Scope & boundaries
This guide covers customer- and employee-facing conversational applications, where containment rate and clean handoff to a human decide whether the purchase worked.
It does not cover the text analytics and language tooling underneath it (NLP & Text Analytics Platforms), internal agent runtimes, where durability and observability decide it (AI Agent & Agentic AI Platforms), or the voice channel, where latency and caller authentication are the hard problems (AI Voice Agents & IVR Replacement).
Executive Summary
Choose a conversational AI platform by how much of each conversation you need to control and by what it can change in your systems. Every platform here can hold a fluent conversation.
A conversational AI chatbot platform is the software used to build, run and measure chatbots and AI agents that talk with customers or employees in text or voice and complete tasks in the systems behind them. Buyers also search for it as a conversational chatbot platform or a conversational automation platform. IBM defines conversational artificial intelligence (AI) as technologies, such as chatbots or virtual agents, that users can talk to. IBM defines a chatbot as a software application that communicates with people through text or voice.
The platforms differ in three decisions. The first is control: how much of each conversation follows scripted flows that you author, and how much is generated by a large language model (LLM). The second is where the platform runs: inside a cloud provider, inside a productivity suite, inside a contact center product, or on its own. The third is the pricing unit: a session, a request, a credit, or a monthly plan.
This guide explains what these platforms do, how scripted and generative agents differ, the security and disclosure rules a customer-facing bot must meet, six approaches to buying, the eight vendors in the table that represent them, and how pricing works. The contact center that a bot hands over to is covered in the Contact Center Platforms Buyer Guide. Agents that answer employees’ IT and HR requests are covered in the AI Service Desk Agents Buyer Guide, and voice agents that replace an IVR in the AI Voice Agents Buyer Guide.
Why the Platform Choice Matters More Than the First Bot
Most organizations start with one bot for one channel: a website assistant, a WhatsApp channel or an internal helpdesk. The platform chosen for that first bot usually becomes the place where the next twenty are built, so it decides what those bots can reach, who can change them and how they are governed.
The words in this market overlap. IBM describes conversational AI as a broad category of technologies that enables systems to understand and respond to human language, and a chatbot as one application of that technology. IBM notes that some organizations use virtual agent interchangeably with chatbot, often in customer service. IBM states that not all chatbots use artificial intelligence (AI), and that traditional chatbots follow predefined rules, decision trees and scripted conversation flows. IBM states that AI agents can plan, make task-specific decisions and carry out multi-step workflows with limited human involvement. Most platforms in this guide now sell AI agents that can call tools and update records, not only answer questions. Ask each vendor which actions its agents can take in your systems, and who approves a new one.
A bot that can act can be misused. OWASP states that a prompt injection vulnerability occurs when user prompts alter the LLM’s behavior or output in unintended ways. OWASP describes a scenario in which an attacker injects a prompt into a customer support chatbot, instructing it to ignore previous guidelines, query private data stores, and send emails, leading to unauthorized access and privilege escalation. OWASP states that it is unclear if there are fool-proof methods of prevention for prompt injection. OWASP recommends human-in-the-loop controls for privileged operations to prevent unauthorized actions. Give a customer-facing agent only the permissions its tasks need, and test it with adversarial prompts before launch and after every change.
Customers must know they are talking to a machine. The European Commission states that the transparency rules of the AI Act will come into effect in August 2026. The European Commission states that when using AI systems such as chatbots, humans should be made aware that they are interacting with a machine so they can take an informed decision. Check how each platform labels the bot in every channel, and how it hands the conversation to a person who can see what the bot already did.
Scripted or Generative: How Do Conversational AI Chatbot Platforms Differ?
A scripted, or deterministic, bot follows flows and intents that someone authored. It gives the same answer to the same question and can be audited line by line, but every new question needs new design work. A generative bot answers from instructions, documents and tools through an LLM. It handles questions nobody planned for, but its wording varies and it needs guardrails and testing to stay inside policy.
Every platform in this guide now offers both, and the vendors’ own documents show how they combine them. Google states that Conversational Agents has two editions: Flows (also known as Dialogflow CX), deterministic agents built with intents and flows, and Playbooks, generative agents that can be built with natural language instructions. AWS states that AI agents can handle more complex, open-ended interactions, while Amazon Lex remains a foundational service for handling predefined intents and specific conversational flows. IBM describes using deterministic patterns alongside AI reasoning to control behavior in critical interactions. Cognigy describes designing AI agents in AI Agent Studio, combining LLM reasoning with deterministic steps in one conversation. Kore.ai states that reasoning and deterministic flows work together.
The split matters for cost as well as control. Some vendors price a generative turn higher than a scripted one, and some bill a whole conversation at the generative rate once any generative step is used. The pricing section gives the published rates.
Use scripted flows where the answer must not vary: payments, identity checks, regulated disclosures and anything that changes an account. Use generative answers for the long tail of questions answered from documents. Ask each vendor to show both in one conversation, including the hand-back from a generative step to a scripted one.
| Scripted flows | Generative agent | Hybrid | |
|---|---|---|---|
| How it is built | Intents, entities and flows authored by a designer | Instructions, knowledge sources and tools given to an LLM | Flows for fixed paths, an LLM for the rest |
| Same question, same answer | Yes | Not guaranteed | Yes on scripted paths |
| New question nobody planned for | Falls back or hands over | Answers from its sources | Answers from its sources |
| What to test | Intent recognition and flow coverage | Grounding, policy adherence, prompt injection | Both, plus the switch between them |
| Typical use | Payments, identity, regulated steps | Product and policy questions | Most customer service deployments |
How do you evaluate Conversational AI Chatbot Platforms?
Start with actions. List the ten requests your customers or employees make most often and the system each one changes: an order, a booking, a ticket, an account. Ask each vendor to build three of them against your sandbox, with authentication, and to show what the bot does when the system returns an error.
Test grounding next. Load a sample of your own knowledge base, ask questions it answers and questions it does not, and count the wrong answers given with confidence. Ask whether each answer can cite its source.
Test handover. When the bot cannot finish, the person who takes over should see the transcript, what the bot already did and why it stopped. Ask each vendor to hand a conversation to your contact center or service desk, not to its own demo desktop.
Check model choice and data handling. Some platforms run one vendor’s models, others let you choose a model per task or bring your own. Ask where conversations are stored, for how long, and whether they are used to train anyone’s model.
Check testing and monitoring. A platform should run a set of test conversations before each release and show, after launch, which conversations ended without resolution and why.
Channels come last because most platforms cover the common ones. Confirm the channels you need today, such as web chat, WhatsApp, SMS, a collaboration app or voice, and the languages your customers use.
| Capability | What it does | Buyer translation |
|---|---|---|
| Actions and integrations | Reads and writes your systems through connectors and APIs | Which of your top ten requests the bot can finish end to end |
| Knowledge grounding | Answers from your documents and sites | How often it answers wrongly, and whether it cites a source |
| Scripted and generative control | Fixed flows beside LLM-generated answers | Whether regulated steps can be locked down |
| Guardrails and permissions | Limits topics, data and actions | What an attacker could make it do with a crafted prompt |
| Handover | Passes the conversation to a person | What the person sees, and in which desktop |
| Testing and analytics | Simulated conversations, resolution reporting | Whether you can prove a change did not break anything |
| Channels and languages | Web, messaging, voice, collaboration tools | Whether every channel uses the same bot and the same rules |
Vendors in Conversational AI Chatbot Platforms
Representative vendors include Ada, Amazon Lex + Amazon Q in Connect (Amazon Connect AI agents), Google Customer Engagement Suite / Conversational Agents (CX Agent Studio), IBM watsonx Assistant / watsonx Orchestrate, Kore.ai, Microsoft Copilot Studio, NiCE Cognigy and Yellow.ai. The list is alphabetical, is not a ranking, and is not for sale. How we choose vendors
The approaches below follow two questions: where does the platform run, and who is expected to build on it? Two pairs of vendors share an approach. Several product names in this market have changed, so the table prints the earlier name with the current one.
Two cloud providers sell agent builders billed by use. Google has described Conversational Agents, Agent Assist, Conversational Insights and Google Cloud Contact Center-as-a-Service as the four products within the Customer Engagement Suite. Google describes Customer Experience Agent Studio (CX Agent Studio) as a minimal code conversational agent builder. Google states that CX Agent Studio represents the evolution of Dialogflow CX. AWS describes Amazon Lex as an AI chat builder that allows users to interact with any application using natural language voice or chat. AWS states that Amazon Lex natively integrates with Amazon Connect, its cloud contact center, for bots that handle customer queries across channels, including chat or phone. AWS states that functionality previously known as Amazon Q in Connect is simply one of its AI agents.
A productivity suite vendor sells a low-code agent builder for employees and customers. Microsoft describes Copilot Studio as a graphical, low-code studio for building and managing AI-powered agents and workflows. Microsoft states that agents can work with employees and customers in multiple languages across Microsoft Teams, Microsoft Copilot, websites, mobile apps, and other channels.
An agent management platform runs agents built in it and elsewhere, and absorbs an earlier assistant product. IBM describes watsonx Orchestrate as an agent management platform designed to help build, deploy, orchestrate, manage and govern AI agents wherever they are built or run. IBM’s documentation describes upgrading an existing watsonx Assistant instance or assistants to take full advantage of the agentic AI capabilities in watsonx Orchestrate, and states that existing assistants, dialog flows, actions, and integrations continue to operate without disruption after the upgrade. IBM states that AI agents can be connected to the channels customers use, across web chat, phone, SMS and WhatsApp.
A standalone agent platform defines agents in its own language and runs them on any model. Kore.ai describes ABL as a typed, schema-driven language purpose-built for agentic AI that lets enterprises define agent behavior, tools, guardrails, orchestration, and handoff logic. Kore.ai states that compliance rules are enforced at the engine level, outside the LLM. Kore.ai states that application definition is independent of the LLM: swap models and the ABL still executes.
A contact center vendor now owns one of the established conversational AI platforms and still sells it on its own. NiCE announced on September 8, 2025 the closing of its acquisition of Cognigy, following receipt of all required regulatory approvals. NiCE states that Cognigy will be available both as part of the unified CXone platform and as a standalone offering. NiCE Cognigy states that it works with any CCaaS platform, is built for Amazon Connect, Avaya and Genesys, and is native inside NiCE CXone. Cognigy states that customers can pick the LLM for each job from OpenAI, Anthropic, Google, AWS or their own, and switch without rebuilding agents. Cognigy states that AI Agent Evaluation runs simulated conversations against agents before launch and after every change.
Customer service AI agent vendors sell to customer experience teams rather than to developers. Yellow.ai describes building, testing and deploying AI agents with natural language prompts. Yellow.ai describes a multi-LLM architecture for flexibility, with no vendor lock-in. Yellow.ai lists 150+ pre-built enterprise integrations, including Salesforce, Zendesk, Genesys, and Nice. Ada describes orchestrating multiple LLMs to understand, act, and resolve complex inquiries. Ada lists voice, email, chat, Messenger, WhatsApp, SMS, Instagram, in-app and custom channels. Ada states that Playbooks let a customer’s AI agent follow multi-step SOPs using real-time data, without relying on rigid, scripted frameworks. Ada states that Ada and Dialpad are joining forces to bring people and AI together to do more for every customer. Ada states that its platform stays open and continues to integrate with existing contact centers, business applications and partners.
Messaging platforms built for contact centers compete for the same budget. LivePerson states that customers can orchestrate AI and human agents across every channel and bring their own LLM. LivePerson lists connectors for common third-party bot platforms, including IBM Watson, Google Dialog Flow, Amazon Lex and Microsoft bots. LivePerson describes simulating thousands of real customer interactions to validate AI agents before they ever interact with a customer.
| Vendor | Approach | Where it fits |
|---|---|---|
| Amazon Lex + Amazon Q in Connect (Amazon Connect AI agents) | Agent builder from a cloud provider | Lex for scripted intents, Connect AI agents for open-ended ones; priced per request |
| Google Customer Engagement Suite / Conversational Agents (CX Agent Studio) | Agent builder from a cloud provider | Successor to Dialogflow CX; $0.50 per chat or voice session |
| IBM watsonx Assistant / watsonx Orchestrate | Agent management platform with an assistant upgrade path | In-place upgrade from watsonx Assistant; plans from $578 a month |
| NiCE Cognigy | Conversational AI from a contact center vendor | Works with any CCaaS platform; sold inside CXone or on its own |
| Ada | Customer service AI agents for CX teams | Playbooks for multi-step procedures; joining forces with Dialpad |
| Yellow.ai | Customer service AI agents for CX teams | Agents built from prompts; multi-LLM; 150+ integrations |
| Microsoft Copilot Studio | Low-code agents inside a productivity suite | Teams, websites and apps; Copilot Credits pooled across the tenant |
| Kore.ai | Standalone agent platform with its own agent language | Agents defined in ABL; compliance rules enforced outside the LLM |
Eight vendors across six approaches; two approaches have two vendors each. Each vendor was selected to show one way of buying, after the approaches were set,. A place in the table is not for sale. Contact center suites, messaging platforms and newer AI agent start-ups also sell conversational AI, and any of these vendors can make its own case in the Spotlight below.
How much should you budget for Conversational AI Chatbot Platforms?
Conversational AI platforms are priced in four units: a session, a request, a credit, or a monthly plan. The unit decides what happens to the bill when a bot gets busier or starts using generative answers, so model a year of your own traffic in each vendor’s unit before you compare.
The cloud providers publish rates per session or per request. Google lists CX Agent Studio at $0.50 per chat session and $0.50 per voice session, in USD. Google lists Conversational Agents chat requests at $0.007 each for Flows and $0.012 each for Playbooks. AWS prices an example bot that processes 8,000 speech requests and 2,000 text requests in one month at $0.004 per speech request and $0.00075 per text request. AWS states that new Amazon Lex users can process up to 10,000 text requests and 5,000 speech requests per month for free for the first year.
Credit-based pricing charges more for generative and agentic steps than for scripted ones. Microsoft sells Copilot Studio as a tenant-wide license with capacity packs of 25,000 Copilot Credits at $200.00 per pack per month, and also as a pay-as-you-go meter. Microsoft lists a classic answer at 1 Copilot Credit, a generative answer at 2 Copilot Credits and an agent action at 5 Copilot Credits. Microsoft states that the Copilot Credits in a Copilot Studio license are pooled across the entire tenant. Model how many generative answers and agent actions a typical conversation uses, not only how many conversations you expect.
Plan-based pricing starts from a monthly fee with usage limits. IBM lists an Essentials plan starting at $578.00 USD per month and a Standard plan starting at $6,360.00 USD per month. The specialist platforms in the table do not list prices on the pages cited here; ask each for a quote priced on your own conversation volume and channels.
Model costs and telephony can sit outside the platform price. If you bring your own LLM, its token charges are billed by the model provider. Voice bots also carry carrier minutes and phone numbers.
| Model | You are charged for | Grows with | Where it goes wrong |
|---|---|---|---|
| Per session | Each conversation, chat or voice | Conversation volume | Long conversations split into several billed sessions |
| Per request | Each user turn or API call | Turns per conversation | Bots that ask many clarifying questions |
| Per credit | Each answer and action, weighted by type | Generative answers and agent actions | A bot moved from scripted to generative answers |
| Monthly plan | A tier with usage limits | Tier upgrades | Limits reached mid-year, forcing a jump in tier |
| Model tokens | LLM usage when you bring your own model | Prompt and answer length | Long knowledge passages sent with every turn |
| Telephony | Carrier minutes and numbers for voice bots | Voice minutes | Voice added after a chat-only quote |
How long does implementation take for Conversational AI Chatbot Platforms?
Most of the work in a conversational AI deployment is integration and testing, not conversation design. Connecting the bot to the systems it changes, deciding what it may do without a person, and building a test set take longer than writing the first flows.
Rank requests by volume from transcripts and ticket data. For each, decide whether it is scripted, generative or handed to a person, and which system it must change. Agree the disclosure wording and escalation rules with legal and risk.
Connect the systems behind the top requests with authenticated access and the narrowest permissions that work. Load and clean the knowledge sources. Build a test set of normal, edge and adversarial conversations.
Launch to one channel and a share of traffic, with a fast route to a person. Review every conversation that ended without resolution, and rerun the test set after each change.
Add requests, channels and languages in waves. When migrating from an older bot, run both side by side on the same traffic before switching, and keep the old transcripts for the retention period you need.
What should you ask vendors about Conversational AI Chatbot Platforms?
Ask each vendor to answer in writing and to demonstrate on your own data.
-
Does your contact center or cloud already include an agent builder?Yes Start with it, and test whether it covers your channels and systems before adding a specialist.No Shortlist standalone and customer service platforms that integrate with your contact center.
-
Will developers build and maintain the agents?Yes Include the cloud providers’ builders and the agent platforms that expose code and APIs.No Weight low-code studios and platforms sold to customer experience teams.
-
Must some conversations follow a fixed, auditable path?Yes Require scripted flows beside generative answers in one conversation, and test the switch.No Weight grounding quality, testing tools and the price of generative turns.
RFP packages are a separate paid product. Vendors do not pay to appear in buyer guides.