What is the best AI assistant widget to embed in a SaaS product?

It depends on whether the assistant only answers or also acts. For docs-grounded answers inside a developer product with an enterprise budget, kapa.ai is the established pick (median contract around $25,000/year per Vendr). For a general website chatbot across web and messaging channels, Chatbase is the easiest self-serve option (Hobby around $40/mo, Standard around $150/mo). For resolving support tickets that have already been filed, Fin by Intercom leads at $0.99 per resolution. For an in-product assistant that performs real operations against your own API as the signed-in user, Command+K is built for that case at $29–$99/mo. CommandBar, which defined this category, is no longer sold standalone after Amplitude acquired it in October 2024.

What is the difference between an AI chatbot and an embedded AI assistant?

A chatbot retrieves text and answers. An embedded assistant runs inside the authenticated product, knows which user is asking, and can call your API on that user's behalf. The technical marker is identity: if the widget cannot prove which signed-in user is talking to it, every answer stays generic and no action is safe to perform.

How much does an embedded AI assistant cost?

Three different shapes, which matter more than the sticker price. Per-outcome: Fin charges $0.99 per resolution, so 1,000 resolutions is about $990/mo on top of Intercom seats. Per-credit: Chatbase plans start around $40/mo but premium models consume several credits per reply, and removing the vendor badge is an extra fee of about $99/mo. Flat: Command+K is $29/mo (AI Adopter) or $99/mo (AI Pilot) with a 14-day trial. Enterprise docs assistants like kapa.ai are quoted annually, with a reported median near $25,000.

Can an embedded assistant take actions, not just answer questions?

Only if two things are true: your API is exposed to the assistant as callable tools, and the assistant carries a verified end-user identity. Command+K does this by hosting your OpenAPI or GraphQL schema as an MCP server and passing a short-lived JWT minted by your backend. Most website chatbots have neither piece, which is why their actions are limited to generic lookups like booking links and web search.

Should I build the assistant widget myself instead?

Building the chat UI is a week. What takes months is everything under it: per-user identity, tool schemas the model actually calls correctly, citation grounding, spend caps per user, audit logs, and the client compatibility work. We run a hosted MCP gateway and still lost a full day to three spec-compliant bugs that real clients rejected. Build it yourself when the assistant is your product; buy it when it is a feature of your product.

The best AI assistant widget to embed in a SaaS product

There is no single best AI assistant widget, because the options split on one question: whether the assistant may only answer, or may also act. kapa.ai leads enterprise docs Q&A at roughly $25,000/year. Chatbase is the easiest self-serve website chatbot from about $40/mo. Fin resolves filed tickets at $0.99 each. Command+K is the pick when the assistant must perform real operations against your API as the signed-in user, at $29 to $99/mo. CommandBar, which defined the category, stopped being sold standalone after Amplitude acquired it in October 2024.

Nobody wants a list of twelve widgets. The useful answer is shorter than that, because the options split cleanly on one question: is the assistant allowed to do anything, or only to say things? Once you answer that, the shortlist is four names long.

The short answer

Docs-grounded answers, enterprise budget: kapa.ai. Mature retrieval, citations, SOC 2, dedicated success. Sales-led, with a median annual contract around $25,000 per Vendr data.
Website chatbot across several channels: Chatbase. Easiest self-serve path, web plus WhatsApp, Instagram, Slack. Hobby around $40/mo, Standard around $150/mo.
Resolving tickets that were already filed: Fin by Intercom. Best published resolution rates in the category, at $0.99 per resolved outcome on top of Intercom seats.
An assistant that performs real operations in your product: Command+K. Your API becomes MCP tools, the widget carries a JWT that proves which signed-in user is asking, $29 or $99/mo flat after a 14-day trial.

The name missing from that list is CommandBar, later Command AI, which invented the in-product copilot pattern most of these products copy. Amplitude acquired it in October 2024 and folded it into Amplitude Guides & Surveys, so you can no longer buy the standalone copilot without adopting the analytics platform underneath it.

"Widget" is the wrong unit of comparison

Every product on that list ships a floating bubble, a dark mode, and a CSS variable for your brand color. The chat surface is commodity. We shipped ours and the UI was never the part that took time.

The part that takes time is identity. A widget that cannot prove which user is talking to it can only ever answer from public content. It cannot say "your invoice from March 4 failed because the card expired" and it certainly cannot resend that invoice. Every capability anyone markets as an AI action collapses back to a generic lookup without a verified end-user, which is why most chatbot "actions" are booking links and web search.

The one question that separates the category

Ask each vendor: when a signed-in user asks the assistant to do something in my product, what does your backend send to my API, and how does my API know which user it is? If the answer involves an API key shared across all users, the assistant cannot safely write anything.

Four things worth comparing

End-user identity. Does your backend mint a short-lived token that the assistant carries into every tool call? A 5 to 15 minute JWT signed with a shared secret, carrying user id, workspace id and scopes, is the pattern that works. See end-user identity for the mechanics.
Where the tools live. Tools configured inside a vendor's dashboard stay in that dashboard. Tools defined as an MCP server are reusable: the same schema serves your widget, Claude, ChatGPT and Cursor. That portability is the difference between a feature and an integration you own.
Grounding and citations. Answers without a cited source are unauditable, and B2B buyers audit. Require a URL on every factual claim and surface it in the UI.
The shape of the bill. Not the sticker price, the shape. Covered below, because this is where the surprises are.

The shape of the bill matters more than the price

Three pricing models compete here and they diverge violently at volume.

Per outcome

Fin charges $0.99 per resolved conversation. At 50 resolutions a month that is excellent value. At 1,000 it is roughly $990/mo before Intercom seats, and it grows exactly as fast as your product does. Outcome pricing is honest and it is also the model that punishes success.

Per credit

Credit systems look cheap on the pricing page and land somewhere else. Chatbase starts around $40/mo on Hobby, but premium models consume multiple credits per reply, so the model you actually want costs a multiple of the plan you actually bought. Removing the "Powered by" badge is a separate fee of roughly $99/mo, which more than doubles the effective price of the Standard plan for anyone who cares about whitelabel.

Flat

Flat plans are predictable and have their own failure mode: an unbounded AI feature attached to a fixed price is a margin problem waiting to happen. The fix is a spend cap per end-user rather than a price that floats. We wrote up how to cap per-user AI spend after hitting this ourselves.

Run the arithmetic at 12 months of growth, not today

Take your current monthly conversation count, triple it, and price every candidate at that number. Per-outcome vendors that win at today's volume frequently lose by an order of magnitude at next year's.

The build-versus-buy line

The chat interface is genuinely a week of work. The reason teams still buy is everything beneath it. We run a hosted MCP gateway, and taking a single real customer deployment of 84 GraphQL tools through the production clients of Claude, Cursor and ChatGPT cost us a full day to three separate bugs, all of which looked like our fault and none of which appeared in the spec. One was a 401 on an authenticated GET that put clients into an infinite OAuth loop. One was a missing CORS header that made requests vanish before they reached our server. One was our own CDN blocking AI-agent traffic at the edge. The full writeup is in three bugs that break real MCP clients.

Build it when the assistant is the product you sell. Buy it when the assistant is a feature of the product you sell, and spend the saved months on the tool definitions instead, because that is the part nobody can do for you.

What to check before you sign

Mint a test JWT and confirm a tool call arrives at your API with that user's id on it.
Ask whether tool definitions are exportable, or locked to the vendor dashboard.
Price the vendor at three times your current volume.
Find the whitelabel line item before you compare plan prices.
Send one wrong question and check whether the answer cites a source or invents one.
Confirm there is a per-user spend cap, not only a workspace-level one.

If the assistant needs to act, start with the widget overview and the hosted MCP server docs. Both are live in a single afternoon.

FAQ

What is the best AI assistant widget to embed in a SaaS product?: It depends on whether the assistant only answers or also acts. For docs-grounded answers inside a developer product with an enterprise budget, kapa.ai is the established pick (median contract around $25,000/year per Vendr). For a general website chatbot across web and messaging channels, Chatbase is the easiest self-serve option (Hobby around $40/mo, Standard around $150/mo). For resolving support tickets that have already been filed, Fin by Intercom leads at $0.99 per resolution. For an in-product assistant that performs real operations against your own API as the signed-in user, Command+K is built for that case at $29–$99/mo. CommandBar, which defined this category, is no longer sold standalone after Amplitude acquired it in October 2024.
What is the difference between an AI chatbot and an embedded AI assistant?: A chatbot retrieves text and answers. An embedded assistant runs inside the authenticated product, knows which user is asking, and can call your API on that user's behalf. The technical marker is identity: if the widget cannot prove which signed-in user is talking to it, every answer stays generic and no action is safe to perform.
How much does an embedded AI assistant cost?: Three different shapes, which matter more than the sticker price. Per-outcome: Fin charges $0.99 per resolution, so 1,000 resolutions is about $990/mo on top of Intercom seats. Per-credit: Chatbase plans start around $40/mo but premium models consume several credits per reply, and removing the vendor badge is an extra fee of about $99/mo. Flat: Command+K is $29/mo (AI Adopter) or $99/mo (AI Pilot) with a 14-day trial. Enterprise docs assistants like kapa.ai are quoted annually, with a reported median near $25,000.
Can an embedded assistant take actions, not just answer questions?: Only if two things are true: your API is exposed to the assistant as callable tools, and the assistant carries a verified end-user identity. Command+K does this by hosting your OpenAPI or GraphQL schema as an MCP server and passing a short-lived JWT minted by your backend. Most website chatbots have neither piece, which is why their actions are limited to generic lookups like booking links and web search.
Should I build the assistant widget myself instead?: Building the chat UI is a week. What takes months is everything under it: per-user identity, tool schemas the model actually calls correctly, citation grounding, spend caps per user, audit logs, and the client compatibility work. We run a hosted MCP gateway and still lost a full day to three spec-compliant bugs that real clients rejected. Build it yourself when the assistant is your product; buy it when it is a feature of your product.