Launch HN: Inkeep (YC W23) – Copilot for Support (think Cursor for help desks)
It’s a conversational sidebar you can use as an app for Zendesk or Chrome Extension for any support platform. There’s a demo video at https://vid.inkeep.com/cx-copilot or you can try the live sandbox with example tickets at https://try.inkeep.com/cx-copilot.
Why? Most AI support tools today are focused on trying to have AI answer customer questions before they even reach support teams (‘deflection’). We’d focused on that too. However, we heard from many of our customers that while they care about deflection, they care even more about providing high-quality, fast human support when users need it. Some teams don’t even want customer-facing AI at all and just want AI tools to help their team be more efficient. We created Keep with these scenarios in mind.
Keep does a few neat things we haven’t seen elsewhere:
1. Provides intelligent suggestions: if Keep is confident, it’ll create a draft answer and tell you the sources it used. If the ticket is long, it’ll summarize the conversation so far and outline the remaining to-dos. All automatic and contextual to the ticket.
2. Is fully conversational: ask for clarifications, revise draft answers, and iterate as needed.
3. Uses ‘Generative UI’: suggestions are rendered as glanceable, interactive UI components. For example, a draft answer has buttons like “Shorten” & “Concise” that prompt the AI to revise the answer. UI components are interweaved within normal text.
4. Turns tickets into FAQs: can generate an FAQ from a closed ticket and lets you iterate on it and save it when done.
5. Leverages many content types: uses your docs, help center, previous support tickets, Slack threads, etc.
We were inspired by tools like Cursor, Claude Artifacts, and v0. These experiences go beyond plain-text conversations by interweaving interactive code blocks or UI previews into their answers. This makes answers digestible and intuitive (and fun) to iterate on.
Some technical details, for those interested: We use the Vercel AI SDK to optimistically stream the React components by using our Chat APIs, which are powered by Claude Sonnet 3.5 and our RAG service. Our APIs follow the OpenAI chat completion format so are generally compatible with any LLM tooling. Our `inkeep-qa` API generates draft answers and the `inkeep-context` API generates structured outputs and tool calls (docs: https://go.inkeep.com/ai-api). For an example of how these APIs are used, check out our Intelligent Support Form example (demo: https://try.inkeep.com/ai-form, repo: https://quickstart.inkeep.com/ai-form).
If you want to try Inkeep on your product's content, just fill out the form in our landing page. You’ll get a demo in your inbox powered by your public content — NO “call us”, “book a demo”, or “schedule a meeting” required. Note: we do check that your email domain matches your content to prevent spam.
Curious to hear about your experiences when working with customer/support questions and any ideas on how else the copilot could be useful for those scenarios.
61 comments
[ 5.1 ms ] story [ 139 ms ] threadhttps://aws.amazon.com/q/business/
Our tool focuses on support use cases (customer-facing or internal-facing), which means we can go deep with workflows like detecting gaps in your documentation and focusing our efforts on quality around these scenarios. Generally our support copilot intro'd here also generates dynamic UIs so goes beyond a normal chat interface.
tl;dr we define a JSON schema with a few semantic labels that represent a gradient in confidence. On our end we prompt it with certain examples and guidance for when to use each label. This is generally a better approach than e.g. asking an LLM to give a numeric score.
We also have trained embedding-based classifiers as non-LLM heuristics.
Inkeep works great at Pinecone and meaningfully reduced the number of support tickets with common questions/issues.
[1] https://support.pinecone.io/
Last week, I heard about a company (I think it was an YC company) describing themselves as "open source Cursor".
Also last week, a comment here on HN stuck with me: "I live inside Cursor" https://news.ycombinator.com/item?id=41651380
And now, the Cursor for Help Desk.
I've used Cursor, and I loved it. 10 to 1 over regular GitHub Copilot, and well worth the $20 dollars and I am a hobby programmer (management job during the day).
But... all this? It has become the reference point just like we had "Uber for...", "AirBnb for...". It seems like it happened so fast.
Number one was the "apply" suggestions from chat, but now I think Copilot has it too.
Number two are the suggestions while I am typing to change multiple lines at once. Such a time saver.
Number three is how I can send entire files/folders as reference in the chat.
Also, it feels a bit snappier? The suggestions seem to come a bit faster than Copilot.
In terms of correctness / good good, they're both equally good (and bad).
To me it's all about how much time I am saving. When I sit down 30-90 minutes 3-4 times a week to code, I just feel like it helps me to get more stuff done.
VS Code may have caught up now, but haven't looked back.
Robert spoke at our meetup and is awesome. https://www.youtube.com/watch?v=35JdjmiDvWI
we answer 250k+ customer-facing questions/mo today for teams who really care about quality (Anthropic, Clerk, Pinecone, Postman) - we're brining that same care and high bar to our copilot for support teams.
the generative UI and conversational aspect is quite different than other copilots we've seen.
I'm not sure how they do it but the answer quality and the UI is meaningfully better than all the other "chat with your docs"-type products I've tried.
In other words the promised outcome isn't very original but they've nailed the execution.
1. FAQ/Knowledge bases with search functionality.
2. Conversational mediums and agent notifications (e.g., live chat widget, messenger support).
3. Ticket management systems and agent management, which is the core of Zendesk/Intercom. This is the most difficult to operationalize as it requires process architecture, SLA management, etc.
4. Orchestration and workflow management, which can be done inside #3, though some products are available as well.
Most new post-LLM startups target #2 but face platform risks as they rely on companies covering #1, #3, and #4 (e.g., Zendesk, Intercom, Gorgias).
I feel InKeep doing some combination of #2 but emphasising that you can support client whenever they are (ie Github, Discord, Slack) instead of asking them to submit tickets in the website widget.
Another issue for AI support startup is the verticalization/horizontal trap. Most LLMs require solid tuning per client, especially for enterprises like us. Startups often avoid this initially, opting for a more horizontal general path (e.g., AI support for Shopify merchants). This is where enterprise players are more beneficial. Thus companies like ServiceNow, Zoom, and Oracle offer products for support and implementation services.
Neat business imo.
https://docs.inkeep.com/faqs/comparisons
Might help.
how do you ensure that companies don't use this to make it impossible to actually contact a human? or do you feel that isn't in scope for a product that's encouraging companies to make it impossible to contact a human?
Agree customer-facing AI has to be done in a tasteful and mindful way.
"Escelating to humans from Inkeep AI Slack bot" : "Escelating" --> "escalating"
Upvoted – looking forward to supporting you guys more
Pretty sure most people will go to whoever has a free tier, and even that space will be competitive.
Nick & Rob are very strong leaders.
The learning I've had is that whilst the majority of queries go through standard search patterns (i.e. users search for something that's covered by documentation), a subset of queries are not answerable by our documentation but only implied by it. I've direct experience that Inkeep is serving a large part of that user segment and reducing our support burden.
As a very recent/specific example from last week, we had a community user generating a terraform provider for an internal use-case. By putting error messages from our CLI tooling into Inkeep's "Ask AI" feature, they discovered a nuance in "x-speakeasy-match" (the error message implied it created a circular reference, but didn't spell that out) and self-served a solution.
Inkeep effectively turned our documentation into a guided tutorial on our product, specific to the customer. Pretty strong ROI.
https://ai.gov.uk/projects/caddy/
Our ultimate goal was to make our experience explicitly not feel like you're talking to AI.
So rather than trying to intercept questions from being posted to our forums, we trigger Inkeep _after_ a question is posted. If we're able to find an answer with a high degree of confidence, our "AI user" (Max) will show an answer within about 30 seconds.
The OP can then provide feedback that we're using to train further answers.
If the answer is marked as helpful, we display the answer publicly (and disclaim it as an AI response)[2]. If the answer is marked as _unhelpful_, the answer only shows to the OP and we review the feedback to figure out how we can improve (ie: do our docs need to be improved so Inkeep has better source material?).
It's been fun getting creative with the Inkeep team on a solution that worked for our specific use case. I'm planning on rolling out Inkeep more broadly in other areas of our site as we verify that our highly confident answers are genuinely useful to our users.
IMO Inkeep has been the first AI solution that hasn't sucked – and that's high praise coming from me!
[1] posthog.com/community [2] https://posthog.com/questions/autocapture-event-bubbling