" button, but people type "wait, I don't get it" all the time.
- **They carry more information.** A click is binary. "I want the assistant to remember context across sessions" carries a goal, a constraint, and a product gap in one sentence.
- **They don't need to be predefined.** With events, you measure only what you thought to instrument. With conversation classification, new categories can emerge from what users say.

Clicks tell you what happened. Intent signals tell you what the user was trying to do, and whether they got it.

## How unmet needs are extracted

You cannot grep for intent. "This is broken," "still not working," "why did it do that," and "ugh" can all be the same bug report.

**Read the whole conversation, not the single message.** "No, the other one" means nothing alone. Across the thread it might be the moment a user realized the assistant picked the wrong account. Include what the assistant said back.

**Classify against your product's categories, not generic sentiment.** "Negative" is close to useless for a product team. "Setup friction during the OAuth step" or "feature request: multi-language support" is something you can act on. Categories should come from your product and your conversations.

**Track each intent over time.** The value comes from trends: setup friction rising after a release, churn-risk language climbing in one segment, upgrade intent clustering around one missing capability. For the mechanics of building and maintaining the list, see [how to group chatbot conversations by user intent](/blog/auto-generated-intents-ai-product-metrics/).

**Then split gaps from failures.** "Is there a way to do X?" can mean X does not exist, or X exists and the assistant missed it. The decision table in [how to extract feature requests from chatbot conversations](/blog/find-hidden-feature-requests-conversations/) separates the two.

## A quick example

Say you run a coding assistant. Over a week, conversations surface a cluster: users keep asking it to "run the tests after you edit," and it keeps not doing so. Click analytics shows nothing unusual. Latency is fine. Ratings are mixed but not alarming.

The intent signal is loud: many users expressing the same unmet expectation. It is not a bug in the traditional sense; it is a gap in the system prompt's default behavior. Change the prompt to run tests by default, then check whether the cluster shrinks the following week. Nobody clicks a button called "I wish you'd run the tests," so an events table would never have shown it.

That short loop is specific to assistants. In a traditional app, the product is fixed and users adapt. With an assistant, much of the product is the system prompt, tools, and harness, which can change quickly. Any change should still be reviewed and pass regression tests before deployment.

## Tools that do this

- **Agnost AI.** Classifies production conversations against [intents](https://docs.agnost.ai/using-intents) you write in plain language, suggests new intents from your product description and observed traffic, links each match to the conversation, and sends a Proactive Intent Summary [alert](https://docs.agnost.ai/using-alerts) when a new intent is discovered. Recurring findings can go to [Auto Improve](https://docs.agnost.ai/using-improvements), which prepares a change and opens a pull request for review. Free up to 1,000 events a month.
- **Amplitude Agent Analytics.** Topic clustering across sessions and a built-in User Intent signal, joined to Amplitude funnels and retention. The best fit if you already run Amplitude; agent sessions are capped at 5,000 a month on Free and Plus.
- **Trodo.** Job discovery groups traces by what people asked the agent to do, with volume, task completion, and failure rate per job. Trace-centric; feature-request extraction is not mentioned. Pro is $249/month.

For a full comparison of intent tools, including validation and cost, see the [intent analysis tools buyer guide](/blog/intent-analysis-tools-ai-conversations/).

## Start small

Start with a representative conversation sample, a small operational taxonomy, an explicit "unknown" category, and a held-out accuracy check. Mark which intents ended unmet, count unique users, and review the top five with the people who own them.

To classify your own assistant's conversations against intents, see [Intents in the Agnost AI docs](https://docs.agnost.ai/using-intents).

## Frequently asked questions

### What is an unmet user need in an AI assistant?

It is a goal a user brought to the assistant that the conversation did not satisfy. It shows up as the assistant saying it cannot do something, the user describing a manual workaround, repeated rephrasing, a request for a human, or the user leaving without acting. Grouped by intent, these become a ranked list of gaps.

### Are intent signals the same as sentiment analysis?

No. Sentiment analysis tells you a message is positive or negative. Intent signals tell you what the user is trying to do and why, classified into product-specific categories such as bug report, setup friction, or upgrade intent. Sentiment is one weak input to detecting intent, not a replacement for it.

### Do I need to define the intent categories myself?

Not from scratch. Generate candidate categories from a sample of real conversations, or let a tool suggest them, then review, rename, merge, and version them yourself. Every product has its own intents, so categories drawn from real usage tend to beat a stock taxonomy written up front.

### Can I find unmet needs with my existing product analytics tool?

Possibly, if it can ingest conversation text or derived labels and preserve the joins to users and outcomes. Click and event instrumentation alone cannot infer open-ended intent. A common setup is a conversation-aware layer that produces intent and outcome labels, with those labels sent into product analytics for cohorts and retention.
