Choose an AI chatbot platform by what it must own: knowledge ingestion, retrieval, conversation storage, identity, actions, human queue, evaluation, and export. Compare hosted agents, inbox suites, and custom orchestration.
A chatbot platform is a system of record for the conversation. The model is replaceable; the place the transcript lives is not. When you evaluate a platform, the real question is which component owns each piece of the customer journey, and whether that owner can handle the failure modes that decide whether your support actually holds together.
This is an architecture and procurement explainer, not another ranking of the best chatbot. It walks through the capabilities a platform must own, the differences between hosted agent platforms, helpdesk suites, and custom orchestration, and one concrete example that traces a single customer question through the stack.
TL;DR: An AI chatbot platform is not the model. It is the system that owns knowledge ingestion, retrieval, conversation and session storage, user identity, actions, human queue, evaluation, and export. Match the platform to where the conversation already ends: a website platform when there is no helpdesk, an inbox suite when agents already live in a ticket system, a store platform when the transcript belongs beside the order, and custom orchestration only when vendor platforms cannot meet a written requirement. Pricing is metered differently across vendors, using message credits, AI credits, outcomes, tickets, or sessions, so compare the unit and the overage terms, not just the sticker price.
A useful way to evaluate any AI chatbot platform is to ask which component is responsible for each of these eight functions.
Knowledge ingestion. The platform must connect to the sources your answers depend on: website pages, PDFs, Notion, Google Drive, Confluence, YouTube, help center articles, and past conversations. It must handle indexing, chunking, and updating when a source changes. If you source article changes, who presses the reindex button? Some platforms offer periodic content updates, some require manual reindexing, and some auto-resync. The platform should make this an operational task, not a one-time setup.
Retrieval. When a customer asks a question, the platform must retrieve the relevant context from its knowledge base and pass it to the model. Retrieval quality determines whether the agent answers accurately or hallucinates. If retrieval fails, the agent should recognize that it lacks enough information and say so, rather than invent an answer.
Conversation and session storage. The platform must store transcripts, session context, and conversation state. This is the system of record for the customer interaction. If the AI agent and the human agent do not share the same transcript, the next human agent will lack the history they need. A new platform in front of an old inbox splits that history.
User identity. The platform must map a conversation to the customer. For email, that may be the sender address. For website chat, it may be a visitor ID that becomes a known user when they log in. For store platforms, the customer identity should connect to the order and account records. A typed email, order number, account ID, or phone number is a lookup key, not proof of identity. Records tied to a customer, and actions taken on their behalf, require an authenticated customer session or a server-validated identity flow, with object-level authorization on every request. The platform should hand authentication to a secure login or verification screen rather than collecting passwords, admin tokens, or one-time codes in chat. A guest lookup can send a secure link to the stored contact address without exposing records in the conversation.
Actions. The platform must be able to perform tasks beyond answering. Common examples include creating a ticket, updating a CRM record, booking a meeting, checking order status, initiating a return, or modifying a subscription. Actions require integrations, permissions, and confirmation that the underlying operation succeeded. An API 200 or 202 response, or a timeout, is not proof that a business action completed: asynchronous jobs can sit pending or fail, so the platform should read the operation's state from the authoritative record before telling the customer it worked, and use idempotency or duplicate prevention before retrying. A generated message and a completed request are different outcomes.
Human queue. When the agent cannot resolve a request, it must hand off to a human with full context. The handoff should include the transcript, the customer identity, and any information the agent already gathered. The queue itself, including assignment, routing, SLA tracking, and the agent workspace, is part of the platform's responsibility.
Evaluation. The platform must provide a way to measure whether the agent is working. Resolution rate, containment, customer satisfaction, and topic analytics are common metrics. The definitions matter: a generated response is not the same as a completed request, and a closed session is not the same as a resolved issue. The platform should let you define the scope, review failures, and improve the knowledge base.
Export. The platform must let you export transcripts and data. This matters for auditing, compliance, and for moving to another platform if you change vendors. If you cannot export your conversations, they are trapped.
A platform that cannot answer four basic questions is a demo: Where is the transcript exported? Who can delete it? What happens when a source article changes, and who presses reindex? What is the overage unit, and what is the cap set to today?
There are three common shapes for AI chatbot platforms. They differ in who owns the conversation, the customer, and the queue.
Hosted agent platforms (Chatbase, YourGPT) provide the widget, the transcript, the knowledge base, and usually a basic inbox. They are a good fit when there is no existing helpdesk and the conversation starts and ends on your website or messaging channels. They focus on building and deploying agents quickly, with options for actions, human handoff, and analytics. See the AI chatbot hub for the full comparison.
Inbox suites (Zendesk, Intercom, Freshdesk) center on the ticket, the agent workspace, and routing. The AI agent is one part of a larger service platform. The agent, the macro, and the assignee share the same ticket. These are the right choice when agents already live in a helpdesk and you want the AI conversation to feed directly into that queue without a second transcript. The help desk comparison covers the numbers.
Store platforms (Gorgias) are specialized for ecommerce. The transcript sits next to the Shopify order, customer profile, and store data. The AI agent can look up orders, handle returns, and assist with purchase-related questions because it has access to the store's data. This is the right shape when the support question is inseparable from the order. Our ecommerce AI support guide explains the tradeoffs, and the customer service software for ecommerce table compares options.
The distinction matters because moving the bot to a platform that does not own the queue means two transcripts for one customer.
Consider a Shopify customer who writes, "Where is my order and can I return the shoes if they don't fit?" Trace this question through the three platform shapes.
Hosted agent platform. After the shopper authenticates through the store's customer account flow, the agent can query the order status through its Shopify connection and show the order details or a product card. On platforms with a native Shopify connection, built-in shopping actions such as catalog search, product cards, and add-to-cart exist when Agent Mode is enabled; order-status and return actions are separate capabilities whose configuration you should verify. Enabling the shopping actions does not by itself grant broader Admin API access. If the return action is not connected, the agent hands off to a human, who processes the return in the store's separate system.
Store platform. The agent has direct access to the order and the store's return workflows through the same helpdesk. After the shopper is authenticated, it can check the order status, confirm the return window, and initiate the return within the same conversation, if the return action is configured. The ticket and the order stay together. If the agent cannot resolve the request, the human agent sees the order history and the conversation in the same helpdesk.
Inbox suite. The agent answers from the knowledge base and can check order status if integrated with the store, after the shopper's identity is validated. If a return needs human approval, the agent creates a ticket in the helpdesk. The human agent sees the conversation and can process the return, but the order data may require a separate integration.
The platform shape determines how much of that question can be completed autonomously and how smoothly the human handoff works.
AI chatbot platforms meter usage differently, and the unit shapes the cost.
Message credits (Chatbase) are the counting unit. Chatbase's Hobby plan is $32/month billed yearly with 700 message credits, and auto-recharge adds $40 per 1,000 credits. A message credit is not necessarily one conversation; the platform counts credits based on usage, so check how many credits a typical conversation consumes.
AI credits (YourGPT) are consumed by model calls. Different models consume credits at different rates. Per YourGPT's pricing FAQ, GPT-3.5 uses 1x, GPT-4o-mini 0.5x, GPT-4o 5x, GPT-4 Turbo 10x, and GPT-4 20x. YourGPT's Essential plan is $39/month billed annually with 10 million AI credits. A long conversation using a powerful model consumes more credits than a short one using a cheaper model. Credit consumption is not interchangeable with Chatbase's message credits.
Outcomes (Intercom Fin) are charged per successful outcome at $0.99. An outcome is counted when a customer confirms a resolution, does not ask for more help after Fin responds, or Fin completes a configured Procedure handoff. Disqualifications and self-serve routings are also $0.99, while sales qualifications are charged at $9.99 each. Intercom's Essential seat is $19/month billed annually (a promotional price beside a struck-through $29; the 35% savings applies to new customers under conditions that are not fully disclosed), Advanced is $85, and Expert is $132 per seat. Fin can also run on an existing helpdesk without Intercom seats, with a 50 outcome monthly minimum.
Tickets (Gorgias) are the helpdesk unit. Gorgias pricing scales by ticket volume, not per agent, and AI Agent is included on every plan with pay-per-resolution billing. Under current terms, a fully automated resolution is charged both a ticket fee and an automation fee; accounts created before May 28, 2025, may be on legacy pricing where a ticket counts toward one or the other, so confirm which terms apply. Gorgias' current price sheet is interactive rather than a published tier list, so request a quote for your ticket volume and AI interaction allotment.
Sessions (Freshdesk) are the unit for its Freddy AI Agent. Freshdesk Growth is $19/agent/month billed annually, Pro $55, and Enterprise $89. The first 500 AI agent sessions are included on each of those plans, and additional sessions cost $49 per 100. A Freshdesk email AI session is defined as a 72-hour window, not a single resolved ticket.
When comparing platforms, ask three things: What is the usage unit? How much of a typical conversation does one unit consume? And what happens when you exceed the plan's limit, whether that is overage fees, recharge credits, or a hard stop? Our AI chatbot pricing guide walks through the math.
The decision should be driven by where your team already works and where the customer's issue is best resolved.
Use a website platform when there is no help desk. If you are a small team without a ticket system, a hosted agent platform gives you the widget, the transcript, and a place to start. Chatbase and YourGPT both provide the infrastructure for building and deploying an agent, with different usage units and feature depths. You can add a human queue later.
Use an inbox suite when agents already live in a helpdesk. If your support team works in Zendesk, Intercom, or Freshdesk, the AI conversation should feed into that same ticket. Moving the bot to a separate website platform means two transcripts for one customer. The suite gives you routing, SLAs, and an agent workspace alongside the AI.
Use a store platform when the transcript belongs beside the order. For ecommerce brands on Shopify, Gorgias connects the conversation to the order and customer data. This is the natural fit when most support questions are about orders, shipping, and returns.
Consider custom orchestration only when vendor platforms cannot meet a written requirement. Build your own host only after you have verified that no vendor platform can perform the required action or meet a written data boundary. Custom builds mean you own the retrieval, conversation storage, identity, and queue, and you also own the maintenance, security, and failure recovery. The bar for this decision is high. See our guide on building an AI chatbot for the requirements to test before you commit.
The tradeoff is not "keep your existing helpdesk" as a universal answer. If your helpdesk lacks the AI capabilities you need, or if you have no helpdesk at all, moving to a platform that owns the conversation may be the right call. The reason to move should be a specific capability gap, not a desire to try something new.
This article is based on public vendor documentation reviewed on October 26, 2026. It does not include hands-on product testing or independently measured outcomes. Vendors make claims about resolution rates, integrations, and features; those claims are attributed and scoped, not independent benchmarks. Resolution rates are not directly comparable across vendors because the definitions of a resolution, a session, and an outcome differ. YourGPT, for example, defines an AI resolution as the successful completion of an eligible customer request without a human agent taking full ownership, which is a stricter definition than a contained or closed chat.
When comparing platforms, ask for the current price sheet and the overage rate on a quote. Public pricing pages change frequently, and promotional prices may not be permanent. Confirm the usage unit and the cap before you commit.
Pick the platform shape that already holds your conversations, then verify the specific capabilities you need against the vendor's documentation. The eight ownership components are the checklist: knowledge ingestion and retrieval decide answer quality, conversation storage decides continuity, identity and actions decide whether requests are completed safely, and the human queue decides what happens when the agent cannot finish the job. Pricing should be compared by usage unit and overage terms, not by sticker price, and promotional or quote-based figures should be confirmed in writing.
What is the difference between an AI chatbot platform and a model? A model generates text. A platform owns the knowledge base, retrieval, conversation storage, user identity, actions, human queue, evaluation, and export. The model is replaceable; the platform is the system of record for the conversation.
Do I need a helpdesk to use an AI chatbot platform? No. Hosted agent platforms like Chatbase and YourGPT include a widget, transcript, and often a basic inbox. You can start without a separate helpdesk. If you already have agents working in a ticket system, the AI should feed into that system to keep one transcript.
How do I compare pricing across platforms that use different units? Identify the usage unit, estimate how much of a typical conversation one unit consumes, and check the overage terms. Chatbase uses message credits, YourGPT uses AI credits that vary by model, Intercom Fin charges per outcome, Gorgias charges per resolved interaction, and Freshdesk charges per AI session. Convert each into an estimated monthly cost for your volume.
Can an AI agent access order or account details when a customer provides an email or order number? No. An email, order number, or account ID is a lookup key, not proof of identity. Access to private records and actions on an account require an authenticated customer session or a server-validated identity flow, typically through a secure login or verification screen. For guest lookups, the platform can send a secure link to the stored contact address instead of exposing records in chat.
What happens when the AI cannot resolve a customer request? The platform should hand off to a human with full context. The handoff should include the transcript, customer identity, and any information the agent gathered. Some platforms let the agent request human approval while continuing to own the conversation, avoiding a full escalation.
What should I check before committing to a platform? Ask five questions: Where is the transcript exported? Who can delete it? How are customers authenticated before the agent accesses private records? What happens when a source article changes, and who presses reindex? What is the overage unit, and what is the cap set to today? If the platform cannot answer these, it is a demo.
When should I build a custom chatbot instead of using a platform? Only when vendor platforms cannot perform a required action or meet a written data boundary. Custom builds give you control over retrieval, storage, identity, and the queue, but you also own the maintenance, security, and failure recovery. Verify the requirement is unmet before building.
