Emil

And we're live! Check out the @TechCrunch article here: techcrunch.com/2024/10/03/how… :)

Lots of companies deploying really, really bad and unhelpful AI agents. Meanwhile, Logitech quietly launches an AI assistant that's 99% accurate. A year later, tickets per unit sold are down 10%, help center content grew 30%, and sales starts using it to answer technical questions on calls. When AI works, people actually use it. When it doesn't, trust evaporates.

"We need to clean up our docs before we launch the AI assistant." This is the easiest way to waste 6-12 months. And I say that as a huge proponent of clean docs. AI has gotten so good, if you point it at your docs, it tells you exactly what to fix. That's because users will tell you. They'll talk to the AI, and the AI will surface questions it can't answer. You'll learn more from a day of user questions than a year of "improving the docs." Trust me. Companies don't improve docs and add AI. They add AI, and that improves their docs.

Talked to a Fortune 500 tech company last week that's desperately trying to pull their users AWAY from ChatGPT and towards their official docs. It's an uphill battle. The harsh truth is most engineers ask an LLM before visiting your docs page. But that creates a vicious cycle. 1. Customers ask ChatGPT instead of the docs 2. ChatGPT has outdated, unreliable, generalized information 3. Customers lose trust in the company because they never get the right answers I told the F500 company the only way to break out is to create a better AI experience. An AI trained on your product knowledge is a week or two away, maximum (if you partner with @kapa_ai) and it builds actual trust. Your users are already using AI to learn your product. The question is whether they're using one that actually knows your product, or one that's guessing based on two-year-old training data.

Last week was special for @kapa_ai. GSMA Insights demoed our custom AI assistant, live on the demo floor, to industry-leading companies like Google, Nokia, Ericsson, Qualcomm, and Samsung. You may have heard of MWC, and GSMA is the reason you see incredible new launches out of Barcelona every year. Every attendee is grateful for their hard work, ourselves included. They do more than put on the best conference on the planet, though. They make industry-leading reports on the global mobile ecosystem. Those reports hold insanely valuable (and technical) insights. We created a custom Kapa AI assistant on top of those reports, so anyone can ask questions and get the insights. It was a huge honor to see GSMA demo this AI assistant to their customers, and see the reaction from the conference. Thanks to Matthew Iji and Emanuel Kolta for letting us show you how great an AI assistant trained on GSMA reports could be. Can't wait for next year!

Have you ever noticed out-of-the-box LLMs almost never say "I don't know"? To me, that's a huge red flag if you (or your customers) are using it for production code. It confidently gives you something that sounds right but will break your production environment. A networking hardware company told me this when they were evaluating AI assistants: "If the answer is not in the documentation, the answer is made up and is plausible. Even if not necessarily correct." Sums it up perfectly. "Plausible" doesn't cut it when customers are associating an LLM's answer with your brand.

"AGI is right around the corner" I asked ChatGPT to transcribe a voice message, it said it couldn't do it. I asked Claude to do it, it tried to install OpenAI. We're still early.

A fintech company handling 32,000+ documentation questions per week described their current AI search solution as "fairly amateur." That's why it fell flat. "Technically works" isn't going to cut it for banking operations, when a hallucination means a compliance issue. Generic LLM pointed at complex product docs will never work for this reason. The LLM needs to actually understand the semantic structure and content of the docs!

🔥 Thrilled to announce $250 billion @SAP x @kapa_ai is official - SAP's Gardener just deployed an Ask AI assistant on top of their developer docs. This partnership means a lot to me. If you work in the enterprise, you know SAP. If you live and breathe ERPs, you know SAP. Not as many know they also quietly maintain Gardener - an open-source managed Kubernetes service that's quietly growing like crazy. It also has crazy technical docs. Technical documentation needs an equally technical assistant, and SAP knows that better than anyone. They're trying to grow Gardener's adoption, and in the age of AI assistants, developers will drop a product if docs are inaccessible, too technical, or too hard to use. That's not an issue for Gardener. Developer adoption will flourish with their Kapa AI assistant - it's trained on their docs, doesn't hallucinate, and knows when to say "I don't know." Well done SAP and Gardener. Excited for the partnership!

Talked to 5+ companies that tried using Intercom Fin for technical knowledge in the last week, and it's a bloodbath. One said it "gets confused distinguishing between technical and non-technical customers." Another called it "inadequate for developer questions." I've worked on this problem since 2 months after ChatGPT launched. I saw this coming a mile away. It's simple. General tools don't understand specific problems. A horizontal customer support tool doesn't understand complex technical products. There is a massive, massive difference between "How do I add a new email" and "How do I configure OAuth2 scopes for multi-tenant environments." It's an expensive mistake to think they're close to the same.

@kapa_ai docs.kapa.ai/data-sources/g…

By the end of this week (just 5 days), at least one of you will solve your AI problem thanks to today's launch. Google Drive connector within @kapa_ai AI assistant. Let me explain. Your company's brain lives within Google Drive. Postmortems, onboarding guides, troubleshooting FAQs, etc. Your team needs this every day, and they can never find it. We're fixing that today. Your new AI assistant connects to everything - your public docs, API references, GitHub repos, support content, Google Drive, and more. (Plus connections to Notion, Confluence, Zendesk, GitHub, and 20+ other sources that Kapa already supports.) Now, you have an AI assistant that has a full picture. What this unlocks: 𝗦𝘂𝗽𝗽𝗼𝗿𝘁 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝘀 use Kapa to pull answers for customers from your public docs AND the internal troubleshooting runbook in Drive. 𝗖𝗦𝗠𝘀 ask AI "What's the renewal playbook for enterprise accounts?" and get answers instantly. 𝗔𝗘𝘀 get an AI assistant that understands the complete 360 of a customer. They get this AI assistant through our Slack bot, via our portal, or plugged in to any AI tool via our MCP server. This is one AI assistant that becomes your team's second brain. And it takes less than a day to set up. Less than a week to solve your AI problems. Check it out below - our existing customers like OpenAI, Docker, and Reddit are huge fans.

I'll say it, Slack is where technical knowledge goes to die. But thanks to AI, it's actually your biggest opportunity. Slack is where work actually happens. Technical debates, troubleshooting, decision-making, etc. That's exactly the problem. As soon as you scroll, that knowledge might as well be gone. So you ask the same questions 50 times. Literally everyone I know has done this (including me!) If you're running customer support through Slack, it ends up being a never-ending thread of the same questions you've answered before, or in your docs. Probably the most fragmented knowledge base in your company b/t DMs, threads, group chats, etc. But it's still a huge opportunity for that same reason. Your best documentation is already written in Slack! Unifying all that knowledge is one thing, and you can do that by ingesting all of your messages securely. Still, customers are so used to AI chat, you can't just dump answers in a troubleshooting guide. The hack is an AI system that ingests Slack data → unified knowledge base by combining with other data (like docs, GitHub) → AI assistant. That's what our customers are doing, and they're escaping Slack hell. Biggest opportunity for technical orgs in 2026 is taking control of your Slack Connect channels, and letting them work for your company. They're controlling your support team otherwise.

@tannerlinsley @kapa_ai 100% we're reaching out :) Part of the team is at Mobile World Congress today so we're a little slower than usual - sorry!

Would anyone from @kapa_ai be willing to chat with me today?

Feeling grateful today. The kapa team is growing fast post @YCombinator. AI agents for technical knowledge bases are changing how our customers run their companies. There's a lot of uncertainty right now around AI's impact. Will teams shrink? Where do humans fit in? We have 200+ customers, including huge companies like Grafana, Reddit, n8n, HPE Juniper Networks. Their product, support, AI enablement and documentation teams aren't laying people off. They're growing faster than ever too. They get time to focus on real strategic work. Agents handle the rest. Our team worked hard to make this happen. Bullish on 2026 for AI agents. Mega-bullish on @kapa_ai. (We're hiring, btw!)

Read the full guide here: kapa.ai/blog/driving-u…

Most MCP servers sadly never get the chance to take off. I studied 10+ unicorn startups with broad MCP adoption, and found the 3 strategies they use to onboard developers. This is the hard part. No matter how much of your knowledge base is served via MCP, the hard part comes after the server goes live. Here's what I recommend. 1. Add to your AI chat widget This is the highest-leverage place to promote your MCP. Your developers are already using your "Ask AI" widget on your website. They already showed you they want AI-powered answers. Adding a built-in MCP install menu makes it easy to one-click install MCPs in Cursor, VS code, or any other tool. ClickHouse, n8n, and Temporal all do this well. 2. Write a "Build with AI" page I've noticed this is a huge difference maker for our customers. You need a permanent home for your MCP server. The most effective approach we've seen is a dedicated documentation page that explains what your MCP server is, why it's useful, and how to set it up in every major AI tool. It should cover a clear explanation of what connecting does, setup instructions for the major tools, and example questions to try. You can honestly just follow RedPanda's lead here - they did an excellent job. 3. Announce it! Adoption will be much slower if you don't promote what you've made. Planet Labs posted an announcement in their community forum. Silicon Labs posted on LinkedIn. Expo created a full walkthrough on YouTube. That's not to mention newsletters, webinars, and changelogs. Remember, you're playing the long game here. Every new developer who joins your community, reads your docs, or attends your events is someone who hasn't heard about your MCP server yet. Keep mentioning it. We wrote a full guide on this below - check it out to supercharge your MCP adoption.

For retrieval endpoint: docs.kapa.ai/api/reference/… For 1-click MCP server: docs.kapa.ai/integrations/m…

Ever wonder why Claude Code is so good? It has access to your codebase. It knows your project. It has context. Now every company is trying to build the same thing for their product. Companies like Grafana and Amplitude have built AI sidebars that can do things like create_dashboard(), run_cohort_analysis(), and more. But here's the thing. Just like Claude Code needs your codebase, these copilots need to know your product. And that knowledge? It lives scattered across your docs, API references, help centers, wikis, GitHub repos, and forums. Without it, your copilot can't reason. It just guesses. We've solved it with one simple call. POST /retrieval { "query": "what alert evaluation intervals are supported?" } That returns the most relevant chunks from your entire product knowledge base, ranked, with sources. Behind that one call: state of the art low latency multi step agentic retrieval with query decomposition and reranking. 50+ managed data source types, auto synced, always fresh. Optimized for product knowledge. Hardened across 200+ production deployments over 3 years. You don't see any of that. You just get the right chunks back, fast. Plug it into your copilot as an API call. Or use search_product_knowledge() as a hosted MCP tool call. Either way, your agent goes from generally smart to product expert. You build the copilot. We handle the knowledge layer. Link below to get started.

60% of executives worry about AI hallucinations. I see results like this, and I don't blame them. But it's a misconception that you can't do anything about it. We've deployed hundreds of customer-facing AI agents on top of technical docs, and our customers don't worry about hallucinations. There's a lot of work we do to get there, but my number one tip is simple. Let the AI say "I don't know." Too many hallucinations have this root cause. If your AI assistant can't say "I don't know", it will make up things to please the user. @kapa_ai knows when to answer with facts, and knows when to admit it doesn't have facts. Overlooked aspect of hallucinations, IMO.

Takes 30 seconds to upgrade. Follow this link: #do-i-ever-need-to-update-my-slack-bot-installation" target="_blank" rel="nofollow noopener">docs.kapa.ai/integrations/s… :)

🔥 We just shipped 3 HUGE updates to the best AI agent Slack bot. Kapa 2.0 gets a massive upgrade. We literally rebuilt it from the ground up, and here's what's new: 1. Drop files directly into the chat. Logs, screenshots, RFPs, CSVs, whatever context you need. Over 100 file types supported, so your AI agent can see what you're seeing. 2. Tag "@kapa" into ANY thread. Yes, any thread. See a question in Slack that Kapa could answer? Just tag it. It will read the full thread context and jump in. 3. Real-time streaming responses. Just like ChatGPT, we stream answers live. See the answer as it appears, not after 10 seconds of waiting. Sounds small, feels huge. Companies like Airbyte, and Mixpanel use Kapa's Slack bot to handle hundreds of community questions every week. Coralogix uses it internally to unblock support engineers and CSMs in real time. This is the biggest update we've ever shipped for it. Your team will notice the difference immediately. Use the link below to upgrade - it takes 30 seconds.