Best Conversational AI Platforms in 2026, Compared With Real Prices

The first table gives you a starting point by situation. The second compares every platform on one screen. Everything after that is the reasoning, the arithmetic and the per-platform detail.

Quick picks by situation

Conversational AI platforms compared side by side

Prices were checked on each vendor's own page, and on AWS, Azure and Google Cloud Marketplace, on September 18, 2026. Ratings are stars and review count; "whole company" means G2 has no listing for the agent product alone, so we show the company's. On Gartner, "Not in top 20" means the product isn't among the 20 most-rated on the conversational AI platforms market page and we didn't look it up further.

What a conversational AI platform is, and what it isn't

Gartner defines conversational AI platforms as platforms "primarily used for developing applications simulating human conversation across multiple channels", which is broad enough to hold all six types above. A working definition has three parts: a way to define what the agent knows and may do, a runtime that holds the conversation on one or more channels and calls your systems, and tooling to test, monitor and improve it. The things the term gets confused with each have one or two of those parts.

Our overview of AI agents for customer service goes deeper on the support slice. This page stays at the category level.

The six kinds of conversational AI platform

The type decides who builds the agent and what you pay for. Here are the six, including the internal-use and outbound buyers the ranking listicles mostly skip.

Three questions sort most buyers in under a minute. Is the conversation with customers or with employees? Is it mostly on the phone or mostly in text? And will engineers own the agent after launch? A customer-facing, text-first support team without engineers lands in the support-agent row. An IT team fielding password resets in Teams lands in the last row. A voice-first team with developers lands in the voice-API row.

The architecture choice underneath: flows or prompts

Most platforms here offer some mix of two ways to define an agent. In a flow, you draw the conversation as a graph of intents, steps and branches, and the agent follows it. In a prompt (Google and Voiceflow call it a playbook, Botpress an autonomous node), you write instructions in plain language, attach tools and knowledge, and let a large language model decide each turn.

Google makes the difference visible because it prices the two separately. Its Conversational Agents product has two editions: Flows, "also known as Dialogflow CX", deterministic agents built on natural language understanding, and Playbooks, generative agents built from instructions. A chat request costs $0.007 on Flows and $0.012 on Playbooks, and voice costs $0.001 or $0.002 a second, according to Google's pricing page.

Retell puts the same choice in front of you when you create an agent: "Single prompt" or "Conversational flow", which Retell labels production-ready and deterministic.

A flow is predictable and auditable: you can point to the branch that sent a customer to a refund and prove it only fires after an order lookup. It also takes weeks to build and breaks when customers say things nobody drew a branch for. A prompt handles unplanned phrasing on day one and takes an afternoon to write, but "never promise a refund over $200" becomes a sentence you have to test for, not a branch you can inspect. Our own test below shows the risk. Regulated voice lines (payments, identity checks) still tend to run on flows. Support chat and email increasingly run on prompts with tool permissions and a test set. Ask which of the two a platform treats as the default, because that's the one your team will maintain.

How conversational AI platforms charge, and whose incentive each model serves

Learn the billing unit before the demo, because it decides what you pay for when the agent fails.

Outcome pricing sounds like the vendor only earns when you win, and partly that's true, but it depends on what counts. On Fin's pricing page, a resolution, a "procedure handoff" to a human and a disqualification each bill at $0.99, and a qualification bills at $9.99, ten times a resolution. So a well-built handoff is billable, and a qualified sales lead is the most valuable event on the menu. Sierra's explainer on outcome-based pricing lists resolved conversations, saved cancellations and upsells as outcomes, so the vendor also earns on revenue events.

Credit packs hide the unit one level down. Microsoft's billing rates page charges 1 credit for a scripted answer, 2 for a generative answer, 5 for an agent action, 10 a minute for classic voice and 35 a minute for generative voice. The same conversation costs more the more it does, and when a tenant hits 125% of its prepaid capacity, Microsoft disables its custom agents until capacity is added.

Per-minute pricing has the opposite blind spot. It's honest about cost and tells you nothing about whether the call went well, so the incentive to make calls shorter sits entirely with you.

What 3,000 conversations a month costs, platform by platform

The ranking pages show starting prices, which don't tell you what a real month costs. Here is one volume per channel, worked through on every platform that publishes enough to work it, using each channel a vendor prices and the cheapest plan or pack combination that covers the volume.

Chat assumptions: 3,000 conversations a month, 8 customer turns each (24,000 turns, and 24,000 agent replies), half resolved by the agent (1,500). Voice assumptions: 3,000 calls averaging 4 minutes (12,000 minutes, 24,000 turns). Prices checked September 18, 2026. The chart plots every row in the tables below it.

Chat, from published prices

Chat, vendor numbers with a caveat

Chat, annual contracts listed on AWS Marketplace

Chat, estimates from outside the vendor

Voice, 12,000 minutes, from published prices

Voice, caveats and contracts

The cloud rows look cheap because the price leaves out the people: a Google Flows agent at $168 a month still needs engineers to design, test and maintain the graph, and Lex's $18 buys a runtime with no channel attached. The ElevenLabs chat row is a floor for the same reason in a different place: its per-message fee comes before the model. Retell's per-message rate includes the model, but the $12 assumes its smallest one; the same month on GPT 5.4 mini is $180. For voice, four of the eleven published rows come in under $1,000, six sit between $1,000 and about $1,740, and Copilot Studio's generative voice is the outlier at $3,400, so the choice turns on control, latency and the LLM more than list price. The two annual contracts on AWS cost two to three times the published rows at this volume, and cover a lot more volume than 3,000 chats. Rerun the arithmetic with your own turn counts and resolution rate before comparing quotes.

How we researched this, and the agent we built twice

We pulled the live Google results for "conversational ai platform" on September 18, 2026, and read the pages ranking in the top ten: listicles from Retell, Moveworks and Cekura, the ElevenLabs, LivePerson, Sierra and Voiceflow pages, AWS's definition page, and Gartner's Peer Insights market page, which we read in a browser. Retell wrote the page that ranks first and lists itself first. The same day we checked every vendor's pricing page, searched AWS Marketplace, Azure Marketplace, Google Cloud Marketplace and Vendr by vendor name, read G2 and Gartner in a browser, read Trustpilot, and checked the acquirers' newsrooms for ownership changes.

We built the same minimal agent on two platforms, ElevenLabs Agents and Voiceflow, on their free plans that day. It's a support agent for Northwind Candles, a fictional shop: a short prompt with shipping times and a 30-day return rule for unused items, no tools, no order lookup, and an instruction never to promise a refund. On each we sent the same two messages: "my order hasn't arrived, can I just get a refund?", then an order number plus a question about returning a candle that had been burned once. Neither agent had a tool, and we ran text turns only, no voice turn on either. Both free plans offer tools and ElevenLabs is a voice product, so that's a limit of our test, not of the products. We did not build an agent on any of the other platforms here. Botpress's sign-up rejected the throwaway address we used, Retell's asks for a work email behind a CAPTCHA, and Vapi's form didn't accept our input, so we stopped there. The remaining screenshots come from vendor pages, marketplace listings and public documentation.

Both agents handled the policy questions correctly. Neither promised a refund, and both refused the burned candle under the unused-items rule. Both also said something untrue. ElevenLabs replied "I have noted your order number 4417", and Voiceflow said "I've noted order 4417, and a teammate will check on it". Neither had a tool to note anything. The prompt told the agent a human would follow up, and the model narrated that follow-up as if it had already happened. ElevenLabs's automatic summary then recorded that the agent "offered to escalate the order status to a human teammate". In production, that summary would describe a handoff with no ticket behind it.

Both products log each conversation in detail. ElevenLabs showed the conversation at 60 credits, $0.00195 of LLM cost and a 618 ms first reply. The default model on a new ElevenLabs agent was Qwen3.5-397B, not an OpenAI or Anthropic model.

Publishing a change on ElevenLabs opens a review screen with a diff of the prompt and a commit message, so every edit is versioned and you can see what changed before it goes live.

Voiceflow pre-fills a structured prompt (Role, Goal, Tone, Guardrails) and shows an estimated cost and latency for the chosen model above the editor: about $0.017 a message and 1,700 ms on Voiceflow Core 4.2. The run log billed the reply that claimed to note the order at $0.005 and timed it at 1.1 seconds to first text, and the whole test used $0.03 of the $1.00 free credit.

The lesson applies to every prompt-based platform on this page. Your test set has to check claimed actions against actual tool calls, because a well-behaved agent will describe work it never did.

How to evaluate a conversational AI platform

Channel and builder narrow the list. These seven checks separate what's left.

1. Compliance, by plan. SOC 2 and ISO 27001 are common; HIPAA, PCI and SSO are often reserved for the top tier. Voiceflow's in-app plan table shows SOC 2 on every paid plan but HIPAA, SSO and PII redaction on Enterprise only. IBM's HIPAA-ready option sits on watsonx Orchestrate Premium, per its pricing page.
2. Languages. Ask for the list of languages the agent is tested in, and whether one agent switches languages or you build one per language.
3. Latency, for voice. Measure time to first word on your own calls. The reply times our text tests logged ran from 0.3 seconds (ElevenLabs's model time on one reply) to 1.9 seconds (Voiceflow finishing a reply), and a voice turn adds speech recognition and synthesis on top.
4. Testing and evaluation. Look for simulated conversations, a way to turn a bad transcript into a test, and versioned publishing, so a prompt change can be reviewed and rolled back.
5. Human handoff. Check what the agent passes to a person: the transcript, the collected fields, and the reason. Our test shows why this matters, since an agent can claim a handoff that never happened.
6. Analytics you can act on. Resolution rate by topic, cost per conversation, and the conversations the agent got wrong, not only volume charts.
7. Data residency and deployment. Some buyers need EU hosting, a private cloud or on-premises. Kore.ai's AWS listing says it deploys in the cloud or on-premises, and Bland's Enterprise plan offers on-prem or VPC.

Implementation time runs from an afternoon for a prompt agent to months for an enterprise flow program with integrations. Ask the vendor who does the work, and budget for it separately from the license.

Enterprise CX platforms

These are suites for organizations with many use cases, a platform team or integrator, and procurement that expects an RFP.

Kore.ai

Kore.ai sells customer-facing and employee-facing agents on one platform, built by your developers or an integrator. Its docs confirm Automation AI bills in 15-minute sessions, so a 31-minute call is three. The website quotes privately, but AWS Marketplace carries two usable numbers. Its general AI for Service listing charges $0.20 a session on top of a 12-month contract whose listed dimensions are $40,000 for enterprise support and $40,000 for a non-production environment. Its AI For Banking listing sells BankAssist in monthly tiers from $350 for 1,000 sessions to $17,500 for 50,000, $0.35 a session at every tier, with support included.

Kore.ai says it was named a Leader in Gartner's 2026 Magic Quadrant for Conversational AI Platforms and in Forrester's Q2 2026 Wave for customer service, per its analyst page. Peer ratings split: 4.6 from 505 on G2 but 4.3 from 167 on Gartner Peer Insights, the most-rated product in Gartner's market. G2's leading con tags are "Learning Curve" (46), "Usage Limitations" (43) and "Slow Performance" (39), and a junior developer at an enterprise wrote that "the platform can occasionally present unexpected bugs or inconsistent behavior" (G2, July 27, 2026).

Skip it if you have one use case or no platform team: at $40,000 a year before the first session, the contract pays off only when several departments share it. More in our Kore.ai guide.

NiCE Cognigy

The best peer rating among the enterprise suites belongs to NiCE Cognigy: 4.8 from 156 on Gartner Peer Insights as of September 18, 2026, up from the 138 NiCE cited in January. G2 has only 13 reviews (4.6). NiCE closed its acquisition of Cognigy on September 8, 2025, per NiCE's announcement, which says Cognigy stays available standalone as well as inside CXone. The homepage shows Toyota, Lufthansa and DHL and says Cognigy was named a Leader in the Forrester Wave 2026 for conversational AI platforms for customer service.

The website quotes privately, but Cognigy's AWS Marketplace listing prices a Basic tier at $43,080 for 12 months, covering platform, setup, standard support and 60,000 conversations a year, which is about $3,590 a month or $0.72 a conversation at full use. Adding five Voice Gateways takes it to $53,916, and an enterprise license for up to 10 million conversations lists at $1,000,000 plus a $200,000 ramp.

The few G2 complaints are about depth: an assistant manager at a mid-market company listed "limited analytical possibilities, less possibilities of advanced chat flow" (G2, September 18, 2024). It fits a contact center running voice and digital channels at scale, particularly one already on NICE. Skip it if your volume is well under 5,000 conversations a month, since the smallest listed tier still costs more than most per-use rows in our model. More in our Cognigy guide.

Yellow.ai

Of the enterprise suites, only Yellow.ai lets you start without a sales call. Its pricing page offers a Free plan with one AI agent and 500 chat sessions a month, then $0.99 a resolution, with Enterprise by quote, and its AWS listing adds annual plans at $10,000 (Basic, one use case) and $25,000 (Standard, four), each with tiered conversation limits the listing leaves unstated. On August 3, 2026, Yellow.ai announced a plan to go public through a $550 million merger with Bluerock Acquisition Corp., per its PR Newswire release.

It rates 4.4 on both G2 (107) and Gartner (102), and multilingual, multichannel deployments are its home ground. G2's con tags are small but consistent: "Steep Learning Curve" (7), "Learning Curve" (6) and "Difficult Implementation" (3). The sharpest recent review is about service: a small-business owner wrote that "it's extremely difficult to contact them, and you can't speak with anyone" (G2, July 31, 2026).

The free plan is a real way to test the builder, but one agent and per-resolution overage make it a trial, not a production tier. Skip it if you need a named support contact from day one and can't negotiate one into an Enterprise contract. Our Yellow.ai guide has the third-party estimates for Enterprise.

LivePerson

LivePerson's homepage ranks on page one for this search. SoundHound AI completed its acquisition of LivePerson on September 4, 2026, and SoundHound already owns Amelia, so two of the older enterprise platforms now share an owner. Its strength is messaging at brand scale. The pricing page names Bronze, Silver and Gold plans with no figures, and its Azure and Google Cloud listings route to sales. G2 gives 4.3 from 207, with "Learning Curve" (21) and "Expensive" (11) among the cons, and its homepage cites a Niche Player placement in Gartner's 2025 Magic Quadrant. Ask about the roadmap after the acquisition before a multi-year term.

Cloud components from Google, Amazon and Microsoft

Their per-request prices are among the lowest here, and they're the most work per agent. You assemble the product from parts.

Google Conversational Agents

No platform here prices more transparently than Google Conversational Agents (formerly Dialogflow CX). You build in Google Cloud's console, either as a Flows graph or as Playbooks written in plain language, and the per-request and per-second rates make both cost tables above reproducible. New users get $600 of Flows credit and $1,000 of Playbooks credit for 12 months, per the pricing page. Channels, handoff, analytics and the website widget are yours to assemble or buy.

G2 shows 4.4 from 144 reviews for Google Cloud Dialogflow, with "Cost" (5) and "Learning Difficulty" (4) among the cons, and Gartner Peer Insights shows 4.4 from 19 ratings for Google in this market. The cost complaint is about visibility more than the rate: a small-business reviewer in insurance wrote that "it's hard for me to get a clear early warning when chatbot traffic starts driving up cloud bills" (G2, August 21, 2026). Set a budget alert in Google Cloud before launch.

It suits a team with Google Cloud engineers who want to own the graph and can live with building everything around it. Skip it if nobody on the team writes code: the $168 month in our model assumes the engineers are already on payroll, and the new-user credit expires after 12 months whether you've launched or not.

Amazon Lex and Amazon Connect

Amazon splits the job between two products. Lex is the conversation engine, at $0.00075 a text request and $0.004 a speech request per its pricing page, and new AWS customers get up to $200 of Free Tier credits. Connect is the contact center around it: $0.010 a chat message sent or received and $0.038 a voice minute, per Connect's pricing, which presents its AI features as part of those rates, with a cheaper "Customer Basic" option for teams not using them. Your AWS engineers build it, and the data stays in your AWS account.

Lex rates 4.4 from 52 on Gartner Peer Insights and 4.2 from 43 on G2. Reviewers keep landing on maintenance: an IT and cloud infrastructure manager at a small business wrote that "customizing complex conversational flows can be challenging" (G2, June 29, 2026).

The appeal is metering you can model exactly and no new vendor to add to procurement. The cost is engineering: Lex's $18 month buys no channel, no analytics and no handoff until you build them, and Connect's $480 chat month assumes you want Amazon's contact center, not your current one. Skip it if your team isn't already on AWS or your support conversations live in a help desk you plan to keep.

Microsoft Copilot Studio

Microsoft sells Copilot Studio as $200 a month per pack of 25,000 Copilot Credits, per Microsoft's pricing page, with a pay-as-you-go meter as an alternative, and it burns credits at different rates by feature, as covered above. Makers build in a low-code studio inside the Microsoft stack. It serves two buyers. For employees, usage is included for people licensed for Microsoft 365 Copilot. For customers, it's a credit meter, and the voice mode you choose decides the bill: classic voice at 10 credits a minute comes to $1,000 for our 12,000 minutes, generative voice at 35 comes to $3,400.

Ratings: 4.4 from 155 on G2 and 4.2 from 71 on Gartner, with "Limited Customization" (11) and "Expensive" (11) among G2's cons. A freelance AI app developer wrote that "there isn't much transparency or any real version control, and it also doesn't seem to handle more complex tasks very well" (G2, May 23, 2026).

The watch-out is enforcement: at 125% of prepaid capacity, custom agents are disabled until someone adds credit, so set per-agent caps before launch. Skip it if you're not a Microsoft 365 shop, because most of its advantage is the identity, Teams and data connections you already pay for.

Agent builders

Voiceflow

Voiceflow mixes playbooks for LLM-driven tasks with workflows for scripted steps in one visual builder, and it ships Zendesk, Salesforce, Shopify and HubSpot among its global tools. One technical builder on a CX team can run it. Its public pricing page routes businesses to a demo, but the plan screen inside a free account shows Pro at $60 a month and Business at $150 ($54 and $135 billed annually), each including the same amount in usage credit. Enterprise is a different scale: Vendr's buyer data puts the median Voiceflow contract at $258,625 a year.

G2 rates it 4.6 from 113 reviews, with "Missing Features" (25) and "Usage Limitations" (24) leading the cons; Trustpilot shows 4.1 from 15, and Gartner lists it with no ratings yet. Price is the recurring complaint from small teams: a forward-deployed engineer wrote that "the price was too high for a chatbot I need to develop myself" (G2, September 1, 2026), and a senior software engineer wanted more than two seats on the $60 plan (G2, June 16, 2026).

In our test the builder was the clearer of the two about cost before anything ran, and its log priced each reply. It fits a CX team that wants chat and voice from one canvas and doesn't need HIPAA or SSO, which are Enterprise-only. Skip it if several people need to edit agents on a small budget, because seats, not usage, will set the price. Our Voiceflow guide covers the product in depth.

Botpress

Botpress publishes its prices, which puts it ahead of most of this list for a buyer trying to budget. Plus is $150 a month and Team $750 a month (both billed annually), including 250 and 1,500 conversations and $25 and $150 of AI usage; extra conversations come in packs of 100 at $65 on Plus and $50 on Team. G2 shows 4.5 from 512 reviews, with "Learning Curve" (59) and "Poor Documentation" (28) among the cons. Its sign-up rejected our disposable address. More in our Botpress guide.

Voice-agent APIs

Retell AI

Developers build Retell AI agents through its API and dashboard, as a single prompt or a conversational flow. Its pricing page advertises $0.07 to $0.31 a minute, and the live calculator defaults to $0.115 on GPT 4.1 ($0.045 LLM, $0.055 voice infrastructure, $0.015 text-to-speech) before telephony. Chat is priced per AI message by model: the pricing card headlines $0.002, but the table starts at $0.0005 on GPT 5 nano, then $0.001 on GPT 4.1 nano and $0.002 on GPT 5.4 nano. At $0.0005 our 24,000 replies cost $12 with the model included, the lowest chat row in our cost tables. New accounts get $10 of credit and 20 concurrent calls, and sign-up asks for a work email behind a CAPTCHA.

It has the largest review base of the voice APIs: 4.8 from 2,638 on G2 and 4.9 from 814 on Trustpilot, against a single rating (4.0) on Gartner. The complaints are mild and about growth: a retail reviewer who set up agents for two companies wrote that "pricing can also grow depending on call volume" (G2, June 26, 2026). Retell also wrote the page ranking first for this keyword, so read its ranking of itself with that in mind.

Skip it if nobody on your team will work in an API: the chat price is low, but the product, the docs and the calculator are built around developers shipping phone agents, and a text-only team without one gets more from an agent builder.

Vapi

Vapi is the modular option: you pick the transcriber, the model and the voice, and Vapi adds a $0.05 a minute hosting fee to the provider costs it passes through (the same $0.05 appears on its AWS listing). Its calculator estimates $82 to $129 for 1,000 minutes, and plans run from $29 a month (Core) to a $999 monthly minimum (Pro), per Vapi's pricing page. Reviews are thin and poor: 3.0 from 3 on G2 and 2.4 from 15 on Trustpilot, where one reviewer wrote that "VAPI has cost us $50k in damages so far from downtime caused by bugs". Load-test it before live calls.

ElevenLabs Agents

ElevenLabs Agents comes from the text-to-speech company, and voice quality is the reason to pick it. You build an agent in its dashboard or through the API, pick the model (our test agent defaulted to Qwen3.5-397B) and a voice, and deploy it to phone, web voice or chat. Its agents pricing tab bundles call minutes into plans: Pro is $99 for 1,238 minutes, Scale $299 for 3,738 and Business $990 for 12,375, and every plan charges $0.08 for each extra minute, so at our 12,000 minutes Pro plus extra minutes (about $960) undercuts Business. Text messages cost $0.003 each and start on the $6 Starter plan, which makes our $78 chat floor. The LLM and telephony are billed on top.

G2 has no separate listing for the agents product. The company as a whole rates 4.5 from 1,213 reviews there, with "Expensive" (171), "Needs Improvement" (163) and "Pricing Issues" (148) among the cons; Trustpilot shows 3.0 from 1,150, where billing is the sore point, and Gartner Peer Insights lists ElevenLabs only in its AI Dubbing market (4.5 from 17). The complaint that matters most for a support line is about reading data aloud: a software engineer at a mid-market insurer wrote that "the engine mispronounces complex alphanumeric policy numbers unless we explicitly format them phonetically" (G2, September 17, 2026). Order numbers raise the same problem, so test them.

It fits a team that wants a natural-sounding phone agent and has a developer to wire up the model, telephony and any tools. In our test its logs showed the cost and model time of every reply, and publishing went through a reviewed diff. Skip it if your conversations are text-only: you'd be choosing a voice company's product and pricing for the one channel it isn't built around.

Bland

Outbound sales and appointment calling is a sub-market of its own, and Bland is built for it. It publishes a Start plan at $0.14 a minute with no platform fee and a 100-call daily cap, and a Build plan at $299 a month plus $0.12 a minute, with LLM, transcription and voices included, per its pricing page. G2 shows 5.0 from 11 reviews, too few to lean on. For outbound, compliance with calling rules is on you whichever platform you choose.

Synthflow, on both the Retell and Cekura lists, now quotes enterprise contracts "starting at $30,000 annually" on its pricing page.

Support AI agents

Fin

Salesforce now owns Fin, the product formerly sold as Intercom Fin: it completed the acquisition on September 10, 2026, and says Fin continues to work with the help desks customers already use. You configure it from your help center content and procedures rather than building it from scratch. At $0.99 an outcome with a 50-outcome monthly minimum on help desks other than Intercom, it's the most transparent of the support agents, and it's also sold through AWS Marketplace as usage-based Intercom credits at $0.01 each.

G2 shows 4.5 from 3,912, though that covers the whole Intercom product, and Gartner Peer Insights shows 4.6 from 18 for Fin AI Agent. The recent complaints are about long conversations: a customer support specialist at a mid-market company wrote that "sometimes Fin can be repetitive in the way it handles a conversation, which can make the interaction longer than necessary", leaving a long history for the human who takes over (G2, September 11, 2026). A mid-market software reviewer called its workflow builder "very stiff" (G2, August 25, 2026).

Remember the billing menu too: a Procedure you configured to end in a handoff bills like a resolution, at $0.99, while a conversation simply passed to your team without an outcome is not charged. Skip it if most of your volume needs multi-step actions in back-office systems, or if you can't live with the outcome definitions on its pricing page. Our Fin guide predates the acquisition but covers the product.

Decagon

Decagon keeps prices off its site. Its AWS listing has one 12-month "Platform Access" dimension, "Pricing determined by private offer", shown at $1,000,000, and its Google Cloud listing says to contact the vendor. The neutral number is Vendr's: a median contract of $432,750 a year, ranging from $105,000 to $923,183. Quiq, a competitor, estimates a $50,000 platform fee plus usage, which now looks like a floor. G2 shows 4.8 from 30; an enterprise developer wrote that "it can still struggle with more complex or unusual customer questions" (G2, August 10, 2026). It's priced for budgets of six figures a year. See our Decagon guide.

Sierra

Sierra's homepage also ranks on page one for this search. Its AWS listing says "pricing is determined via private offer", with a $1,000,000 placeholder and outcome-based terms. G2 shows 4.5 from 130; a small-business CTO wrote that "customizing conversation flows for edge cases sometimes requires going back to their team" (G2, August 27, 2026). The negotiation over what counts as an outcome is where the real price is set. It's the wrong pick if you want to edit the agent yourself the same afternoon. Our Sierra guide goes deeper.

Parloa

Parloa is the voice-first entry in this type, built for contact-center telephony. Its homepage lists ISO 27001, SOC 2 Type 2, PCI DSS, HIPAA and DORA, and shows BarmeniaGothaer among its customers. It rates 4.5 from 37 on G2 and 4.5 from 48 on Gartner Peer Insights. Its Azure listings are $1 private-offer templates, and Vendr shows a median of about $5.7 million a year without saying how many purchases that rests on, so treat it as a sign of deal size only. Expect no number before a sales call. Our Parloa guide collects the estimates. Ada is covered in our Ada guide.

Macha

Macha belongs in this type with a narrower scope. It's an AI agent layer that runs inside the help desk you already use (Zendesk, Freshdesk, Gorgias, Front, Intercom or HubSpot today), reading tickets, replying and taking actions in your connected tools. You wouldn't use it to build a phone IVR or a sales agent for your website. It fits teams already running one of those help desks whose conversations are support tickets, and who want the agent working inside the ticket rather than in a new channel beside it. Pricing is one plan set by monthly ticket volume, from $299 for 750 tickets (about $0.40 a ticket), and each thread with one person counts once however many replies it takes. Setup and monitoring by the Macha team are included on every plan, which answers the "who builds it" question for teams without engineers. See how Macha works on your help desk.

Employee and IT agents

Google's AI Overview for this search asks whether you need conversational AI for customer support or internal use. For internal use, the buyer is IT, the channel is Slack or Teams, and the work is password resets, access requests and HR questions.

Moveworks (ServiceNow EmployeeWorks)

Moveworks is the specialist. ServiceNow completed its acquisition on December 15, 2025, per ServiceNow's release, and the product now appears as ServiceNow EmployeeWorks, as the Moveworks homepage shows. IT teams deploy it with the vendor into Slack, Teams and a portal, where it answers questions and files requests across business systems. It's one of the few employee platforms with a price you can read: its AWS Marketplace listing charges $150 per user a year on a 12-month contract for 1,000 to 2,500 users, so the smallest band starts at $150,000 a year. Vendr's buyer data puts the median at $130,000 a year across 31 purchases.

G2 rates EmployeeWorks 4.4 from 128, with "Complexity" (5), "Usage Limitations" (4) and "Delayed Response" (3) among the cons, and Gartner Peer Insights shows 4.4 from 40 ratings, listed under ServiceNow. The harshest recent review came from an enterprise in defense and space: "It seems to need heavy backend support and a lot of lifting just to get things working, even for small tasks like placing requests" (G2, May 5, 2026).

If you run ServiceNow, the acquisition is a tailwind; if you don't, you're buying into ServiceNow's roadmap. On February 25, 2026, ServiceNow announced FedRAMP Moderate authorization for Moveworks, which matters for US public-sector buyers. Skip it for customer-facing conversations, and below about 1,000 employees, where the per-user minimum outweighs the gain. Our Moveworks guide covers the deal and what changed.

IBM watsonx Orchestrate

IBM is one of the few employee-platform vendors that publishes prices. watsonx Orchestrate Essentials starts at $530 a month and Standard at $6,360 a month, with Premium (data isolation and a HIPAA-ready option) by quote and a 30-day trial, per its pricing page. Its AWS listing sells the same Essentials price, $6,360 a year, as 600 "Resource Units" a month without saying how many conversations a unit buys. The product is aimed at work that crosses HR, procurement and IT systems, where an agent acts in those systems on an employee's behalf.

Ratings are steady: 4.4 from 393 on G2 and 4.3 from 120 on Gartner Peer Insights, the third most-rated product in Gartner's market. G2's cons lead with "Learning Curve" (22) and "Complexity" (17), and the recent reviews say where the time goes. A small-business reviewer in accounting wrote that "getting from an empty environment to a first working custom tool involves more moving parts than it should", adding that their first project lost days to it (G2, September 3, 2026).

It fits enterprises, especially IBM and SAP shops, automating employee work across several back-office systems, with a team that can spend the first weeks on setup. Skip it if you need something running this week with nobody to own the setup: the reviews describe days lost before the first custom tool worked.

Copilot Studio (above) is the default for Microsoft 365 organizations, since employee use is included for licensed users. Aisera, now owned by Automation Anywhere, rates 4.3 from 101 on Gartner; see our Aisera guide.

Contact-center suites and also considered

If you're replacing the contact center itself, Genesys Cloud CX (4.6 from 79 on Gartner), Talkdesk (4.5 from 53) and Sprinklr Service (4.6 from 105) bundle conversational AI with routing and workforce tools; buy them for the suite, not the agent.

Also considered from Gartner's top 20 by rating count: Omilia (4.7, 77), a voice specialist for banks and telcos; boost.ai (4.7, 71), strong in regulated Nordic banking, with our boost.ai guide; Amelia (4.2, 71), now SoundHound AI; CBOT (4.8, 55), OneReach.ai (4.5, 53), Tars (4.5, 44), DRUID AI (4.8, 43) and Avaamo (4.7, 41), each with a narrower regional or vertical footprint than the entries above.

Who should not buy each type

An enterprise CX platform is the wrong choice if you have one use case and no engineers; you'll pay for breadth you never deploy. Cloud components are wrong if nobody will build the channel, handoff and analytics around them. An agent builder is wrong if nobody will own the canvas after launch, because agents drift as policies change. A voice API is wrong for a text-only support team with no developer, since you'd be assembling a product out of parts. A vendor-built support agent is wrong if you need to change behavior weekly without filing a request. An employee agent is wrong for customer conversations. And Macha is the wrong fit if you don't use a help desk, or if the conversations you want automated are sales calls and phone IVR.

Frequently asked questions

What is the best conversational AI platform? It depends on the channel and who builds. Among the products here with 500 or more G2 reviews as of September 18, 2026, Retell AI rates highest (4.8 from 2,638) for voice, then Kore.ai (4.6 from 505) for enterprise builds, with Botpress (4.5 from 512) and Fin (4.5 from 3,912) close behind. On Gartner Peer Insights, NiCE Cognigy leads the most-rated products at 4.8 from 156. Pick the type first, then compare within it.

What is a conversational AI platform? Software for building, running and monitoring AI agents that hold conversations over chat, messaging or voice and connect to your systems. Gartner's definition covers applications "simulating human conversation across multiple channels".

Is there an AI that I can just talk to? Yes. Consumer apps such as ChatGPT, Google Gemini and the Copilot app for personal Microsoft accounts have voice modes you can talk to on a phone or computer, with no setup. If you mean an agent your customers can call or chat with about your business, that's what the platforms on this page build; the voice-agent APIs (Retell, Vapi, ElevenLabs Agents) are the quickest route to a phone line.

What is the difference between a chatbot and a conversational AI platform? A chatbot is one agent deployed on one channel. A conversational AI platform is the tooling to build, run, test and monitor many agents across chat, messaging and voice, and to connect them to your systems.

What are the 5 major AI platforms for conversational AI? If you mean the largest companies, they're Google (Conversational Agents), Amazon (Lex and Connect), Microsoft (Copilot Studio), IBM (watsonx Orchestrate) and Salesforce, which now owns Fin. All five publish prices for these products, from Google's $0.007 a request to Fin's $0.99 an outcome.

What does Gartner say about conversational AI platforms? Gartner publishes a Magic Quadrant for Conversational AI Platforms. Kore.ai says on its analyst page that it was named a Leader in the 2026 edition, and LivePerson's homepage cites a Niche Player placement in the 2025 edition. Forrester's 2026 Wave for conversational AI platforms for customer service names Kore.ai (Q2 2026, per the same analyst page) and NiCE Cognigy (per Cognigy's homepage) among its Leaders. On Gartner's Peer Insights market page, the most-rated products are Kore.ai Artemis (4.3 from 167), NiCE Cognigy (4.8 from 156) and IBM watsonx Orchestrate (4.3 from 120), read on September 18, 2026.

Is there a free conversational AI platform? Several have free tiers or credits: Voiceflow ($1.00 of credit), ElevenLabs Agents (Free plan), Yellow.ai (500 chat sessions a month), Botpress Free, Retell ($10 of credit), Vapi ($5), Google ($600 Flows and $1,000 Playbooks credit for 12 months), AWS (up to $200 of Free Tier credits for new customers) and a 30-day watsonx Orchestrate trial. Macha offers $50 of free usage with no card. None covers production volume.

How much does a conversational AI platform cost? At 3,000 chat conversations a month, published prices run from about $12 (Retell chat on GPT 5 nano, its cheapest model) to about $1,500 (Botpress Team), with Amazon Lex's runtime alone at $18 and ElevenLabs chat at $78 before the model. Annual contracts listed on AWS Marketplace start at $6,360 a year (IBM watsonx Orchestrate Essentials) and $10,000 (Yellow.ai Basic), but neither listing says how many conversations that buys. The two that do work out to about $3,590 a month (NiCE Cognigy, 60,000 conversations a year) and $3,933 (Kore.ai AI for Service with its support contract). Voice at 12,000 minutes runs from about $96 (Lex, before telephony) to $3,400 (Copilot Studio's generative voice) on published prices, and four of the eleven published voice rows cost under $1,000; Cognigy's annual contract with Voice Gateways works out to about $4,493 a month. Decagon's buyers pay a median of $432,750 a year, per Vendr.

Can I download a conversational AI platform and run it myself? Most are cloud services. Kore.ai's AWS listing says it deploys in the cloud or on-premises, and Bland's Enterprise plan offers on-premises or VPC deployment. For the rest, expect a hosted service with regional data options on higher tiers.

What is the 30% rule in AI? There's no standard definition; the phrase gets used for several different rules of thumb. For a conversational AI rollout, the number that matters is your own resolution rate on a sample of real conversations, measured before you sign, so start with one category and count.

Who owns Cognigy, Fin, Moveworks and LivePerson now? NiCE completed its purchase of Cognigy on September 8, 2025. ServiceNow completed its purchase of Moveworks on December 15, 2025. Salesforce completed its purchase of Fin, formerly Intercom, on September 10, 2026. SoundHound AI completed its purchase of LivePerson on September 4, 2026.

If your conversations are already tickets in Zendesk, Freshdesk, Gorgias or Front, start a Macha trial with $50 of free usage and point an agent at one ticket category first.

Sources: Google Conversational Agents pricing, Amazon Lex pricing, Amazon Connect pricing, Microsoft Copilot Studio pricing, Microsoft Learn, Copilot Studio billing rates, IBM watsonx Orchestrate pricing, Kore.ai AI for Service on AWS Marketplace, Kore.ai AI For Banking on AWS Marketplace, Kore.ai usage plan docs, Kore.ai analyst recognition, NiCE Cognigy on AWS Marketplace, NiCE, closes acquisition of Cognigy, NiCE Cognigy, Yellow.ai pricing, Yellow.ai on AWS Marketplace, Yellow.ai merger release, PR Newswire, LivePerson, LivePerson pricing, SoundHound AI, completes acquisition of LivePerson, Voiceflow pricing, Vendr, Voiceflow, Botpress pricing, Retell AI pricing, Vapi pricing, Vapi on AWS Marketplace, Twilio US voice pricing, ElevenLabs Agents pricing, Bland pricing, Synthflow pricing, Fin pricing, Fin on AWS Marketplace, Salesforce, completes acquisition of Fin, Decagon on AWS Marketplace, Vendr, Decagon, Quiq, Decagon pricing, Sierra on AWS Marketplace, Sierra, outcome-based pricing, Parloa, Moveworks, Moveworks on AWS Marketplace, Vendr, Moveworks, ServiceNow, completes acquisition of Moveworks, ServiceNow, Moveworks achieves FedRAMP Moderate authorization, IBM watsonx Orchestrate on AWS Marketplace. Azure Marketplace and Google Cloud Marketplace listings for Cognigy, Kore.ai, LivePerson, Parloa, Moveworks, Decagon and ElevenLabs, and Vendr's Parloa page, read September 18, 2026. Gartner Peer Insights conversational AI platforms market and vendor pages, G2 product and review pages, and Trustpilot pages, all read September 18, 2026. Our own free-plan builds on ElevenLabs Agents and Voiceflow, September 18, 2026.