AI for Chatbots: What It Is and Which Tools to Use in 2026

By Samuel Rose, Founder of Agensi. Published September 28, 2026. Last updated September 28, 2026.

An AI chatbot is software that holds a conversation in ordinary language. The AI part is a large language model: a system trained on a very large amount of text that predicts a good next sentence given everything said so far. The chatbot part is what wraps it: a place to type, a memory of the conversation, and, for a business bot, a connection to your own documents so the answers come from your help center rather than from the model's general knowledge. That last piece is the whole difference between the two halves of this page.

If you want an assistant for yourself, start with ChatGPT or Claude and stop reading after the first group. If you want a bot that answers your customers from your own help center, the second group is for you, and the one question that decides whether it is usable is whether it can show which passage of your source produced each answer. Our picks: Google Cloud's Conversational Agents if you have an engineer, Intercom Fin if you have the budget, Zendesk if you already have Zendesk, and Voiceflow if you want to build it yourself. This page is not about roleplay or companion apps.

Which one do you actually need

Google's own answer to this search ends by asking two questions: do you want a ready-to-use assistant or a tool to build your own, and what is the main goal. Those two questions are the routing.

A ready-to-use assistant is a product you open and talk to. It knows the world in general, it does not know your business, and it costs $0 to $20 a month per person. If the goal is writing, research or thinking out loud, this is you.

A builder is a platform you train on your own material and put in front of your customers, on your website, in your app or in your help desk. It answers only from what you gave it, or it should, and it is billed per conversation, per resolution or per message. If the goal is support, lead qualification or answering the same questions at 2am, this is you. A small business with under a few hundred conversations a month should also read AI chatbots for small business, which is priced for that scale. For the wider stack around a bot, see best AI tools for business.

Character and roleplay chatbots are a third category with its own audience. They are out of scope here.

How AI chatbots work

Three layers. Language processing: the model reads your words as tokens and works out what is being asked. The model itself: a large language model generates a reply one token at a time, choosing what is most likely given the conversation. Context: everything the model can see when it answers, which is the conversation so far plus, in a business bot, the passages retrieved from your documents. A ready-made assistant's context is mostly the conversation. A builder's context is mostly your documents, and its job is retrieval: find the right passage, hand it to the model, and make the model answer from it rather than from memory. When that retrieval fails, the model still answers, and that is where invented answers come from.

What people use chatbots for

Customer support is the first use in Google's answer and the one this page is built around. Sales and marketing come second: qualifying a visitor, booking a demo, answering pricing questions, covered by the Claude skills for sales and lead generation collection. Content and productivity is the ready-made half: drafting, summarising, research.

How we picked these tools

This is the short form of the Agensi method, the same one on every ranking page we publish. Two of the top three results for this search are ranked lists with no method: one scores itself 9.1 against Claude at 8.5 on its own page, the other puts "tested" in the title and shows no test. A stated method is cheap to publish and the only thing that makes a ranking checkable.

The longlist had 17 tools: the top search results, the platforms Google's AI Overview names on each side of its own split, and the Agensi collections. A tool stays if it passes four checks. Available: buyable without a sales call, or labelled sales-led. Alive: a release or price change in the last 12 months. On the job: it is a chatbot or a chatbot builder. Evidenced: tested by us, or rated at least 4.0 from 50 or more reviews on a named platform. Five fell out: DeepAI Chat and the three voice platforms on the rating bar, and Chatbase on the gate below.

The gate. A builder that cannot show which source passage produced an answer is not listed in the builder group, whatever it scores. This is a gate and not a change to the weights, because the weights are fixed across every page we publish. It excludes Chatbase, whose documentation we read describes matched Q&A pairs but no source display for generated answers. If that changes, it comes back.

We have not run our benchmark on these tools. The benchmark is a support bot built on one 40-page help center, then twenty real customer questions, five of which cannot be answered from that source, scored on grounded answers, correct escalations and invented answers, costed at 2,000 conversations a month, two seats and one help desk integration. When we run it, each tool gets a score out of 100 across six dimensions, weighted in this order: output quality, total cost at realistic usage, job fit, sustained user experience, setup and integration, data and security. Until then there is no score column. The ranking is our judgement on pricing pages, vendor documentation, third-party ratings and dated user reviews.

Disclosure: Agensi sells skills that install into Claude, Cursor, Codex CLI and Copilot, and this page links to our collections. We sell no chatbot platform and no vendor paid to be listed. Reviewed by Samuel Rose, Founder of Agensi. Refreshed on a major release, a price change, or every six months.

Prices and ratings were checked on September 23, 2026 on US pricing pages.

AI for chatbots: ready-made assistants

1. ChatGPT

ChatGPT is first because it is the widest product: text, files, images, voice, research, and the largest ecosystem of connected apps. The free tier gives unlimited text chats on the base model.

Watch: consumer plans use your conversations for training unless you turn off "Improve the model for everyone". Plus has message caps at busy times, and reviewers in September 2026 describe it becoming "unresponsive or behaves oddly" and sounding "more confident than the evidence really supports".

Price: Free. Plus $20 a month. Pro $100; the $200 Pro tier paused new sign-ups on September 10, 2026. G2 4.6 from 3,016 reviews.

2. Claude

Claude is second by a hair and first for long documents and careful writing. Projects on the paid tiers hold your context, and the same plan runs installable skills, which is what the last section of this page is about.

Watch: usage limits. Two reviewers in the same week of September 2026 wrote "the main thing I dislike is the usage limit". Training on consumer accounts is a setting you choose; when it is off, retention is 30 days.

Price: Free. Pro $17 a month annual or $20 monthly. Max from $100. G2 4.6 from 469 reviews.

3. Gemini

Gemini is the pick if your life is in Gmail, Drive and Docs, because it is already there, and the cheapest paid step-up on this page at $4.99.

Watch: the free tier gets "varying access" to the Pro model. Reviewers in September 2026 call answers "inconsistent, especially with complex or multi-step tasks".

Price: Free with a Google account. AI Plus $4.99 a month. AI Pro $19.99. G2 4.4 from 614 reviews.

4. Perplexity

Perplexity is the research assistant: every answer carries numbered citations to the web. It is fourth, not higher, because citations link to pages rather than passages and reviewers report that they are not always right.

Watch: "phantom citations linking to wrong pages", per a reviewer in August 2026. Training is on by default with an opt-out in settings. The free tier allows three Pro searches a day.

Price: Free. Pro $20 a month per the vendor's own announcement. Max $200. Education Pro $10 with student verification. G2 4.4 from 355 reviews.

5. Microsoft Copilot

Copilot is the assistant that comes with Microsoft 365 Personal and Family, and it is on this list for the same reason Gemini is: you may already own it.

Watch: the AI features belong to the subscription owner only and cannot be shared with the family plan's other members. Confirm the G2 listing before quoting it; we saw the rating on a category page rather than the product page.

Price: Included in Microsoft 365 Personal at $9.99 a month; Premium $19.99. G2 4.4 from 361 reviews.

DeepAI Chat is the cheap multi-model option at $9.99 a month with overage at about a third of a cent per message. It has no G2 or Capterra listing, so it is named here and not ranked.

If you use Claude or ChatGPT for writing, the free Humanize Writing skill, in Claude skills for writing and content, turns a robotic draft into prose a person would send.

AI for chatbots: platforms you build on your own data

The gate again, in plain words: if the platform cannot show you which passage of your help center produced an answer, you cannot check it, your support team cannot correct it, and you should not put it in front of customers. Every tool below states what it shows.

6. Google Cloud Conversational Agents

Google's platform, formerly Dialogflow CX, ranks first because it does the most for the least and shows the most: the docs let you configure how many source links and citations appear, expose the snippet of the top source, and return a grounding confidence for each answer. Shows the source: yes, passage snippet plus link.

Watch: it needs an engineer. Three reviewers in August and September 2026 said the same thing about the UI being "confusing" or "overly complicated" once flows get advanced. Billing is per request, not per conversation.

Price: Flows $0.007 per request, Playbooks $0.012, with $600 to $1,000 in trial credit. At about five turns per conversation, 2,000 conversations is roughly $120 a month. G2 4.4 from 144 reviews.

7. Intercom Fin

Fin is the strongest managed agent here and the most expensive. It answers from your Intercom content, runs procedures against other systems, and hands off to a person. Shows the source: yes, inline links in chat, and an answer debugger for the operator listing the content it used.

Watch: hallucination reports have not gone away; a reviewer in June 2026 wrote "I don't like how much it hallucinates and gives people incorrect answers". Fin may use anonymised customer data for fine-tuning unless you opt out. Billing is per outcome, and a handoff to a human counts.

Price: from $0.99 per outcome plus $29 per seat, or a $49 standalone base with 50 resolutions. If half of 2,000 conversations end in an outcome, about $1,050 a month. G2 4.5 from 3,915 reviews.

8. Zendesk AI Agents

If you already run Zendesk, this is the answer, because the AI agent is in every Suite plan and reads the help center you already have. Shows the source: yes, "Display sources for generative replies" is on by default and customers can click through to the articles.

Watch: CSV files are not listed as sources. The per-resolution price is not published; the docs call their figures "placeholders" and point to sales. A reviewer in May 2026 said the AI features "are often paid add-ons and may feel limited".

Price: Suite Team $55 per agent a month paid yearly, with a small resolution allowance per seat. G2 4.3 from 7,079 reviews.

9. Voiceflow

Voiceflow is the builder for a team that wants to design the conversation itself, with a knowledge base, a choice of models, chat and voice. Shows the source: yes, the agent can include the source URL with each reply, and you set how many chunks it retrieves.

Watch: the vendor pricing page is now demo-gated, so the self-serve prices come from G2's copy of it. Billing is credits plus model token charges, and one reviewer in June 2026 wanted "more seats".

Price: Starter free with 100 credits. Pro $60 per editor a month with 10,000 credits. G2 4.6 from 113 reviews.

10. Botpress

Botpress is the developer's builder, with the cheapest entry and the most work. Shows the source: partly. The knowledge agent returns citations, but for documents they reference the page and title only, and showing them to users takes custom work, per the vendor's own community.

Watch: cost. Three reviewers in 2026 said the same thing, one that the pay-as-you-go plan "makes me believe I might overspend".

Price: Free for 25 conversations a month. Plus $150 a month annual with $25 of AI usage included; extra conversations $65 per 100. G2 4.5 from 512 reviews.

11. Tidio Lyro

Lyro is the small-business bot, trained on your FAQ and URLs, inside a live-chat suite with a real free tier. Shows the source: partly. Visitors see a "read more" link under each answer; operator-side attribution is not documented.

Watch: Lyro conversations are billed separately from the chat plan, and the free allowance is 50 for life. A reviewer in July 2026 wrote that costs "get unpredictable on higher-traffic sites".

Price: Lyro from $32.50 a month for 50 conversations. At 2,000 conversations you are on the Plus plan from $300 a month plus usage. G2 4.6 from 1,962 reviews.

Chatbase, popular and cheap at $40 a month for 700 message credits, is excluded by the gate: its documentation describes matched Q&A pairs but no source display for generated answers. Ask the vendor, and if it shows sources, it belongs at about position 9.

Before you build, the RAG Knowledge Base Auditor skill reviews your help center for the gaps, duplicates and stale pages that make a bot invent, and the Customer Support Ticket Resolver skill drafts the escalation logic. Both in Claude skills for business operations. If you are building the bot yourself, read how to create an AI agent first, and browse Claude skills for agents and orchestration for the grounding and escalation procedures.

Voice and phone

Not scored. Retell AI is the best-rated voice agent platform we found, G2 4.8 from 2,638 reviews. Vapi and Bland AI have too few reviews to judge, 3 and 11. We have not priced any of the three, and the pricing model for voice differs from chat, so do not carry any number from this page across.

Free tiers: what you actually get

ChatGPT: unlimited text on the base model, limited uploads and images. Claude: free chat with usage caps. Gemini: free with a Google account, limited access to the Pro model. Perplexity: three Pro searches a day. Copilot: chat with limits, numbers not published. Conversational Agents: no free tier, but $600 to $1,000 of trial credit. Intercom: trial only. Zendesk: 14-day trial. Voiceflow: 100 credits. Botpress: 25 conversations a month. Tidio: 50 chat conversations a month and 50 Lyro answers for life. Enough to build the bot. Not enough to run it.

What users actually report

For each tool we read two to three dated reviews from the last 12 months on G2, retrieved September 23, 2026, and coded each against the six themes in our method. Too small to score, so we do not score it.

On the ready-made side the theme is limits: usage caps on Claude, message caps on ChatGPT Plus, three Pro searches on free Perplexity. On the builder side it is two things. Cost that grows with volume, in the same words at Botpress, Tidio and Intercom. And invented answers, which is the reason the gate exists: "I don't like how much it hallucinates" at Intercom, "phantom citations" at Perplexity. Nobody in the sample complained that a bot was too slow to set up. They complained about what it said and what it cost.

Why one bot grounds and another invents

Disclosure, again: Agensi sells skills, so we have a commercial interest in this argument. Judge it on the mechanism.

Our benchmark has five questions the help center cannot answer, on purpose. They are the test. A bot that grounds says it does not know and escalates. A bot that invents produces a confident paragraph about a refund policy you do not have. The platform decides whether you can see which happened; the source passage display is the difference between a bot you can correct and one you cannot. But two teams on the same platform, with the same 40 pages, get different results, and the platform is not why.

The first team uploads the help center and switches the bot on. The second team writes the procedure: which page wins when two conflict, the exact phrase for "I don't know", the three conditions that trigger a handoff, the tone, what never gets promised. Then it audits the help center for the gaps the bot will fall into. The output changes because the inputs changed, not because the model did. Configuration is the product. The AI agent configuration guide walks through it.

That procedure is what a skill is: a folder of instructions and reference files an agent loads when the job comes up. It installs into Claude Code, Cursor, Codex CLI or Copilot in about 30 seconds, and every listing on Agensi passes a security scan before it goes live. We have not yet run the benchmark as a scored before-and-after. When we do, the result goes here with the date, the version, and what each platform did with the five questions it could not answer.

How to choose

Decide which half of the page you are on. If a builder, count your monthly conversations honestly and price the top three at that volume, not at the entry price. Ask each vendor to show you a source passage for one answer, live. Then check the help desk integration last, because it is the one that decides whether the bot's handoffs land anywhere: native if you are on Intercom or Zendesk, through a workflow automation skill if you are not.

Mistakes to avoid

Buying on the demo. A demo runs on a clean help center. Yours has a 2023 pricing page still live.

Pricing at the entry tier. 2,000 conversations is $120 on one platform and $1,050 on another.

Turning off source display to make the bot look cleaner. Then nobody can check it.

Putting customer data into a consumer assistant that trains on it. ChatGPT's consumer plans and Perplexity do by default.

Skipping the security read on anything the bot can act through. Claude skills for security holds the red-team and audit skills for it.

The four skills named above are a start, and every one of them sits in a collection linked from its chapter.

Samuel Rose is the founder of Agensi, a marketplace for AI agent skills built on the SKILL.md open standard. Published and last updated September 28, 2026.