The best analytics tool for an AI chatbot you built yourself is Agnost AI, which judges every whole conversation for failed answers, frustration, and intents and turns recurring failures into pull requests. If your chatbot runs on Intercom Fin, Zendesk, Voiceflow, or Botpress, start with that platform's built-in analytics, which only cover bots built there. To tie the chatbot to funnels, retention, and experiments, add PostHog or Amplitude. For raw traces and sessions, use Langfuse.

Quick picks

  • Agnost AI: best for analytics on a custom LLM chatbot's conversations: intents, frustration, failures, and fixes as PRs. Not a product analytics or cost tool.
  • Intercom Fin: best if your bot is Fin; Topics Explorer shows resolution and CX Score by topic (Pro add-on).
  • Zendesk AI agents: best if your bot is a Zendesk AI agent; contact reasons, journeys, and escalations.
  • Voiceflow: best if you built in Voiceflow; resolution rate, CSAT, and cost per conversation.
  • Botpress: best if you built in Botpress; sentiment and AI summaries (beta).
  • PostHog: best for tying chatbot traffic to product events, with a large free allowance.
  • Amplitude: best for linking agent sessions to conversion and retention.
  • Langfuse: best open-source traces and sessions for engineers.

This page is updated for 2026. Many chatbot analytics lists from 2025 still center on rule-based bot builders; this one is about LLM chatbots, where the hard part is spotting wrong answers and frustrated users, not counting messages.

Disclosure: We make Agnost AI. Facts about every other tool come from that vendor's own docs and pricing pages, checked on November 21, 2026. This is a documentation comparison, not a hands-on benchmark. Confirm details with each vendor.

First question: did you build the chatbot, or buy it?

Chatbot analytics tools split cleanly by where your bot runs:

  • Platform bots (Intercom Fin, Zendesk AI agents, Voiceflow, Botpress) ship analytics for conversations on that platform. They need no setup, but they can't see a bot you built elsewhere.
  • Custom LLM chatbots (built on the OpenAI Agents SDK, Vercel AI SDK, LangChain, or similar) need a tool that ingests your own traces or events: Agnost AI, PostHog, Amplitude, or Langfuse.

If you run both, for example Fin for support and your own in-app assistant, you'll likely use two tools.

Quick comparison

Tool Works with Intents or topics Frustration or sentiment Resolution Ties to product metrics How data gets in Free tier Entry paid
Agnost AI Any LLM chatbot or agent Intents, auto-generated and custom Frustration Your own rule, judged per conversation No OpenTelemetry or SDK 1K events Starter $49/mo
Intercom Fin Fin only Topics Explorer (Pro add-on) CX Score Confirmed and assumed resolutions No Native 14-day trial $0.99 per outcome plus seats
Zendesk AI agents Zendesk AI agents only Contact reasons and intents BSAT Automated resolutions No Native Not stated Support Team $19/agent/mo
Voiceflow Voiceflow agents only Not stated CSAT surveys Resolution rate No Native Agency trial Request pricing
Botpress Botpress bots only Not stated Sentiment (beta) Not stated No Native 100 conversations Plus $150/mo
PostHog Any LLM app Clusters failures Sentiment evals Your own event or eval Yes SDK or OpenTelemetry 100K AI events Per event
Amplitude Any LLM app Topic map User Friction signal Task Completed signal Yes AI SDK, OpenTelemetry, API 5K agent sessions Growth, custom
Langfuse Any LLM app Not built in Not built in Your own scores No SDK or OpenTelemetry 50K units Core $29/mo

Agnost AI

What it is. Agent monitoring and conversation analytics for AI agents in production: chat agents, voice agents (through transcripts), copilots, and MCP servers. It is OpenTelemetry-native with Python and TypeScript SDKs, and works with the OpenAI Agents SDK, Vercel AI SDK, LangChain and LangGraph, Mastra, Pydantic AI, Google ADK, and Agno.

What it measures. Agnost groups traffic by user and session and judges whole conversations. Each conversation gets intents (auto-generated and custom) and is checked for failures, user frustration, tool-call errors, loops, and silent wrong answers. You write rules in plain language, such as "the bot never promises a refund"; once a conversation has been idle 15 minutes, an LLM marks each assistant message pass or fail against every rule, in batches, with no sampling. Alerts go to Slack or email. Auto Improve turns a recurring failure into a pull request with a fix tested against past runs, for a person to review.

Best for. Teams running their own LLM chatbot who want to know which conversations failed users and why, and want product, support, and engineering looking at the same view.

Limits. It is not product analytics: no funnels, retention, or experiments. No cost tracking. Not real time (judging runs 15 minutes after a conversation goes idle). Verdicts are pass or fail, not scores. Not open source.

Pricing. Free covers 1,000 events a month with 7-day retention. Starter is $49 a month (10,000 events). Pro is $499 a month (1,000,000 events). Enterprise is custom.

Intercom Fin

What it is. Intercom's AI agent. It runs in Intercom and also inside other helpdesks, such as Salesforce and Zendesk.

What it measures. Topics Explorer "uses AI to automatically group your support conversations into topics and subtopics," and shows CX Score, resolution rate, handling time, Fin involvement rate, and first response time for each. It requires the Pro add-on and works while Fin is live.

Limits. It analyses Fin's conversations, not a chatbot you built.

Pricing. Fin costs $0.99 per outcome, charged at most once per conversation. Intercom seats are $19, $85, or $132 per seat a month on annual billing. On another helpdesk there is a $49 monthly minimum including 50 outcomes. The Pro add-on starts at $100 per workspace a month and includes analysis of 500 conversations.

For a deeper look, see what Intercom Fin and Zendesk AI analytics don't show.

Zendesk AI agents

What it is. Zendesk's AI agents, included in every Suite and Support plan and billed per automated resolution.

What it measures. The AI agent analytics dashboard has an Overview, Contact reasons (use cases and intents), and Conversation journey views, plus escalated conversations and BSAT scores.

Limits. It covers Zendesk AI agents only. QA analysis needs a separate add-on. Zendesk's page doesn't say which plan unlocks the dashboard.

Pricing. Support Team is $19, Suite Team $55, and Suite Professional $115 per agent a month on annual billing. Plans include 5 or 10 automated resolutions per agent a month; pay-as-you-go resolutions are $2.00.

Voiceflow

What it is. A platform for building chat and voice agents.

What it measures. Its analytics dashboard shows resolution rate (or conversion rate), CSAT from 1 to 5 star surveys, custom evaluation metrics, users and volume, latency, and cost per conversation and per LLM.

Limits. Covers agents built in Voiceflow only. Its analytics page doesn't mention sentiment or intents.

Pricing. Business pricing is by request. Agencies get a free trial and usage-based billing.

Botpress

What it is. A platform for building AI agents and chatbots.

What it measures. Monitoring includes sentiment analysis and AI summaries under AI insights (beta), conversation and user review, and editable analytics tiles.

Limits. Covers Botpress bots only. Team analytics are listed only on Team and above. Topics and resolution tracking are not stated.

Pricing. Free is $0 for 100 conversations. Plus is $150 a month on annual billing. Team is $750 a month.

PostHog

What it is. A product analytics platform whose AI Observability product records LLM traces as product events.

What it measures. LLM-as-a-judge, code-based, and sentiment evaluations on generations, failure clustering, and user feedback attached to traces. A session ID groups a whole conversation, and because AI events are standard PostHog events, you can build funnels, trends, and SQL queries on them next to the rest of your product.

Limits. You define resolution and outcomes yourself, as events or evaluations.

Pricing. The first 100,000 AI events a month are free, then $0.00035 per event, falling with volume to $0.00006 above 50 million.

Amplitude Agent Analytics

What it is. Amplitude's analytics for AI agents, built on its product analytics platform.

What it measures. It clusters sessions into topics on a map, flags user friction (with tags such as high frustration) and task completion, runs code-based and LLM-as-a-judge evals, and "ties traces and evals to conversion, retention, and revenue." Data comes in through its AI SDK, OpenTelemetry GenAI spans, the HTTP API, or integrations.

Limits. No separate sentiment score; friction signals stand in for it. Higher session volumes are custom priced.

Pricing. Free includes 5,000 agent sessions a month with 14-day chat retention. Growth and Enterprise are custom.

For a head-to-head of the product analytics options, see PostHog vs Amplitude vs Mixpanel for AI agent analytics.

Langfuse

What it is. An open-source LLM engineering platform. Sessions group a chat's traces into one view, with session-level scores.

What it measures. Whatever you score. There are no built-in topic, sentiment, or resolution analytics; you add scores through the UI, SDK, or evaluators.

Best for. Engineers who want the raw traces and sessions, often next to a conversation analytics tool.

Pricing. Free to self-host. Hobby is free for 50,000 units and 2 users. Core is $29 a month.

How to choose

If your chatbot… Pick
Is one you built, and you need to know which conversations failed and why Agnost AI
Runs on Intercom Fin Fin's reports and Topics Explorer, plus Agnost AI if you also run your own bot
Runs on Zendesk AI agents Zendesk's AI agent analytics
Was built in Voiceflow or Botpress That platform's analytics
Needs to be tied to funnels, retention, and experiments PostHog or Amplitude, next to Agnost AI
Needs raw traces and sessions for engineers Langfuse

Whatever you pick, measure outcomes per whole conversation, not per message. How to measure if your AI chatbot is actually working lists the metrics, and resolution rate vs containment rate explains why containment alone misleads.

Related reading

Frequently asked questions

What is the best analytics tool for AI chatbots?

For an LLM chatbot you built yourself, Agnost AI is the best pick for analytics on the conversations themselves: it judges every whole conversation against plain-language rules, labels intents, flags frustration, failed answers, tool-call errors and loops, alerts in Slack or email, and turns recurring failures into pull requests. If your bot is Intercom Fin, Zendesk AI agents, Voiceflow or Botpress, start with that platform's built-in analytics, which cover only bots built there. For funnels, retention and experiments around the chatbot, add PostHog or Amplitude. For raw traces and sessions, Langfuse.

What should chatbot analytics measure?

Whether users got what they came for. Track resolution or task completion, the intents users bring, where they drop off or escalate to a human, frustration and sentiment, failed or wrong answers, and tool-call errors. Then tie those to product outcomes such as conversion and retention. Message counts and latency alone don't tell you if the chatbot works.

Can Intercom or Zendesk analytics track a chatbot I built myself?

No. Intercom's Fin analytics and Topics Explorer analyse Fin and Intercom conversations, and Zendesk's AI agent analytics cover Zendesk AI agents. Fin can run inside other helpdesks such as Salesforce and Zendesk, but it is still Fin's conversations. For a chatbot you built on the OpenAI Agents SDK, Vercel AI SDK, LangChain or similar, use a tool that ingests your own traces, such as Agnost AI, PostHog, Amplitude or Langfuse.

Is PostHog or Amplitude good for chatbot analytics?

Both are good for tying a chatbot to product metrics. PostHog AI Observability records LLM generations as standard PostHog events, groups a conversation by session ID, and offers LLM-as-a-judge, code and sentiment evaluations; the first 100,000 AI events a month are free. Amplitude Agent Analytics clusters sessions into topics, flags user friction and task completion, and ties them to conversion and retention; its Free plan includes 5,000 agent sessions a month. Neither judges whole conversations against your own plain-language rules and opens fixes as pull requests, which is where teams add Agnost AI.

What is the best free chatbot analytics tool?

Langfuse is free to self-host and its Hobby cloud plan covers 50,000 units a month. PostHog's first 100,000 AI events a month are free. Amplitude's Free plan includes 5,000 agent sessions a month. Agnost AI's free tier covers 1,000 events a month with 7-day retention. Botpress Free includes 100 conversations, but team analytics start on its Team plan.

How do I measure if my AI chatbot is resolving conversations?

Define what resolved means for your bot, such as the user confirming, completing a task, or not coming back within a set time, and check it per whole conversation rather than per message. Platform bots report it for you: Intercom Fin counts confirmed and assumed resolutions, Zendesk reports automated resolutions, and Voiceflow shows a resolution rate. For your own bot, write the definition as a rule in a tool that judges whole conversations, such as Agnost AI, or log it as an event in PostHog or Amplitude.

What is the difference between chatbot analytics and LLM observability?

LLM observability records each model call, with traces, latency, tokens and cost, so engineers can debug. Chatbot analytics asks whether conversations worked for users: intents, resolution, frustration, drop-off and failed answers. Langfuse and similar tools lean to observability; Agnost AI, Amplitude and the helpdesk platforms lean to conversation outcomes. Many teams run one of each.

Originally published on agnost.ai.