← All transcripts

Circleback CEO Ali Haghani: Recording Company Meetings Will Become The Norm Transcript, AI Summary & Key Points

Y Combinator · 4 days ago · Science & Technology · 15:17 · EN-US

💡 Answer

Yes. As LLMs become more capable and AI agents do more company work, companies will increasingly default to recording and sharing conversation context, provided they can control who receives sensitive information.

🧠 AI Summary

AI agents are shifting work from manually writing code and handling operations toward orchestrating, reviewing, and improving agent output. Company conversations are becoming an important source of context because agents otherwise work in isolation. As software becomes cheaper to produce, deciding what to build and designing interoperable systems matter more than shipping speed. AI use should remain subject to human review for emails, production copy, data access, security, and architecture.

🔑 Key Points

  • CircleBack records and transcribes meetings, writes notes, assigns action items, supports automations, and lets teams query their conversations as a company memory.
  • Ali spends approximately 10-20% of his day with OpenClaw because it works continuously, while he spends more time reviewing code than writing it.
  • AI agents handle most end-to-end coding work at CircleBack, while humans and specialized agents focus on code, architecture, product, and copy review.
  • CircleBack uses evaluations to test whether prompt and model changes improve outputs without creating regressions, such as redundant action items or overly long topics.
  • Ali does not allow agents to send emails independently, put unreviewed copy into production, architect systems involving data access and security, or make key architectural decisions.
  • The opportunity cost of not recording company conversations is already high because agents lack context about earlier discussions and customer feedback.
  • As software production becomes cheaper, the most important engineering decisions concern what to build and how to design systems that can interoperate.
  • Ali is consciously reducing his coding time because the company's bottleneck is bringing on more strong people and increasing the throughput of product decisions.

✅ Actionable items

  • Use evaluations for outputs with quantifiable right-or-wrong criteria and rerun them after model or prompt changes.
  • Prevent agents from sending emails autonomously; allow them to prepare drafts for human review.
  • Require human ownership of architecture for work involving data access and security, then use agents to implement the design.
  • Reserve a no-meeting day for focused work and aim to ship the item being worked on end-to-end that day.
  • Automate repetitive operations work so more time can be spent on high-leverage responsibilities.
  • Design AI workflows so each task makes the next execution easier or produces a better output.
  • Record relevant company conversations while controlling access so captured information is shared only with the appropriate people.

💡 Business ideas

AI meeting intelligence and company-context platform00:15

Record and transcribe company conversations, turn them into notes and action items, make the information searchable, and expose relevant context to people and AI agents while controlling who can access it.

For
Teams and companies, including traditional businesses, growth-stage startups, and Y Combinator companies.
Solves
Important customer, employee, interview, and company context is scattered across conversations and unavailable to AI agents, leaving them to work in isolation and making follow-ups and company-wide visibility difficult.
  • CircleBack's interview workflow: unresolved interview action items are used to track who needs a response and which employees are interviewing candidates.
Soon you can unlock the full business plan.

Behind this: 14 build steps · 4 tools and how each is used · how to validate demand · 3 more real examples · 6 things the video never answers.

Inquire for details

Soon you can unlock the full business plan.

Inquire for details

🤖 AI in practice

Used for

Record and transcribe meetings, write notes, and assign action items. 00:15
Extract custom information from meetings and update business applications such as a CRM, issue tracker, or Slack. 00:25
Answer questions across a company's recorded conversations so the meeting history functions as a company knowledge base. 00:35
Track recruiting follow-ups and interview progress as an applicant-tracking workflow. 00:56
Prepare a founder for an interview by compiling relevant context and suggested topics. 04:18
Capture an impromptu thought or conversation without handling a phone. 04:49
Review customer conversations, retrieve requested product feedback, and draft follow-up communication. 07:05
Investigate customer emails and inbound sales leads through shared company workflows. 05:28
Generate and implement software changes end-to-end, then have the results reviewed. 08:41
Review pull requests for code quality, architecture, product behavior, and copy consistency. 09:11
Evaluate and improve meeting-note prompts and models without introducing regressions. 09:41
Automate repetitive founder operations work. 11:55

Agents

  • OpenClaw — Handle ongoing company workflows for finance, customers, people operations, customer emails, and inbound sales. 2 held 05:28
  • Coding orchestrator — Launch multiple software-development tasks for Ali to review later. 2 held 08:41
  • PR review agents — Review pull requests for technical and product consistency. 2 held 09:11

Advice

  • Record company conversations by default and make the relevant context available to the people and agents that need it. for Companies and teams using AI agents.
    As LLMs and agents do more company work, agents need context about earlier discussions and customer statements instead of working in isolation.
  • Put evals in place for AI behavior that has a quantifiable right or wrong answer. for AI product teams and developers.
    Evals catch redundant or incorrect outputs, such as marking a task as an action item after it was already completed during the meeting.
  • Check AI outputs after every prompt or model change and use incremental evaluations to prevent regressions. for AI product teams and developers.
    Improving one part of an output can worsen another, creating a whack-a-mole cycle without systematic checks.
  • Use AI extensively to automate operational work and share effective tools and workflows with the team. for Startup teams and company operators.
    Automating low-value tasks frees people to focus on higher-leverage work, while internal show-and-tells spread useful practices.
  • Keep humans responsible for architecture and require review before AI-generated work is accepted. for Software engineering teams.
    AI agents can build out an implementation, but the engineer or other responsible person should design the architecture.

What it could not do

  • Agents must not send emails autonomously. — Ali permits agents to prepare drafts but requires a person to review and send them.
  • AI-generated product copy cannot go directly into production without review. — Agents may produce a first pass, but the team requires human review for clarity.
  • Agents are not allowed to independently architect systems involving data access or security. — AI can review this work and build an implementation after a human has designed the architecture.
  • Recording and sharing conversation context creates a risk of inadvertent disclosure. — Companies need confidence that captured information is shared only with the right people, especially for private one-on-one conversations.

🧰 Tools & AI usage

  • CircleBack — AI note-taker and company conversation system used to record and transcribe meetings, create notes and action items, search conversations, prepare for meetings, and update connected apps.00:15
  • Slack — Used for shared customer-support and inbound-sales threads, investigations, and company visibility.00:06
  • OpenClaw — Used to run always-on agents for finance, customers, and people operations.00:05
  • Telegram — Used as the environment for creating and interacting with separate operational agents.00:05

AI is used for

  • Capture and organize company conversations — Record and transcribe meetings, produce notes, assign action items, and preserve company context.00:15
  • Prepare for meetings — Combine emails, calendars, and past meetings to generate interview preparation.00:18
  • Support customer and sales operations — Investigate customer emails, organize inbound sales work, and make activity visible to the company.00:14
  • Write software end-to-end — Have agents implement work while humans revisit and review the results.00:39
  • Review pull requests — Check code quality, architecture, product decisions, and copy for consistency.00:09
  • Evaluate prompts and model changes — Measure output quality and prevent regressions in notes and action-item extraction.09:21
  • Automate operations work — Reduce miscellaneous founder tasks and increase time spent on higher-leverage work.11:58

📄 Transcript

Searchable transcript of Circleback CEO Ali Haghani: Recording Company Meetings Will Become The Norm — Y Combinator (15:17). Search for a phrase, then click its timestamp to jump straight to that moment in the video.

Captions sourced from the original video on YouTube, published by Y Combinator. The video, its captions and all related intellectual property remain the property of their respective owners; AINotes claims no ownership. Provided for research, accessibility and search — see the Transcript Notice and Copyright Policy.

00:00 Hey, I'm Ali, co-founder of CircleBack. We were in the YC Winter 24 batch and excited to show you my setup. >> [music] >> CircleBack is an AI note-taker that captures and organizes everything from your conversations. It records and transcribes your meetings, writes the notes, and assigns the action items. It also lets you build automations to pull custom info from those meetings and update other apps that you use like your CRM, your issue tracker, your Slack.

00:31 You can also ask it questions across all of your team's conversations, connect your agents to it, and essentially use it as your company brain that remembers everything. >> Take me through your hardware setup. What are the things that you cannot live without? >> So, I try to have a pretty minimal setup both in terms of like tools and software that I use and things on my desk.

00:52 I still am a big fan of mechanical keyboards and I kind of kept going down the rabbit hole of what's the most elite keyboard. And this one's actually not a mechanical keyboard. It's a membrane keyboard. Um, this style of keyboard is called Topre, which I think is for Tokyo Precision. First off, it was really hard to get and when I received it in the mail, all the instruction manuals for it were in Japanese.

01:15 So, there's still a lot of things that I I'm sure I can do with it that I don't know about. Um, but it's it's quite nice to type on because it's got a very nice typing feel, but it's also not super loud. And it has a very like early registration of each key press, so you don't need to like bottom out every key. I also do love like blue and brown switches of mechanical keyboards, but they're a little louder and obnoxious in like an open office, which is what we have.

01:42 >> Can you share more about the little vehicle that is on your desk? >> So, this little guy is um when we're doing our onboarding, we're all kind of like obsessed with how different products do their onboarding. I got served an ad for Temu for this. And I got hypnotized like all the way through the onboarding process. I actually ended up with like six things in my cart with $400 of like cash back with Temu.

02:06 And I was about to actually press checkout cuz they make it so easy. Like there's the right amount of dopamine injection at every level. But I ended up like waking up to like what was going on and removed everything but kept this guy. And now it's something that I like when we're like talking about something I like play with while uh while we're like discussing or if I'm like dictating something and like at the same time.

02:28 Just something to fidget with. Then I also have the AirPods Max. I've had these forever. It's honestly more of like a status not in terms of like signaling to the rest of the team that like you're locked in. Um I don't wear them often but when I do that's kind of what it signals. It's kind of like if you were on MSN or AIM and like you're setting your status as locked in.

02:50 >> Can you share some unexpected ways you are using CircleBack? >> Yeah. So I kind of use it as my ATS. I have this view for meetings and action items where I can kind of keep on top of everything that we need to do. So this view is basically any action item that belongs to someone at CircleBack that is not done and is from an interview. So I can kind of use these to keep track of okay, who are the people that we need to get back to and who is being interviewed by who at the company.

03:19 And then sometimes when I'm really excited about a candidate I will actually go to their person page to get a full timeline of the touch points that they've had with the rest of the team whether it's emails or meetings. And then also make sure that like they have an upcoming interview booked or if they already had their previous interview go and look and see like how did it go?

03:37 Um like and how they performed on it and generally I also use CircleBack to kind of keep track of people and companies that we're talking to. And it So for example with the Y Combinator company page in CircleBack I can see who are all the people that we've talked to at YC. Is there anything that we owe them in terms of like our action items? What are the things that they're doing?

04:01 And like a full timeline of conversations that we've had with anyone from YC. And also like upcoming meetings so I can see this. Shoot, is there we also have our post batch group office hours tomorrow. Uh just a unified view of everything that's going on. One thing that I was doing literally last night in preparation for this interview was I asked CircleBack to prep me for this and it went through all the emails that we've exchanged, past conversations that you and I have had to get a full understanding of like what is

04:28 this going to be and then it gave me this prep that, you know, side sources from my calendar of like where is this going to be, emails that you had sent of like potential things to prepare for and then also went through my past meetings of like cool things to showcase given the understanding that I had of the questions that were shared. So, one thing that I use multiple times a day is just pressing and holding the action button.

04:53 I have it set to start a recording on CircleBack. That's super helpful if I just want to quickly capture something. I don't have to fiddle with my phone. I just press this one button. We're also releasing a watch app which will make it so you just press the complication on your watch and it starts a new recording. That's also super helpful for me if I don't have my phone cuz I'm like walking around the office and it's at my desk or I'm on a walk and I don't want to be holding my phone the whole time.

05:18 That's something that I've been using a lot as we've been beta testing it. >> How much of your day do you spend in OpenClaw Telegram versus coding tools? >> Probably like 10-20% with OpenClaw cuz a lot of things is just always on and it's just working. That's the state ideally I want them to be in. And then with coding tools I would say more of my time.

05:38 I've been spending a lot of my time uh reviewing code recently so not writing as much or playing up as many PRs. >> Can you share more about the different agents that you created in Telegram? >> Yeah, so I have my like main claw agent and then I have one for finance, one for customers, one for people ops. These are just like boring names I came up with cuz I couldn't think of something more creative when I made them.

06:04 We're also in the middle of moving all of these processes that currently live in my open claw to actual Slack channels at the company. So anytime we get a customer email, there's like a new thread on Slack with the investigation and you can kind of do the whole back and forth there. Same with inbound sales. And that's good because it makes it a company effort and also gives everyone visibility of like what are the things that are that our customers are reaching out to us about or what types of customers are reaching

06:34 out for sales and like what do they care about and if they if it doesn't pan out like why is it why is that the case? We have I would say a lot of more traditional businesses that like sit in cubicles and where the only known Microsoft product they use and then we have a lot of growth stage startups and YC companies that also use our product. So it's also one of the reasons we continue to do a weekly run rotation when where an engineer is handling all the incoming support and talks to our customers.

07:00 So this is for example an a fireside chat that we had with a customer in June. I can see quickly like what were the things that we talked about. It's also super helpful cuz when they share feedback while they're at our office, CircleBack captures it. And these were the some things that we were going to look into for them. And then I can actually go back and look at the recording of the meeting that we imported into CircleBack from here and jump to a specific part.

07:23 I can also ask a question here, what was on Chris's wish list for CircleBack? And then it'll comb through everything that we talked about in the 30 minutes and give us a quick rundown and then I can also follow up with making sure like hey, have we done any of these? If so, let's draft an email and send it to him that we shipped these things. >> What percentage of your day do you spend in a terminal versus everything else?

07:47 >> I love writing code and like building the product, so as much as I can, but recently I've been like consciously weaning myself off of it because the bottlenecks of the company we have a lot of great engineers on the team now and like our bottleneck is bringing on more great people to increase our throughput of like product decisions and everything else with the company as well.

08:09 But Thursdays is my no meeting day which I look forward to every week and those are the days where I like just go like headphones on and work on a lot of things that's been on my mind and try to like end-to-end actually ship the thing within that day whatever it is that I'm working on. >> And what are the tools that you're using on a day-to-day basis to write code?

08:28 >> I was like writing a lot of code myself even like through Christmas and then I started using Cloud Code again around that time. I think a lot of people went through this around then. And now I'm at a point where I mostly use like an orchestrating agent orchestrator to fire off a couple of things and then revisit and review them. For some things I still like manually edit code for copy I like to be able to like quickly change it and see how it looks in context on the product or for a small styling things if we're

08:56 playing around with some things it's just quicker to edit it edit the file directly yourself, but for most other things we've built enough scaffolding and infrastructure that AI agents are really good at like doing things end-to-end. And then the bottleneck then becomes code review which we've also done a lot to kind of spawn different types of agents depending on what the PR is and making sure that everything not just from like a code quality standpoint, but also architecture and like product and even like the copy of

09:27 what we're putting in goes through that AI review that makes sure everything is consistent with how we usually do things. >> Walk us through you write and test prompts. >> We have evals in place for things where it's quantifiable or there's a right or wrong answer. Like what things should come up as action items. Like if I say during a meeting, "Hey, I'm going to send you this right now."

09:50 and there's signal that like the thing was done during the meeting, that should not be an action item. It's already done. It would be redundant. It's unnecessary noise. The other things that we really try to do is the writing style actually getting CircleBack notes to seem like they're written by a person. And also we're kind of opinionated in like what good looks like as we should be I think.

10:11 It's a It's one of the primary service areas of the product. But for example, like we really try to never have our notes say the word discussed cuz it's fully pointless. Obviously, things were discussed. It's a meeting. It doesn't add any value. So, that's I think somewhere in our evals as well when we like update models or change prompts to see how is it faring with every incremental change that we make.

10:36 Because if you don't have good systems in place to make incremental improvements, you're constantly playing this game of whack-a-mole of like maybe we made the overview of the notes better, but now the topics are too long or things like that. >> What are your thoughts on token maxing? >> We have a big culture of not just doing the thing like the task at hand, but also doing it in a way that makes it so the next time you or someone else goes to do that same thing, it's easier or the output is better.

11:01 So, in that way a big part of it is using AI extensively for the things that we do, automating a lot of ops tasks that like no one should be doing. Right now, we have an unlimited budget for like using AI. Um and we trust the judgment of everyone on the team. Maybe that doesn't scale when we're like 10x the size, but right now we just care about using the best tools to produce the highest quality output.

11:27 And we also like during our show-and-tells are constantly like showing each other like, "Hey, this is a cool tool that I started using this week and here's how I'm using it." >> What's different about your flow today versus 6 months ago? >> I think I'm not really in the editor much anymore. Like I use a very lightweight I use that right now as my editor just to go in and like review diffs really quickly or make small changes, but most of my work is like directly through agents when writing code.

11:53 I think I'm doing a lot less of the things that I didn't like as part of being a founder which is the ops related work where I was like, "I don't need to be logging into our payroll software to be doing this." So, a lot of those things I've automated just in terms of being able to do more of the things that I'm most high leverage at and really good at.

12:13 I'm doing more of that now than I was 6 months ago because I'm not constantly underwater from all the miscellaneous things that surround running a company. >> What are the things that you do not let agents do? >> Never send an email on its own. A draft is okay. I think copy in the product generally like we want to make sure it's super clear. It can take a first pass, but never something that goes into production.

12:37 And then anything with data access and security. It's good for reviewing and doing like that's one of our sub agents that does a pass, but we basically don't want to let the agent architect the thing, but like the engineer or whoever's working on it should architect it and then have the AI agent actually build it out. >> Can you share more about the thesis around companies needing to record everything moving forward?

13:04 >> I think as LLMs become more capable and as AI agents do more work at companies, I think that's definitely a given. The context of everything that's going on becomes super important because otherwise your AI agent is working in isolation. It doesn't know what was said earlier today or what a customer said. And so, the opportunity cost of not recording is already pretty high and it's only going to become higher and higher.

13:29 And as a result, I think more and more companies are going to default to recording everything and default to sharing the context from conversations that can be shared across the company. And also, it's something that we think a lot about, not just making it easy to capture everything that happens in conversations, but also giving our customers confidence that the things that are captured will be shared with the right people, uh and nothing will be shared inadvertently so that you feel comfortable using the product for

13:58 every single one of your conversations without the worry of like, "Oh, is this thing that I said in this one-on-one going to be shared with the whole company?" >> So, as someone who's building on the frontier, where do you think software engineering is headed? >> I think it's an incredibly exciting time to be building software because the bounds of what's possible is now more so your imagination, not so much the skills that your team or you have, or the bandwidth that you have.

14:23 And I think as the cost of writing software becomes lower and lower, what matters more is not how fast you ship, but what you decide to build and how you decide to build it in order to be able to continue composing the systems that you build together and having them interoperate and work together. I think that what what I'm most excited about that every day is like how the field of software engineering and building software is changing seemingly like every few weeks.

14:53 And I think it's an awesome time to be a builder, especially at a startup because you get to play around with the latest tools and you're not bogged down by uh the red tape.