Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- Kapa AI at a glance
- What is Kapa AI?
- How easy is it to set up Kapa AI?
- What channels does Kapa AI work in?
- What are the limitations of Kapa AI?
- What knowledge sources can I train Kapa AI on?
- What features does Kapa AI have?
- How do I improve Kapa AI responses?
- What deflection rate can I expect from Kapa AI?
- What AI model does Kapa AI use?
- What languages does Kapa AI work in?
- How secure is Kapa AI?
- Who is using Kapa AI?
- Does Kapa AI have a free trial?
- Is Kapa AI worth it?
- What are the pros and cons of Kapa AI?
- Pros
- Cons
- FAQs
Created time
Sep 8, 2026 01:32 PM
Title length (<60)
Author
Last optimised
Ecomm?
Tags
Image
Publish date
Aug 16, 2026
Video
Slug
kapa-ai-complete-guide-2026
Featured
Type
Article
Ready to Publish
Ready to Publish
Kapa AI answers technical questions from your own documentation, with citations, and it stops a lot of them before anyone files a ticket. Inside your helpdesk it drafts a reply for a human to send.
Kapa AI is strongest where the answer already exists in public technical material: it searches the documentation, source code and community discussions a company has published, and replies with the source attached. That makes it a good match for developer-tools, semiconductor, hardware and software companies, and a poor one for ecommerce or account-specific questions. Where it stops is support: Zendesk is its one helpdesk connection, and I read that app as agent-assist, because a person still sends the reply it drafts.
You run a product with a big documentation site, and someone has asked you to put AI in front of it. Kapa AI comes up early in that search.
The problem is that Kapa's own pages answer a slightly different question than the one you were given. They tell you the answers are accurate and cited. They say much less about whether any of it takes work off your support team.
I'm Mike, co-founder of My AskAI. We sell an AI support agent, so I have a stake in this.
By the end of this you'll know what Kapa does well, where it stops, and whether your tickets are the kind it can help with.
Kapa AI at a glance
Fact | Detail |
What it is | Cited answers grounded in your own technical knowledge |
How you're billed | Free 14-day trial, then two sales-only tiers |
Best for | Technical B2B with large public documentation |
Not a fit for | Non-technical, ecommerce or transactional support |
Biggest limitation | Never replies to customers in your helpdesk |
Our verdict | 7/10 |
We also published a full My AskAI vs Kapa AI comparison, which covers what each one costs at real ticket volume.
What is Kapa AI?
TL;DR: Kapa AI is a retrieval-based answer engine. It turns your technical documentation, code and community threads into a cited assistant, deployed in your docs, your product, Slack, Discord and a hosted MCP server.
Kapa indexes a company's technical knowledge into one searchable store. When someone asks a question, it looks the answer up first, then writes a reply with links back to the pages it used. Where the sources don't cover the question, Kapa's documentation says it will tell you so.
Kapa now sells the layer underneath all of that, as a knowledge retrieval API for AI agents, offering them "accurate, cited context from your knowledge base via API or MCP".
Both framings are live, and I think that's deliberate. Kapa still calls the product an Answer Engine, and still uses that label on its customer stories. So I read Kapa two ways: as the assistant sitting on your docs, and as the retrieval service that feeds any assistant you build.
Kapa's Y Combinator profile lists a team of 22 and names Emil Soerensen and Finn Bauer as co-founders. Its own homepage says more than 200 companies use it, and it names OpenAI, Monday and Logitech among them.
The audience is narrow, and Kapa is open about it. The four industries it lists for itself are developer tools, semiconductors, software, and hardware and industrials (Nordic Semiconductor, Espressif and Silicon Labs are all named customers).
That points squarely at technical B2B products with real documentation. A shop with an order-status problem is a different buyer entirely.
My AskAI is pointed at the other half of that job. We train on your help center, your past tickets and your live backend data, then plug into the helpdesk you already run so the answer lands on the ticket itself.
How easy is it to set up Kapa AI?
TL;DR: Point Kapa at your docs and a basic widget goes live quickly. A custom in-product agent is a frontend engineering job, and there is no self-serve path to a price.
The flow has four steps: connect your sources, let Kapa index them, pick a surface to deploy, then watch the analytics. Connecting a first source takes a few minutes, and Kapa's integrations documentation describes deployment as "a script tag, an install command, a toggle".
The widget on your documentation is where a Kapa rollout starts (and if you only ever do one thing with Kapa, make it that). That is the cheap end of the work, and where the vendor's promise of quick deployment holds up best.
The custom in-product agent is a different job. Kapa's Agent SDK comes as two packages: one handles sessions, streaming, tool calls and the approval step, and the other gives you React components and hooks. It is described as frontend-first, with your server doing nothing but minting a short-lived session token (so the work lands on whoever owns your web app).
What you save is the retrieval work. Kapa's own line for it is "No RAG to build." RAG just means the assistant looks up your documents before it answers, and every agent here arrives with a search tool already wired to your indexed sources.
Building it is still a developer project (nobody is configuring this from a settings screen).
On price, Kapa publishes a tier ladder and no rate card. Kapa invites you to "Start for free. Scale when your agents ship." across three tiers: a free trial, Growth for agents running in production, and Enterprise for governed and global deployments.
Both paid tiers ask you to "Talk to us". You can sign up without talking to anyone.
My AskAI installs from your helpdesk's own marketplace, with no code and no developer, and most teams are live in 10 to 15 minutes. Our prices are published on the pricing page, so you can build the business case before anyone books a call.
What channels does Kapa AI work in?
TL;DR: A docs widget, in-product chat, Slack, Discord, a hosted MCP server, SDKs and an API, plus a pre-ticket support-form deflector. On helpdesks, the story is Zendesk and nothing else.
Kapa publishes eight prebuilt places to run the same agent: the Website Widget, the Slack Bot, the Discord Bot, an Internal Technical Assistant, the Support Form Deflector, the Zendesk Agent, a Chat SDK and a Chat API. Its deployment documentation frames these as one agent tuned per deployment, so what changes between them is where the agent works. The free trial opens the retrieval API and the hosted MCP server alongside those prebuilt surfaces, so a developer can test the layer underneath directly.
Underneath sits the developer layer we keep hearing about from technical buyers: a retrieval API, the Agent SDK, and a hosted MCP server your coding assistant can call. Kapa runs a public documentation server of its own, and customers get their own scoped servers with OAuth. That is how teams like Expo and Redpanda expose their docs to an AI coding agent.
I rate the placement higher than anything else in the product. The answers are embedded in the documentation and reachable from the search bar, and the Slack presence is rich. Developers already live in both of those places, so nobody has to be talked into a chat bubble.
Two of the eight surfaces touch support, and neither answers a customer in an inbox. The Support Form Deflector catches the question on your support form before it becomes a ticket. The Zendesk Agent sits beside your human agents and helps them write.
Kapa's published integration list names Zendesk as its only helpdesk, with no Intercom integration, no voice channel and no email-inbox auto-reply.
We built My AskAI to work inside the helpdesk your team already runs: Zendesk, Intercom, Freshdesk, Freshchat, Gorgias and HubSpot. The replies go out on the ticket, in the helpdesk's own chat, messaging or email channel.
What are the limitations of Kapa AI?
TL;DR: Kapa answers documentation questions. It does not pick up a ticket, reply as your agent, tag it, route it, or handle anything a document can't answer.
Kapa's Zendesk integration installs as a Support sidebar app that puts an assistant grounded in your knowledge base next to every ticket, as its Zendesk Agent documentation sets out.
I count this as the strongest thing Kapa does on the support side. It reads the ticket on its own: subject, status, priority, tags, the whole comment thread, and the customer's recent ticket history. Nothing gets pasted in.
Ask it to summarize a ticket and it writes a handoff note for the next agent: the customer's ask, the recent messages, the status, and what is still open. It stages that as an internal note (we stage the same note on handover).
It is careful about internal content too. On a customer-facing draft it can draw on sources marked internal while stripping out any citation link that points to one. On internal surfaces the citations work normally.
Kapa states the boundary itself:
"Drafts are staged into the Zendesk composer for human review before they reach the customer — the agent never clicks Submit." — Kapa AI product documentation
One click puts the draft in the public-reply box, another discards it, and a person sends it. Kapa tells agents to read every draft and click through the citations first, because the private and public split depends on how each source was tagged. I check the source tags before I trust that split.
Every insert writes an
assisted_by_kapa tag, so you can report on which tickets Kapa touched. Kapa publishes no ticket tagging, routing or triage feature, so nothing on the ticket records why the customer wrote in.The Zendesk app needs a Zendesk Suite Growth plan or higher, or Support Professional, with the Agent Workspace turned on (check your plan before anyone gets excited).
Every published surface answers from the knowledge you connect, so refunds, billing changes, account changes and order lookups sit outside it. A developer can wire your own endpoints in as agent tools through the SDK, and I treat that as a project with a start date and an owner.
Customization does not close a gap in the source material either. Kapa says of its editable instructions: "They do not add knowledge." If the topic is not in the connected sources, the assistant will not answer it, and Kapa's customization documentation spells that limit out.
This is the line My AskAI is built on the other side of. Our agent picks the ticket up, replies as your agent in the helpdesk, tags and routes it, and hands it to a human with a summary when it can't finish the job.
What knowledge sources can I train Kapa AI on?
TL;DR: Kapa's catalog names 23 connectors, and past Zendesk tickets, Salesforce knowledge and cases, Confluence, Notion, Jira and GitHub are all on it.
The connector catalog starts where you would expect: web crawling, Slack, Discord, Discourse, Stack Overflow, file upload, S3, OpenAPI, Google Drive and YouTube. It carries on through Zendesk Tickets, Zendesk Help Center, Salesforce Knowledge, Salesforce Cases, Confluence, Notion, Jira, Jira Service Management, Custom Answers, and GitHub code, issues, pull requests and discussions. Kapa's own documentation puts the wider figure at 50+ sources.
Two of those surprised me.
Zendesk Tickets indexes your historical support tickets, and Salesforce Cases indexes support cases including comments and email threads. Kapa will happily learn from tickets you have already answered.Kapa's published catalog names no connector for SharePoint, OneDrive, Dropbox or Shopify. Its two help-center connectors are Zendesk Help Center and Salesforce Knowledge, so an Intercom, Freshdesk, Gorgias or HubSpot help center has no listed route in.
I check the indexed types first. Long and complex PDFs, images, source code and community threads are all indexed types, and Kapa has published its own work on indexing images for retrieval. Ingestion keeps the index current on its own, re-processing what changed on a per-source schedule, with source groups for multi-product or multi-version setups.
My AskAI reads your help center and your historic tickets. Train on Historic Tickets drafts knowledge articles from the last 5,000 tickets by default (more on request) and puts them in Self-Learning for you to review. If you have no help center at all, that is how you start from scratch.
What features does Kapa AI have?
TL;DR: Three layers. Grounded answers with citations, an analytics loop that shows where your documentation falls short, and a developer layer of SDKs, an API and a hosted MCP server.
Layer one is the answering itself. Kapa owns the prompt and model orchestration, the citation format, the uncertainty reporting and the continuous evaluation, and its defaults stay cautious. When the sources don't cover a question, it says so and stops (we hand the conversation to a person at that same point).
Layer two is analytics. Every retrieval gets logged, so you can see what was asked and what was retrieved.
Coverage Gaps names the topics Kapa was unsure about, and Top Questions shows you what people ask most. I ask for a live look at both in a demo.
Layer three is for developers: the retrieval API, the hosted MCP server, the Agent SDK and the Chat SDK.
Two more pieces sit alongside those layers. The first is Kapa Skills, five installable skills for Claude Code, Cursor and Codex. I judge a feature by where it installs, and nothing in the Skills set installs into a helpdesk.
The most useful of them is Answer RFPs. It pulls every requirement out of an RFP or a security questionnaire, fetches answers live from your knowledge base, grades each one by how well the sources back it, and writes the response out to Word, PDF or Excel.
If you have ever lost a week to a vendor questionnaire, that one is worth a look.
The second is how much of the assistant you can change, and I found the documentation for it specific and easy to plan against. Kapa splits it three ways. There is a fixed harness you cannot touch, covering formatting, citations, prompt-injection protection and uncertainty reporting.
Then come the editable sections, grouped under General, Style & Tone and Guardrails & Boundaries. The default reply style there is "Short, focused, and professional." Your own custom instructions sit on top of those defaults.
Kapa also publishes a retrieval benchmark of its own, scored on Recall@5 against human-annotated questions from four public knowledge bases. It is the vendor's own test on the vendor's chosen measure, so take it with a grain of salt.
Our own feature list starts one step further on, where a ticket needs something done to it. Tasks and Tools let the agent call your APIs to issue a refund, cancel an order or check an account. Guidance sets the rules it follows, and Insights scores every conversation for CSAT instead of a sampled few.
How do I improve Kapa AI responses?
TL;DR: You improve Kapa by improving the documentation underneath it. The Coverage Gaps report tells you which page to write next.
Coverage Gaps looks only at the conversations where Kapa could not give a conclusive answer. Top Questions covers everything asked.
Pick a week, month or quarter, and Kapa clusters the recurring uncertain topics into two parts. I ask for the month view, because a week of uncertain conversations rarely clusters into anything.
A Finding says what people asked about and why Kapa couldn't answer fully. A Recommendation suggests how to update the documentation.
Kapa is careful about how far it will take that. Its own qualifier is that the suggestions "require human review to determine the actual problem" and the right fix. It also says not every gap needs action: some topics are undocumented on purpose, some are noise, some are out of scope.
The output exports to CSV, and the documented way to work it is one cluster at a time against your own documentation repository (a real job for somebody). Kapa offers an installable skill that runs the same analysis inside Claude Code, Cursor or Codex. A project needs a decent volume of uncertain conversations before clusters generate at all, so a small documentation set may see very little.
I put the documentation work at the top of what a documentation answer engine costs to run.
Kapa answers from the pages you connect, so weak documentation produces weak answers. The cost moves from your support team to whoever writes those pages.
The one lever outside the documentation is Custom Answers, the connector for pinning an exact question-and-answer pair against a gap you do not want to write a page for. We've written elsewhere about training AI on your knowledge base, and about the patterns that make a help center easy for an AI agent to read.
We put that improvement loop on the tickets themselves. Self-Learning compares our agent's draft with the human reply that followed a handover, then writes the missing knowledge article for you to approve. Inspect and Logs shows which sources produced any given answer, so you can see what to fix before anyone opens a documentation repository.
What deflection rate can I expect from Kapa AI?
TL;DR: Kapa counts a request as deflected when the person reads the answer and doesn't submit the form anyway. That is ticket avoidance at the form, and it measures something different from a resolved ticket.
Kapa publishes its definition on the deflection-rates page. A support request counts as deflected when Kapa's generated answer satisfies the query and the user does not then submit an identical ticket through the same form. Kapa publishes no deflection figure of its own, so the number you work with will be your own.
Someone fills in your support form and hits submit. Kapa intercepts, opens a modal with a streamed answer, and lets them ask follow-ups (the form is still sitting there behind it).
Close the modal and submit anyway, and it counts as not deflected. Walk away, and it counts as deflected.
Two gates limit any number you are shown.
Kapa generates no answer at all when the main input field holds fewer than 25 characters, "too brief to analyze effectively". It also generates nothing when the same user has already filled the form inside a cache window, two hours by default.
What you get back is a Support preset on the dashboard (Kapa's own label for the view) with four cards: deflection attempts, successful deflections, deflection rate, and "Added to your team". That last card prices each deflected ticket at your support handling rate, to estimate the capacity you got back. There is a deflection-rate-over-time chart alongside them.
I've sat on calls where this one definition changed the whole business case. Closing a modal is the event being counted. A resolved ticket is a different event, measured in a different place, so the two numbers do not swap cleanly when you compare vendors.
Our own AI resolution rate benchmark study covers 195 rated deployments across 38 vendors, with a median of 70% and a middle band of 56% to 80%. Read that as an aggregate across the field, and as directional rather than like-for-like, because every vendor defines its own headline measure its own way.
We count the other event. Our agent picks the ticket up inside your helpdesk, answers it and closes it, so what lands on the report is a resolved ticket with a customer on the end of it. It is worth pinning down which event a vendor is counting before you put two dashboards side by side.
What AI model does Kapa AI use?
TL;DR: Kapa is model-agnostic. It works across OpenAI, Anthropic, Cohere and Voyage, and runs its own models where they win.
Kapa's retrieval FAQ puts it in one line: "Kapa is model-agnostic, meaning it isn't tied to any single model or provider."
It works with "multiple model providers, including but not limited to OpenAI, Anthropic, Cohere, and Voyage". Kapa picks the best performer per job.
The reason it gives is that retrieval moves fast and it would sooner absorb that churn for you. Whether that convinces you depends on how much you want to think about models at all (most support leaders I speak to would rather not).
The data side rides along with it. Kapa says it has "DPAs and training opt-outs with all providers we work with", so your content is not being used to train someone else's model.
We run several models ourselves for the same reason, picking per task. Underneath, your sources get chopped into passages and indexed.
When a question comes in, the agent decides whether to search, the search runs on Kapa's own servers, weak matches get dropped, and the answer comes back with its source links attached. The raw passages never reach the browser.
A customization is tied to a chat model version at creation and cannot be moved afterwards, so your tuned assistant does not change behavior when Kapa releases a new version.
Our own mix is chosen by job: small fast models handle routing, classification and summarizing, and a larger one writes the reply your customer reads. Reasoning models stay out of the live path, because they are too slow for a chat somebody is waiting on. We A/B test the mix against real conversations, so it moves when something better turns up.
What languages does Kapa AI work in?
TL;DR: Kapa states 20+ languages and answers across them, so a French question against English docs comes back in French. English is still where it is strongest.
The claim is published on the Zendesk app's multilingual feature page: "Kapa supports 20+ languages, including English, Spanish, French, German, Italian, Portuguese, Russian, Chinese, Japanese, Korean, and more."
I like where Kapa put the language controls. They sit in the editable sections, and the default arrives switched on: respond in the same language as the question, falling back to English.
The cross-language behavior you would want is on out of the box.
The capability comes from the underlying models and from that language-detection default (our own agent detects the language per message and replies in it). Kapa has not localized the product market by market, and its own default names English as the fallback.
We cover 95 languages, detected per message, so a customer who switches mid-thread is answered in the one they just used. On Intercom, Live Translation puts the conversation into your agent's own language as an internal note, then translates their reply back on the way out. Handover stops being the point where a multilingual thread falls apart.
How secure is Kapa AI?
TL;DR: SOC 2 Type II, zero data retention, SSO and access controls, with one setting to check before rollout. PII masking, which strips personal data out of messages, is off until you turn it on.
Kapa's certifications page states it directly:
"kapa.ai is SOC 2 Type II certified." — Kapa AI product documentation
Kapa says the certification covers security, availability, processing integrity, confidentiality and privacy, assessed against AICPA requirements. Everything else goes through its Trust Center, which takes document requests at a named security address.
This is the question our own security reviews turn on: does the indexed content train someone else's model? Those DPAs and opt-outs cover it, so nothing you index trains a third-party model.
The setting I check first arrives switched off. Kapa's PII masking documentation states that "PII masking is not enabled by default."
You turn it on per project, choose which personal data to strip, and add your own patterns with a regular expression.
When it is on, user messages are scanned the moment they arrive, before anything is stored or sent to a model, and only the masked version is written to storage. It covers the website widget, the Support Form Deflector, the Slack and Discord bots, the chat and retrieval endpoints, and the Agent SDK.
On the Zendesk Agent, user-message masking is not applied (the question is coming from one of your own agents). Raise that with your security reviewer before they find it. Knowledge-source masking is separate and stricter: personal data found while crawling is discarded permanently and replaced with anonymized labels.
Residency and the rest of the paperwork sit on the Enterprise tier. It names SSO, SCIM, advanced access controls, audit logs, retention controls, and EU and regional hosting.
Its security pages name SOC 2 Type II as its certification and point at the Trust Center for anything further. HIPAA and ISO 27001 are not advertised, and no encryption detail beyond transit encryption is published.
We hold SOC 2 Type II and GDPR ourselves, with AES-256 encryption at rest and TLS in transit, and nothing you send us trains a model or gets used for anything beyond serving your own tickets. Our live trust report and security documentation are public, so a reviewer can read them before the first call.
Who is using Kapa AI?
TL;DR: Documentation-heavy technical companies. Mapbox, CircleCI and Planet Labs all publish numbers, and the buyer is usually a DevRel or docs owner.
Three named customers on Kapa's customer-stories page carry a published figure each:
- Mapbox cut monthly technical support tickets by 30%, running Kapa across a Discord bot, a Slack bot, Zendesk and the docs widget.
- CircleCI improved response times by 28% with AI chat, deployed on Zendesk and in its docs.
- Planet Labs reports 99.1% CSAT across more than 50,000 queries, using a docs widget and an internal technical assistant.
The wider list is more of the same crowd: Logitech, Netlify, Monday, Coralogix, Redpanda, Silicon Labs, Espressif, Nordic Semiconductor, Nokia and Reddit. Docker is a Kapa customer too, named on its Y Combinator profile, and publishes its own account of it. Kapa's homepage says it is trusted by more than 200 technical companies.
The deployments name the buyer: documentation widgets, community bots, internal assistants and a Zendesk sidecar. That is a developer-relations, documentation or support-engineering owner.
It is rarely the support manager with a ticket backlog and a headcount problem I spend my days with.
On our side, 200+ ecommerce and SaaS businesses run My AskAI, and our agents have resolved more than 1,000,000 tickets. The customer stories are collected on our reviews and case studies page.
Does Kapa AI have a free trial?
TL;DR: Yes, 14 days, with one index and limited usage. There is no free-forever plan, and anything past the trial runs through sales.
The free trial on the pricing page says "Try Kapa for free for 14 days." You get one index with limited usage, plus the retrieval API and hosted MCP server. Cited context and the prebuilt agents and integrations are included.
I give Kapa full marks for the way in: it asks you to connect your sources and hands you an API key or a hosted MCP server at the end of it.
Kapa publishes no free-forever plan. I think fourteen days is tight for a documentation rollout. It also offers a second way into the trial, asking only for a company email and promising you will "Go live in less than 7 days".
My AskAI's trial is 30 days, with every feature unlocked, unlimited tickets and no card required. You can run it in internal-notes mode alongside whatever you have now, so nobody has to bet the inbox on a demo.
Is Kapa AI worth it?
TL;DR: If your support is developers asking technical questions against a large documentation set, Kapa is purpose-built and hard to beat. If your tickets are refunds, billing and account changes landing in an inbox, it is the wrong tool.
Kapa AI is worth it if:
- Your questions are technical and the answers already live in documentation, code, API references or community threads.
- You want the answers where developers are: in the docs, in Slack or Discord, in the product, and through an MCP server their coding assistant can call.
- You want somebody else to own the retrieval work, from indexing your content through to finding the right passage for a question.
- Someone on the documentation or developer-relations side will act on the Coverage Gaps report, because that is how the whole thing improves.
Kapa AI is not worth it if:
- Your tickets land in an inbox and need an agent to pick them up, reply as you and close them. Kapa's own words: "the agent never clicks Submit."
- Your helpdesk is anything other than Zendesk.
- Your ticket mix is transactional: refunds, billing, account changes, order status.
- You need to know the price before you talk to somebody.
The misunderstanding I want to name is teams reaching for a documentation answer engine and expecting a support agent. Those are two different products that happen to share a demo. If the second one is what you're after, we've ranked the alternatives for support teams in a separate roundup.
This verdict rests on seven things: what Kapa is, how it installs, where it runs, its limits, what it trains on, how it improves and how it handles security. Evidence comes from Kapa's own product documentation, its pricing page, its published customer stories and its Y Combinator profile.
Seven out of ten is where I land. Grounded, cited answering with uncertainty reporting baked into a harness the customer cannot weaken is rare, and so is an improvement loop that names the missing documentation page for you.
Against that, it stops at the helpdesk door and you cannot price it without a conversation. Those two things hold it below a nine.
My AskAI lives inside Zendesk, Intercom, Freshdesk, Freshchat, Gorgias and HubSpot, replies as your agent on the ticket, and hands over to a human with a summary when it can't finish. If you want the two products side by side, our My AskAI vs Kapa AI comparison does that in detail.
What are the pros and cons of Kapa AI?
Pros
- Grounded, cited answers with uncertainty reporting. The low-hallucination claim is backed by the architecture: retrieval first, citations attached, and uncertainty reporting sitting inside a harness the customer cannot override.
- The interaction pattern developers already use. Answers in the documentation, in the product, in Slack and Discord, and through an MCP server their coding assistant can call. Eight prebuilt surfaces, one agent behind them.
- An improvement loop you can act on. Coverage Gaps names the documentation page that is missing and exports to CSV. Kapa says up front that a human still has to decide what to do about each one.
Cons
- It never replies as your agent inside the helpdesk. The Zendesk app stages a draft that a person has to send.
- Zendesk is the only helpdesk on the list, which rules Kapa out for any team on another helpdesk. No Intercom integration, no voice, no email-inbox auto-reply, and no connector for an Intercom, Freshdesk, Gorgias or HubSpot help center.
- Nothing transactional. Refunds, billing, account changes and order status sit outside the published surfaces unless a developer builds them in.
- You cannot price it without talking to sales. Both paid tiers say "Talk to us", so a budget number needs a meeting.
Kapa AI
- Brand: Kapa AI
- Rating: 7/10
- In a sentence: Close to the best thing available for answering technical documentation questions, and the wrong tool for resolving support tickets.
If you want a full comparison of My AskAI against Kapa AI, we've written one.
FAQs
What is Kapa AI used for?
Kapa is used to ground AI answers in a company's own technical knowledge. In practice that means a documentation assistant customers self-serve from, community bots in Slack and Discord, an in-product agent, an internal assistant for staff, a pre-ticket support-form deflector, and a retrieval layer that other agents call through an API or MCP server. The documentation widget is the cheapest of those to stand up, and the rest deploy from the same indexed sources.
Is Kapa AI an AI chatbot for documentation?
That is one way to use it. The documentation widget is a single deployment of Kapa's prebuilt agent, and its own docs describe those agents as "the ways to deploy it". Kapa now describes itself a layer below that, as a knowledge retrieval API any agent can call. In practice I'd still describe it to a colleague as the docs assistant.
Which helpdesks does Kapa AI integrate with?
Zendesk. Kapa's published integration list names no other helpdesk, and its two support-facing surfaces are the Zendesk sidebar app and the Support Form Deflector, which catches questions on your support form before a ticket exists.
Can Kapa AI reply to customers in my helpdesk inbox?
No, and Kapa says so itself: "the agent never clicks Submit." What it does do is read the full ticket, draft a grounded reply into the composer, and stage a summary as an internal note. A human reviews and sends. That is the line we would draw between agent-assist and an AI agent.
What's the difference between deflection and resolution?
Deflection counts a support request the customer never went on to file, and resolution counts a ticket that was answered and closed. Kapa's published measure is the deflection one, defined as the person not resubmitting the form.
Deflection | Resolution | |
What gets counted | A request the customer never filed | A ticket answered and closed |
Where it is measured | On your support form, before a ticket exists | Inside your helpdesk, after one exists |
When you compare vendors, check which of the two each headline number describes, because in our experience they are rarely the same thing.
What's the difference between AI chatbots and AI agents for customer support?
A chatbot answers a question. An agent acts on it: looks the data up, performs the step, tags and routes the ticket, and hands over with a summary when it can't finish. We do the second job through Tasks and Tools, Guidance, tagging and human escalation inside your helpdesk.
What knowledge sources can Kapa AI connect to?
Kapa's connector catalog names 23 connectors. Alongside web crawling and file upload, it covers Zendesk Tickets, Zendesk Help Center, Salesforce Knowledge, Salesforce Cases, Confluence, Notion, Google Drive, Jira, Jira Service Management, Slack, Discord, Discourse, Stack Overflow, S3, OpenAPI, YouTube, Custom Answers, and GitHub code, issues, pull requests and discussions.
What languages does Kapa AI support?
Kapa states that it "supports 20+ languages, including English, Spanish, French, German, Italian, Portuguese, Russian, Chinese, Japanese, Korean, and more." The published default is to reply in the same language as the question, falling back to English, so you get the cross-language behavior without touching a setting.
Is Kapa AI SOC 2 compliant?
Yes. Kapa's security documentation states that it is SOC 2 Type II certified, covering security, availability, processing integrity, confidentiality and privacy, with the supporting documents available through its Trust Center. Ask for the report there. Every reviewer we deal with does.
Does Kapa AI offer a free trial?
Yes, 14 days. It covers one index with limited usage, plus the retrieval API and hosted MCP server. There is no free-forever plan, so anything beyond the trial goes through the two paid tiers, both of which need a conversation with sales before you see a number.
How much funding has Kapa AI raised?
Kapa is a Y Combinator company backed by a seed round led by Initialized Capital. Its Y Combinator profile lists a team of 22.
