Host: Aakash Gupta
Guest: Jacob Bank, Founder & CEO of Relay App
Date: September 16, 2025
Check out the conversation on Apple, Spotify and YouTube.
Introduction: Why AI Agents Are the Future for Product Managers
Aakash: I just saw the future of AI agents and I’m not even overselling it. I had the founder of Relay app show us new stuff around how to use plain English to do vibe experimentation to build out AI agents that make you more productive and help you save money on employees.
What are the limitations today of AI agents?
Jacob: AI agents are still, in my experience, not ready to take on very complex tasks in a totally autonomous way without a human in the loop.
Aakash: I was just mind blown by how you’ve put 12 agents together to replace an employee. What are some of the other ones in here? What do they look like?
Jacob: My follow-up drafter was something that I wish I had discipline around my whole career, but I always forgot to do. I really think it’s like a critical skill for all of us PMs for the future that we can manipulate and construct our own AI agents to match our particular needs. What I want to show you here is this prompt is ridiculously simple. And I’m sure people will argue with me on this, but in my personal experience, a very simple prompt with one to three good examples at the quality of models that we have today is plenty good.
Our entire webinar process – I come up with a topic, I write it in a Google calendar event, I show up for the meeting, and then my AI agents do all of the other marketing around it. It’s pretty awesome.
Aakash: That’s a lot of marketing channels. People say I’m productive, but you are really productive. That’s insane. What’s the next AI agent people need to build?
Really quickly, I think a crazy stat is that more than 50% of you listening are not subscribed. If you can subscribe on YouTube, follow on Apple or Spotify podcasts, my commitment to you is that we’ll continue to make this content better and better.
Meet Jacob: The One-Man Marketing Army (1:49)
Aakash: Jacob, thanks so much for being on the podcast.
Jacob: Thanks for having me, I’m excited.
Aakash: What are we gonna do today?
Jacob: We’re going to see three real AI agents that I actually use and are super valuable and that every single product manager can and should build for themselves.
Aakash: Amazing. Let’s get right into it.
Live Demo: 12-Agent Executive Assistant (2:07)
Jacob: The first thing I want to show you is my executive assistant. I’m the founder of a small startup. I don’t have a human executive assistant, but I’ve built my own with a combination of the 12 AI agents that I’m showing on screen right now. You can see that I have them divided into calendar management, email management, and task management. I’ve built these up over time as use cases have emerged where I wish I had an executive assistant to help me.
Aakash: Before you even get there though, this is like most people are paying Athena or some company like $2,000 a month or something at least. So we’re saving you at the price of a relay subscription. What’s the regular pro price?
Jacob: Yeah, our cheapest here is $20 per month. I would say all of this executive assistant functionality can fit on the $40 a month plan.
Aakash: Okay. So basically $960 a month of savings if you’re able to tune your agent to your liking. So let’s see if we like Jacob’s.
Jacob: Exactly. So let me show you the first one I created, which is the meeting briefing generator. I have a ton of external meetings as part of being founder CEO. Sometimes it’s with a customer or a prospective partner. Sometimes it’s a podcast like this one. And so I created a meeting briefing generator that does the following for me. Before I have an event upcoming on my primary calendar, I like to get my meeting briefing 30 minutes before.
You can see in this flowchart view all of the things my AI agent is going to do for me in preparation for that meeting. First, it’s going to look at all the guests in the guest list. And then it’s going to do research on each guest individually. It’s going to look at all the past emails I’ve exchanged with that guest. It’s going to look at any meetings we’ve shared and find the notes of any meetings. It’s going to go find their LinkedIn profile and then it’s going to combine that into a dossier on that individual.
And if I can show you actually the prompt for the dossier because I think it’s actually pretty cool. What I’ve told my assistant is if I’m meeting someone for the first time, focus on the public information like the information in their LinkedIn profile because I’ve never met them before. So I want to know the basics of who they are and what they’re up to. If it’s someone I’ve met several times before, I probably don’t need their LinkedIn profile information as much, but I really need the context from our previous meetings and emails.
Then once the AI has done the research on each of these individual guests, it will then send me a notification to my meeting preparation channel in Slack with a combined notification. And just for kicks, I want to show everyone the meeting prep notification that I got for this recording that we’re doing now. That’s always a fun behind the scenes.
So I got this about 30 minutes ago. It says I’m going to have a conversation with Aakash. It’s going to be hosted at this Riverside link. It gives me some more information about you, your background, etc. It includes the context of the emails we’ve exchanged about this. It includes a link to the guide you helpfully sent me to help me do a good job in this recording. And it told me I need to be there in the Riverside 5 to 10 minutes early. And all of this was produced automatically 30 minutes before the meeting by my meeting briefing generator agent.
Aakash: Love it. And this all is getting sent to you in the place where you work in Slack.
Jacob: Exactly. I choose to have it in Slack. Other people – and this is – let me pause here because I think this is actually super important. This is why the skill of building your own agents is so valuable. I might like my meeting briefings in Slack. You might like yours over WhatsApp. Someone else might like theirs over email. I like to get mine 30 minutes before. You might like to have yours at the beginning of the day or four hours before. I like to cover the previous meetings and notes with guests. You may want to just focus on their information in your CRM like HubSpot.
And so in addition to having a pre-built agent like this that you can just plug and play, I really think it’s like a critical skill for all of us PMs for the future that we can manipulate and construct our own AI agents to match our particular needs. This is like this is going to be just as important a skill for the next 10 years of our careers as knowing how to make a spreadsheet was for the last 10 years of our careers.
Aakash: Yeah. Everybody got promoted to manager.
Jacob: Exactly. Yeah. And I think we’re all TLMs. That’s more of an engineering title, but I think it applies to PM too where we still need to have our IC core intact because we all need to be doers and we all need to know what good looks like. We need to be able to write good documents, communicate effectively, do research on our own, but then we also need to be able to manage this team of AI agents that works on our behalf.
I think there’s going to be sort of a squeezing that happens where there’s going to be less need for purely bureaucratic people management because we’re going to have smaller teams. There’s going to be less need for super junior execution work. We’re all going to be these super ICs, super delegators that are in the middle. And that’s cool because that’s like the most fun type of job you can have where you’re close enough to the work to really feel it but also have some leverage in the form of help.
Aakash: So if we look at this, let’s walk through some of the key steps here so that people understand them. First of all, you’re using multiple AI models. I noticed in there using ChatGPT and Claude. Why’d you make that decision?
Jacob: Yeah. So when deciding on which AI model to use, there’s a cost quality trade-off and then there are also certain models that are better and worse at specific things. So you can see that I chose to write the individual dossier in GPT-4o1 which is a generally pretty high-quality model but also very reasonable. And then when I write the final notification that’s going to go to me, back when I built this workflow, Claude 3.5 Sonnet was the state-of-the-art and it’s still good enough. I might change it to Claude 4 Sonnet now.
And now that I have this model picker on screen, I think it’s worth it to linger here because what we’ve tried to do as part of the product is help give people intuition about which models are a good fit for which tasks. For example, when I’m parsing information out of a very large PDF, I’ll typically choose a Gemini model. They have the longest context window and they’re very cost-effective. When I have a writing task, I’ll typically pick Claude 4 or Claude Sonnet 4. And when I have like a spreadsheet analysis task, I’ll typically pick OpenAI’s o1.
Aakash: Very interesting. So it’s really about kind of developing your preference for models for specific tasks. And there are some basic guidelines out there like you just mentioned like those models tend to be the best at those specific things. But that’s also going to change perhaps by the time this episode is published.
Jacob: That’s why I’m very hesitant to give specific model advice because anything I tell you will certainly be wrong 3 to 4 months from now. I think the more important thing is that we all develop a skill set of when a new model comes out, run it through a few test cases or go through a couple of our existing agents. It’s very easy to swap out the model like I just showed you. Do a couple – they don’t have to be super disciplined quantitative evals – just like run the meeting briefing generator with a few different models and see which one performs best because these, as you can see, I haven’t tweaked this workflow in a while but I’m sure now that I’ve switched from 3.5 Sonnet to 4o they’re going to be a little bit better.
Aakash: Amazing. So last question about this is the sequence and LinkedIn icons that we see – what are going on with those tools?
Jacob: Yeah, so this sequence here is a sub-workflow and that sub-workflow is a little helper I’ve created that automatically looks up LinkedIn profiles based on email addresses. And it only works for work email addresses, but the technique is – let’s say you’re searching for my LinkedIn profile and you know my email address is jacob@relay.app. The way you would probably do that is you’d go to Google, you’d search Jacob relay.app LinkedIn profile and my profile will be the first one there. So this little sequence or sub-workflow generates a couple Google search queries that are likely to yield the right LinkedIn profile. It executes those Google search queries programmatically and then it gets the most likely LinkedIn profile URL which then gets passed into a data fetching step that takes that likeliest LinkedIn profile URL and fetches the actual profile data from LinkedIn which will include my title, my experience, my education and everything. So that’s a little set of building blocks that all combine to fetch the personal dossier from LinkedIn on the person.
Aakash: Love it. So if we go back to the executive assistant, because I was just mind blown by how you’ve put 12 agents together to replace an employee. What are some of the other ones in here? What do they look like?
Building Follow-Up Drafter Agent (10:37)
Jacob: Yeah. Another good one that’s also in the meeting space is my follow-up drafter. This was something that I wish I had discipline around my whole career, but I always forgot to do. What I want to do is after every meeting within certainly 24 hours but ideally more like two hours, everyone should get a follow-up email from me and that follow-up email should summarize anything that we talked about, definitely summarize any commitments that I made or they made, and include any notes recording or transcript that’s valuable. And I would say before AI my follow through on this was like I don’t know 25% because you’re busy, you’re in back-to-back meetings like you just don’t think of it. And so I built an AI agent to do it for me and now my follow-through is like 95 or 100%. And I can tell you why it’s not always 100% for good reason.
And so here’s how this AI agent works. And again, it’s super simple. A simple step-by-step workflow with two choices that the AI has to make inside of it. I use Fireflies as my AI meeting note-taker. And so the trigger for this agent to wake up is when a new transcript is created in Fireflies. Now when I first built this workflow, I jumped right to step five. I said once the transcript is created, write a follow-up email and then stick it in my Gmail draft folder. And then I realized I needed a couple of these intermediate steps. I needed a step to figure out if we should follow up at all because there’s a bunch of reasons where a follow-up isn’t actually appropriate. It could have been a no-show. It could have been just like a coffee chat catchup that would be kind of weird to follow up with a recording and transcript. It could have been an internal team meeting where the notes are captured elsewhere.
And so the first thing I added was an AI step that says review this transcript AI and decide if a follow-up is even appropriate in this case. And I gave it some guidance that said in general we should follow up whenever we meet a user, customer or prospect. And we should not follow up if it’s a no-show or an internal team meeting. If no follow-up is required, we’re done. End the run. Stop doing work. No need to do anything further. If there is follow-up required, we actually need to go to Google calendar and look up the event that matches this title because it turns out that from Fireflies, we get the title of the event and the content of the transcript, but we don’t actually get the email addresses of all the attendees. So we need to fetch those separately from the Google calendar event. That’s what this find event step is for. And then once we have the transcript, once we know we should follow up, once we have the attendee details from the event itself, then we can ask AI to write the follow-up draft.
And what I want to show you here is this prompt is ridiculously simple. And I’m sure people will argue with me on this, but like in my personal experience, a very simple prompt with one to three good examples at the quality of models that we have today is plenty good. The longer the prompts I make and the more convoluted they become, somehow the worse my output is anecdotally. And so I wrote a very simple prompt. Given the attached transcript and notes, please write a polite follow-up message and keep it concise. And then I gave it a few examples of previous follow-up notes I’ve written that I thought were good and matched my voice. And then it sticks a draft in my Gmail.
Now I chose for this to be a draft in my Gmail rather than an email that gets sent automatically because I still want to have that final human in the loop opportunity to edit the message before it goes out. There may have been some additional cross talk before the transcript started. There may be some other point that I want to emphasize. And this is a general principle for me that whenever I’m sending an email to a customer or a partner or something that’s a high-stakes communication, I always review the result of an AI step before I actually send it. So I think this is actually really important for people to just realize – let’s not use a note-taker that just automatically sends 80% done summaries to everybody. Add in the human loop step so that you can make that final 20% upgrade as needed.
Aakash: And I’m actually curious, you mentioned the cross talk. Would it be possible for us to go into that and modify that workflow to potentially fetch like did we have a conversation in Slack since then or something like that?
Jacob: Oh, and wait, tell me more. So like you want to add to the follow-up any emails that happened as well?
Aakash: Yeah, something like that. Like, yeah, since the meeting happened, just double check if there were any emails sent between each other.
Jacob: Yeah, exactly. We would use the same exact technique that we used for this find event step. We said, “Hey, look up the event that match this title.” Now let’s say we also wanted to find emails exchanged between these participants. I would add another step. I would type find email. And now I’m searching in my Gmail and I’m looking for any emails that match the criteria. And the criteria I want to look for are – did the from address contain any of the event guest emails. If no, that was easy. If no emails are found, continue. If more than 50 emails are found, just continue with the first 50. I don’t think that’s going to – oh, and I actually should – I need to add one more thing because you mentioned a time bound. You want things that have happened recently, right? Not emails that happened a long time ago. So I want to also say and the send date time comes after the current date minus let’s say like I don’t know 2 hours. I want I only want to check for like super recent emails.
And then I can add those emails as context to my AI step. And I can say something like if there was already a follow-up email from the other person, take that into account in my message because if they’ve already followed up and said thanks so much for the meeting I don’t want to send another hey thanks so much for the meeting without referencing that they already said thanks so much for the meeting if you know what I mean. And so that was the process of modifying that workflow to pull in some additional emails and get a new draft.
Aakash: And the way Relay works, do I need to put in my API keys for Anthropic and OpenAI or are those API calls part of what I’m paying Relay?
Jacob: You have two options. By default, we serve a non-technical audience who doesn’t know what an API key is or what an OpenAI platform account is. So we maintain accounts with all the major models, certainly OpenAI, Anthropic, Gemini, but also kind of purpose specific models like ElevenLabs or Assembly for speech etc. And so the default usage pattern is that you’ll use our AI credits. You pay us, we pay the model provider. If you want to use your own credentials for some reason, like you have a fine-tuned model or you have a particular developer tier level or you have some custom agreement with OpenAI that governs your data use, then you could as an advanced user connect your own API key.
Building a Competitor Pricing Tracker (19:53)
Aakash: Nice. So that’s the EA. I think we got people really good insight into how that works. What’s the next agent that we would want to build?
Jacob: So the next one I want to show everyone is a competitor pricing tracker. And the reason I think this one is really cool is the EA use cases are things that you kind of have to do anyway. And the AI is saving you time. With competitive intelligence, these are things that all PMs, at least in my career, you wish you had time to do but never actually have time to do. And so the AI is providing you like a net new capability that you didn’t have before.
And in this case, we have a competitor pricing tracker because our competitors are changing their prices all the time. They’re changing the amount of included credits or tokens or they’re changing their tier structure or they’re adding free tiers or removing free tiers. And it’s really hard in a hot space to stay on the pulse of that. So here’s how the competitor pricing tracker works. I choose to run this monthly because monthly is the right cadence for me. You could run it if you have a super fast-moving market, you could run it weekly or even daily, but for our market, monthly is sufficient.
And then I have a Google sheet and the Google sheet includes – can you see the Google sheet I’m presenting now?
Aakash: Yep.
Jacob: Yeah. It includes all of our main competitors and a summary of their current pricing. Can you see as I’m scrolling through this?
Aakash: Yeah.
Jacob: Then I’ll go back to the agent. So what the agent does is every month it wakes up, it looks at the row in that sheet and then it goes to that pricing page and scrapes the current information from it and then has an AI step that given the raw text of that pricing page provides a detailed human readable summary like the ones I had in that spreadsheet. So that’s the basics. Go to the website automatically every month and scrape down their pricing.
Then this is where it’s cool. If we haven’t found the pricing information for that competitor in the past, like let’s say they just launched or something or we only started adding them to the tracker now, we’re going to add that pricing summary from scratch. If we have already documented pricing for that particular competitor, we’re going to ask AI, we’re going to tell AI, you can see the prompt here, you will be given two summaries of the pricing of this company. First will be the previous summary from last month and second will be the new summary from this month. Please determine whether they’re the same and if something material has changed like tell me and why.
And so then in addition to updating the spreadsheet with the latest pricing, if they’re materially different, we get an update in Slack. So for example, one of our competitors or alternatives, Gum Loop, recently introduced a lower price individual user plan. And so we got an immediate notification from this in Slack, like, “Hey, Gum Loop introduced this new plan. Here’s the pricing details of it. Here’s how it compares to their previous plans.” And so this is just like one more thing you can take off of your plate and delegate to an AI to do for you.
Aakash: Wow. And for any pricing PM, this would be a gold send. But if you’re not a pricing PM, I imagine you could just change the prompt to whatever element it is. So like for instance, when I was at a firm, I really cared about, well, what are the loan terms that they’re giving off?
Jacob: Exactly. Maybe you want to care about recent customer testimonials or recent product features that were launched. Or maybe instead of scraping their pricing page, you want to scrape their YouTube channel. Or you want to scrape the founders LinkedIn posts. Those are all in the theme of competitive research that you can do. Whatever element of your competitor’s strategy you care about, like you can set up this exact structure of AI agent to keep yourself up to date on it.
Aakash: Nice. And can we see what a final summary looks like for the pricing?
Jacob: Yeah. So let me go back to the sheet and let me look at like Lindy for example. So this is the pricing plan summary. They’ve got a free plan, they’ve got a pro plan, they’ve got a business plan, and they have custom plan pricing. And then they have some additional services. And so these are live updated every month. Let’s see. I know Make just did a big pricing change. Let me see if I can capture that here. They basically changed their pricing structure from operations to credits. Let’s see if we’ve picked that up yet. One sec. Okay. So this one is still on operations because they just rolled out the credit change today. But the next time this agent runs, it’ll tell me Make switched their model from an operations-based model to a credit based model. Here’s what you need to know about it.
Live Build: Reddit Brand Tracker (24:31)
Aakash: Amazing. I think this will be really really useful for all sorts of people. What’s the next AI agent people need to build?
Jacob: Okay, the next one I want to show you we’re going to build from scratch. And that is tracking your brand mentions on Reddit. I don’t know if you’ve talked about Reddit much previously, but I think Reddit is the most misunderstood and underappreciated place on the internet. Everyone has this image of Reddit that it’s just like 14-year-old boys sharing memes. And of course, there are communities on Reddit for whom that’s true. But Reddit is like the definitive source of human information in niche space communities and it makes up like a huge set of the sources in LLM calls. So if you want your product to be positively mentioned by ChatGPT, you need to know how you’re showing up on Reddit.
And so what I want to show you here is how to build a weekly brand tracker that’s going to show all the places where you show up on Reddit, give you a summary of how your brand is appearing, and then it’s going to give you links for each of the individual posts in case you want to upvote a post that’s positive or if you want to react to a post that’s negative. So let me show you how to build that. First, like with the competitor pricing tracker that I showed, we’re going to put this agent on a schedule. It’s going to wake up every week and it’s going to check Reddit for brand mentions.
What’s the third agent people need to build? So the third agent we’re going to build is a Reddit brand tracker. And this is super important because Reddit is one of the main sources of information for all LLMs. So if you want your product to show up favorably in ChatGPT, you better have good organic content talking about you on Reddit. So what we’re going to build here is an AI agent that runs every week, and it’s going to look up all the places our brand is mentioned on Reddit. It’s going to send us a summary report along with the links to all the specific posts and comments where we’re mentioned. So if it’s a positive post or comment, you can upvote it. And if it’s a negative post or comment, you might want to react to it and help correct any misconceptions about your product. So let’s just dive in and build it together quickly. It’ll take about 10 minutes.
So because this is going to run weekly, we’re going to start with what’s called a scheduled trigger. A scheduled trigger is very much like a recurring calendar event. So I’m going to click add trigger and I’m going to click scheduled. And what you’ll see on the right is an interface that looks just like a recurrence picker in a digital calendar. It says when is the first instance and then how frequently does it repeat? And I’m going to change daily to weekly because I want to do this weekly. But if you want to do daily brand checks on Reddit, you know, be my guest. You have this flexibility to run it whenever you want.
Then the next thing we want to do is actually pull the data down from Reddit. So I’m going to add a step. I’m going to type Reddit and I’m going to say get posts by search query. And can you give me a brand that we want to search? Maybe one of the products you’ve worked on in the past.
Aakash: Oh yeah, let’s do that. We were just talking about Affirm.
Jacob: Affirm. Let’s try that. Affirm. Now that might be tricky because affirm is a general word. So should we – can we add like one other word to the search?
Aakash: Fortnite. Maybe Fortnite.
Jacob: Yes. Although it’s going to have so much content.
Aakash: Yeah. Yeah. Let’s – let me actually pick n8n. I like n8n. It’s another alternative to Relay App that your audience should totally check out. It’s another agent builder that’s more optimized for technical users. And this is one that I want to monitor because they’re one of our main competitors. So I want to see what people are talking about. And so in Reddit, you can sort content in your search results. And if you’re a Reddit user, these will look familiar to you by relevance, by how hot something is, new, top, etc. Top is the most viewed overall. And so I’m going to pick top because I want to see the posts that have gotten the most traction. And I’m going to look for all the posts that mention n8n in the last week. If no posts are found, which is not going to be the case for this query, continue without a result. And then I can do a quick test just to make sure some posts come back. And you can see here it found 25 posts. An AI Asian army 12 AI tools that I use this n8n workflow. So I can already tell like this is good content here about n8n.
So now I have the raw clay with which my AI can mold something useful. So now next I’m going to create an AI step that analyzes those Reddit posts and writes a summary for me. So I’m going to press the plus button. I’m going to type AI and I’m going to type write. And in this case, I want to have this report sent to me as an email, for example. And I’m going to say, given a list of Reddit posts from the past week about our brand n8n, please write a detailed report of our brand sentiment. Include any key quotes, top use cases, or other notable posts. This is not the best prompt in the world, but this will be enough to demonstrate what the system can do.
And now I want to just pause here for a moment because I want to show how creating a prompt is different in the context of an agent or a workflow than it is in the context of ChatGPT. There’s two major differences. Number one, in the context of ChatGPT, we’re having a live interactive back and forth conversation. So if I don’t get everything right in the prompt the first time, I have plenty of opportunities to refine and converse. When this is running in the agent or the workflow behind the scenes, it’s going to be one shot. So anything you want included in the prompt, you better include here.
The second is because this is one AI step that has steps that precede it and steps that follow it. You need to think really hard about the inputs and outputs. So now I’ve written a prompt but I need to give this prompt the information it needs – the Reddit posts from the previous step to actually produce the report I’ve asked for. So I click data from previous steps and I select Reddit posts and then I can select the model and like I mentioned I like Claude Sonnet for writing. So I’m going to pick Claude 4 Sonnet and you can see here that this is using our default built-in AI credits rather than my own connection. And I’m going to hit use this model. Now the output here is going to be rich text because I want to have this in a nicely formatted email report. Now let me test this step to see what the AI is going to write. And this will take about 30 seconds so we can talk while it’s going.
Aakash: So when it comes to creating this prompt and getting a really good one-shot prompt, what are the personal things you’re doing to harden it besides testing and iterating?
Jacob: For me, testing and iterating is 90% of the battle. I tend not to agonize too much about prompt frameworks or anything like that. But my general technique which I didn’t fully showcase here is I always give like to give one sentence of context on what role the LLM is playing for me and what its goal is. So I should have included something like you are a competitive analyst or a product marketer or a product manager at blah company trying to do blah. That’s super helpful. Then second I describe the task and any particular guidance or parameters of the task. For example the necessity of including key quotes. And then what I would typically do is I would give one or two examples of output that I thought were really good. That’s my general prompt framework. One sentence of context, however many sentences are required, typically between two and six sentences of what the actual task is and then an example or two.
So let’s see what the AI produced here. n8n brand sentiment report past week analysis. Overall brand sentiment strongly positive. Positive sentiment 88% of posts high engagement. The top post reached 178 votes which is like crazy traction on Reddit. You can see that the main detraction from n8n is that it has heavy setup and a high learning curve which is like the main thing we counter-position against from Relay. We try to be much simpler and users seeking help with implementation. It shows me about the top use cases. The AI powered personal assistant was the most popular use case. And actually seeing this from my AI Reddit listener last week is what inspired me to make that post about my executive assistant that you saw a couple days ago. So it was actually my competitive content researcher that gave me the idea to even package up my executive assistant in that way and post about it. You can see some other information here. Here are some cool notable quotes.
There’s a cost-consciousness theme because n8n is very good value, especially if you self-host. There are some technical developments. Like this is crazy good for writing a two sentence prompt that we just kind of hacked together quickly. And I think this will be really valuable for PMs. So now the last thing we need to do is we set up the trigger, which is going to run every week. We fetched the post from Reddit. We had the AI write the summary. And then the last thing I want to do is send this to myself as an email. Yep. And so I’m going to say subject n8n report. And then in the body of the email, all I have to do is pull the output of step three into the body of step four. Now this workflow is done. I turn it on and every week I’m going to get a report like that in my inbox of the most relevant posts about n8n in the past week. It’s a thing of beauty. And the thing is like if you don’t want it weekly, like you said, you can get it monthly or whatever.
Managing AI Agent Notification Overload (35:38)
Aakash: So I think that there’s two sort of interesting implications with these AI agents. The first is that you’re going to be getting a lot more pings whether those are Slack pings or email pings. And I think most people are going to have to go from a little bit of pure maker day – to use the Paul the Y Combinator founders framing to somewhat of a manager day, right? Because you’re going to be getting all these pings. Yeah. So how do you handle all those pings and all this human in the loop review that ends up coming up because of these agents?
Jacob:ย Yeah. So two points there. One is for these scheduled tasks, I try to keep them on a really organized cadence. So I always have my competitive research land in my inbox on Fridays. I have, excuse me, my content ideas for the next week land in my inbox on Saturdays because I often like to make a post on Sunday. Then I have our support summary from the week land in my inbox on Monday because on Tuesday we have our team meeting where we talk about support. So I’ve kind of for all the ones that run on a scheduled basis, those are actually pretty easy to deal with because I kind of set them on a cadence that matches the rest of my work. For the AI agents that are operating ad hoc – for example, we have agents that every time we publish a YouTube video automatically generates a LinkedIn post – those are natural to do kind of a human in loop in the flow of work because typically I’ve just posted the YouTube video and then I get a ping five minutes later saying now do you want to approve the LinkedIn post as well?
Then the second technique I wanted to mention is I’ve also built an agent or set of agents that then synthesize and aggregate all the other stuff. So one of my examples is like I’m subscribed to many newsletters, including yours. And there’s a lot of content hitting my inbox every day. And so I have a newsletter summarizer agent that every day summarizes all the newsletters that have come in that day. It sends me a digest at 5:00 p.m., which is I know a good time for me at the end of my workday. It includes like the three sentence summary of each with a link to drill in if I want to. So I’ve in addition to my AI agents creating information overload, I also use them to help me manage external information overload.
Current AI Agent Limitations Explained (37:57)
Aakash: Love it. And then there’s been insane hype around agents. We hyped agents a little bit throughout this episode. I want to end on the limitations because there are a lot of limitations. Obviously, you know, your job hasn’t been replaced.
Jacob: No, my job hasn’t been replaced quite yet. Who knows about content writer? Might be a little bit sooner than CEO, but what are the limitations today of AI agents?
So AI agents are still in my experience not ready to take on very complex tasks in a totally autonomous way without a human in the loop. And there’s two components to that that are important. One is there’s a spectrum between a workflow-y experience and an agency experience. Now a workflow – the technical definition of a workflow is that the flowchart of the task is predefined by the workflow creator. You’ll notice that everything I showed you here was technically a workflow. We say step one then step two then AI step three then step four then step five. Whereas in an agent you would just say hey agent your goal is to monitor my brand mentioned on Reddit. Your tools are you can fetch Reddit posts from time to time and write summaries from time to time. Good luck. Go.
I personally have not seen how our customers have success with that level of breadth and autonomy for agents unless they are very good at prompting. So my recommendation to everyone just getting into it is don’t just try to jump right from being a ChatGPT beginner to being like a master agent prompter. Go via AI powered automations and workflows on the journey there.
And then second, for anything that is high stakes, keep a human in the loop. I never put anything on my LinkedIn profile that I haven’t reviewed first. I never send anything to a customer that I haven’t reviewed first. The way I would think about human in the loop, the necessity of human in the loop is think about two axes. One is how good is the AI likely to be at that task? And the second is how high stakes is the task. So if the AI is very good and the stakes are low, do it fully autonomously. That competitor pricing tracker example is a perfect example. AI is very good at it and the stakes are low. Like if I get the competitor’s price in the report wrong one month, it’s not the end of the world. So for that, I had no human in the loop. Fully autonomous.
For the meeting follow-up, the AI is pretty good at it, but not perfect. But the stakes are high. Like if I have a customer meeting or a partner meeting, I really want to make sure that follow-up email is high quality. So I always have a human in the loop for those. So those are the two things I would think about for people who are just getting into agents. Number one, lean towards workflows more than truly agentic experiences to give yourself the control of the flow and the steps. And then number two, for anything high stakes or that you’re not confident AI can do well, leave yourself in there as a human in the loop until you build confidence in the system.
Relay’s $2M ARR with 10-Person Team (40:49)
Aakash: Amazing. For people who haven’t been tracking your journey along the way, how can you share some stats? How big are you guys?
Jacob: Yeah. So we have low thousands of customers. So we’re not, you know, we’re not like a Lovable or a Bolt or a V0, but we have material traction. Our lane in the market is we are trying to serve – be the most intuitive, easiest to use product for less technical audiences. So that means less technical folks at technology companies. And by less technical I mean people who aren’t comfortable writing a ton of code. So that’s marketers, PMs, support people, sales people, HR, operations.
Roughly half of our customers are non-technical roles at technology companies. And then roughly half of our customers are real-world non-technical businesses – real estate lawyers, we have contractors, we have dry ice salesmen, we have professional basketball teams, we have carbonated beverage manufacturers. Like there’s quite a long tail of real world small and medium businesses. So we lean small and medium business. We lean towards a less technical audience. Whereas n8n, for example, really emphasizes it’s the tool for technical people who know their way around a JSON object or an HTTP request. And so if that’s you, you should totally pick n8n. But if you want a simple, easy to use, intuitive way to get started and build the workflows that we showed, that’s what we aim for relay.app to be.
Aakash: Okay. And are you fundraising? How are you building this thing?
Jacob: So we raised a seed round of $8.8 million that was co-led by Coastal Ventures and Andreessen Horowitz. Coastal Ventures was the VC that backed my previous company and then I met Andreessen Horowitz along the way and our current plan is to become profitable and just run a great business. And the reason I say that is I currently don’t feel like we’re capital constrained with anything that we’re trying to do. We have a very small team. We’re 10 people, six engineers, two designers, a product manager, and myself. And even though my background is in product, I’m doing essentially everything but product at the company – marketing, support, sales, HR, finance, operations, etc.
And here’s my claim. We’re going to see way smaller teams in the future than we saw in the past. We’re already seeing this trend right now. And, you know, we’re at thousands of customers now with 10 people. I can totally see us getting to tens of thousands or even hundreds of thousands of customers with the exact same headcount. Maybe one more engineer, maybe one person to coordinate all of our support agents, but I see no reason why we need to be a 30 person team or 100 person team because you know my experience I worked at Google for many years and my experience of working at Google was things are hard not because there’s a lot of absolute work to be done but because there’s a lot of people that need to agree to get anything done.
And so our current plan is to just – and when I say we want to become profitable and not raise future funding that doesn’t mean that we have a lack of ambition. I want to build a product that’s used by millions and millions of people. We just need to do hardcore product work for the next several years with a small really motivated team and then figure out how to speak about it effectively and market it to the world.
Aakash: Because some of your competitors they are fundraising a lot more. How do you think about that? Does that add pressure to it?
Jacob: It’s something I go back and forth on in my mind. I’m always happy when I see a competitor fundraise. It’s like – and I’m happy for two reasons. One is because it’s validation that the space is important and that we’re competing in an area that matters because it means that – so for example n8n just raised a huge round at a 2.3 billion valuation. Lindy has raised a big valuation. Gum Loop recently raised a Series A. Relevance AI recently raised a Series A. Cassidy recently raised a Series A. Wordware raised you know a huge seed round coming out of YC. I see all of that as – the only thing I read from that is there’s evidence that this is a really compelling and exciting space which is the place that I want to be. I don’t want to be competing or I don’t want to be around the fringes where there’s no one else competing. Like I want to compete with the best for a market that matters. So that’s one interpretation and the second interpretation is it also helps me reinforce the lane that makes sense for us.
So we are not going to win today by hiring the biggest enterprise sales force. That’s not our strategy. That’s not what we’re trying to do. We are trying to win by building the most intuitive, simplest, and easy to use product. And curious to hear your experience. In my experience, I’ve never seen a bigger team build a simpler or more intuitive product. I’ve seen bigger teams build more features and more go-to-market motions and spend more money on ads. And so for me, our difference in fundraising strategy is also informed by our difference in product strategy, which is for me, you know, I saw this tweet that really bugged me. It was from a founder who was like, I want to prove I can build a great company without VC money just like people prove that they could climb Everest without having oxygen. I don’t think that’s the right mindset. Like building a startup is hard enough. Like we don’t need to make it any harder just to like show how tough we are. That’s not the reason we don’t want to raise money. It’s not because I want to prove that I can climb Everest without oxygen. I just think for our particular strategy to maintain the most intuitive, easiest to use product, keeping a small product team is like the best way to achieve that goal. It’s not we’re not depriving ourselves of oxygen. We’re giving ourselves the right kind of oxygen if that makes sense.
Aakash: And you mentioned it is such a red-hot space. It is on the fringes for people who haven’t really grooved it. And actually I polled my community two weeks ago now and I asked them how many of you have built an AI agent you know 2% of them had.
Jacob: Yeah. So even amongst people are reading – it’s so early but let me make my best bull case for this market. I’ve worked on productivity tools my whole career. Productivity tools my whole career were a place where you go to click buttons to do your day job. You go to Gmail you click some buttons and you send some emails. You go to Salesforce and you click some buttons and you update some opportunities. You go to Google Docs and you type some buttons and then a PRD comes out. A productivity tool was always just like an interface in which you click buttons to do things. And yes, AI powered productivity tools are really cool. ChatGPT is a place where you go and click buttons and oh my god, it’s AI powered. So the results that come back are like pretty phenomenal and amazing and co-pilot experiences whether it’s Lovable or the cursor autocomplete or Grammarly are amazing because you’re in this canvas of your IDE and now you have this coding assistant that you write faster but fundamentally you are still going to a place and clicking those buttons.
An AI agent is working for you automatically behind the scenes. This is the first time we’ve achieved a technology that is like actually equivalent to hiring someone. Like I am our entire marketing team. And you see me on LinkedIn, you see me running webinars, you see me sending emails to our newsletter, you see me doing customer support. Like there’s no way I could do all of this if I didn’t have my AI team behind me. We would probably need a four or five or maybe even 10 person full-time team to do this.
And so I don’t think that has quite clicked in people’s mind how big the opportunity is. Like we are finally taking the leap from being a nicer environment to go click buttons to something that actually does work on your behalf. And it’s hard like it’s a mental leap to figure out like what is the job description for this AI agent? How do I give the AI agent the training that it needs? How do I give it the instructions that I need?
And I think there’s a lot of people that are professionally over complicating building AI agents. And so one of my goals in showing you the Reddit listener that we built in what 5 minutes is like it’s not as hard as people think. Like if you have some basic skills with ChatGPT and you’ve built some basic workflows in Zapier at some point in your career, you have all the tools you need to build your first one two 5 10 100. I have about 250 AI agents working on my behalf right now. And so that’s why I think like we’re still very much in the visionary phase of the market from a customer adoption perspective, but like as soon as we hit the early majority, it’s going to start to explode. My hope and prediction.
Why PMs Are Losing to Sales/Marketing Teams (49:14)
Aakash: What are the pockets? Who are the people? What are the industries? You know, you mentioned things like real estate, lawyers. What are those places that they are actually uptaking AI agents faster? I was surprised PM was so low.
Jacob: I think the fastest by far is in go-to-market functions. That’s certainly customer support and customer service. I mean look at Intercom Finn and Decagon and Sierra like that is already a red-hot market with you know many millions of ARR going around. So I think support teams are very ahead of the game because they already have a mental model for how a human gets worked on within a support team and what the metrics are and those are places where an AI can often provide faster higher quality service at a lower cost.
And then I think sales teams and marketing teams are also adopting AI tools way faster than PMs are. Sales teams especially whether it’s for prospecting or lead qualification or lead routing or in marketing competitive research content research content generation I think from what we see in our data like those three functions are the furthest ahead and I hope this lights a fire under your audience because I know PMs like we always think of ourselves as the innovators and the technology people and every company I’ve always worked at there’s been this sort of like friendly I don’t know about rivalry or like you know joshing each other between the go-to-market side of the business and the product engineering side of the business and the product engineering side of the business is always like we’re the technologists we’re the visionaries we’re so far ahead and like you salespeople you don’t get it and I think in when it comes to adoption of AI tools I think it’s the opposite right now like I think the customer-facing teams are leading the charge and PMs need to get our act together.
Now part of that is because AI agents like the ones I showed you are best fit for like frequently recurring tasks that share a very common pattern. And I think those are more common in sales marketing support than they are in PM. But I think that also is just a lack of imagination that we just haven’t tried hard enough to think about the elements of the PM job. And the way I always think about it is like we’re all going to use AI in three modalities. We’re all going to use chat bots. ChatGPT, Claude, Gemini, Grok, pick your favorite. We’re all going to use co-pilots. Every single one of us should be living using Lovable. Every single one of us should be using Descript for video editing or you know pick your equivalent in category and we’re all going to be using agents and the challenge for all of us is to figure out like what is the right tool for the job. So if I’m writing a PRD I don’t want an agent to write a PRD for me. It’s not that frequent of a use case. I don’t need that level of autonomy. I want to use something more like a Claude Project or a custom GPT and I want to have an interactive experience. On the other hand, if I want to get a weekly report of all the open P1 bugs affecting a certain class of VIP customers, like have an agent do that. Like, there’s no reason for you to be manually doing that same SQL query and writing that same email to your leadership every single week. Like, make an agent do that. You review it.
Aakash: Yeah. There were so many jobs I had early on in my PM career which were literally like take emailed scheduled SQL query, put it into Google sheet, update PowerPoint slide.
Jacob: I mean every single PM we all have 10 of these, right? Like whether it’s what have competitors launched recently, what pricing changes have happened recently? What are all the features we’ve launched recently? Put it in our release notes. What are all the bugs that we fixed recently? What are all the bugs that have come in? Who are all the new customers this week? Who are all the upgrades this week? Who are all the churn this week? Anytime you find yourself writing “weekly” in the in like an email subject like that should be a mental like that should be an immediate cue like I should have an agent writing this for me.
Aakash: Love it. So it’s all of the recurring tasks. So I guess if we had to summarize like the top I don’t know 10 for PMs. I’m just thinking off the top of my head and then let help me finesse this. Right. There’s probably like we just said like your executive assistant right? So that’s your meeting. And that’s a whole bucket of managing your meetings, managing your calendar, managing your tasks, finding time for your tasks, notifying you about tasks that have done, managing your emails, summarizing your newsletters, filtering out cold spam email. Like I would say, you know, I have my executive assistant of 12. I think that’s about right because I don’t think anything I do in email, calendar or task is unique to me. Like those are things that we pretty much all do in our work. So that’s one executive assistant is one big bucket. Then competitive understanding is another big bucket where like you said you might want to be tracking competitor interest rates or competitor pricing or competitor feature launches or you know because content marketing is so important for us whenever a competitor has a LinkedIn post or a Twitter post or a Reddit post or a YouTube video that hits I want to know about that. So there’s a whole cluster around competitive tracking then there’s a whole cluster around support and user feedback. So, for example, synthesis from user interviews, synthesis from user feedback in product, synthesis from support tickets, synthesis from organic brand mentions on Reddit like I showed you, like customer understanding is another big bucket. And I think the third big bucket is project management and the cadence of the PM job. Metrics reports, monthly product newsletter, weekly, you know, release notes for the iPhone app, quarterly update for VPs or other stakeholders, the sort of, I guess you could call it like product rhythm of business, rhythm of business type use cases. Those are the four big buckets I see PMs benefiting most from with agents. And within each bucket, you could have anywhere from one to 20 agents within each bucket, depending on how deep down the rabbit hole you want to go.
Choosing Between Agent Platforms (54:53)
Aakash: So PMs often they don’t have the ability to choose which tool. So let’s address this specifically for the product leader. If I’m the product leader, I’m figuring out should I get my 20 PMs a Relay license or a Lindy license or a Make license or n8n license? What’s the difference between these tools?
Jacob: So first I think you should offer people all of the above because the way these tools charge is it’s not per seat. It’s based on usage. And so the base subscription for any of these plans is super cheap, like $20 to $100 a month for a team plan where like anyone can use it, have a login, it only gets expensive when you start using it a lot, but that’s when it starts accruing a lot of value. So if I were a PM manager, I would not say we all use Zapier or we all use Relay or we all use n8n. I would say like, hey, swipe your credit card, pay 20 bucks for whichever one you want, and then yeah, if there becomes like a winner emerges, then you can convert it to a team plan. But this is not the kind of case where like you have to buy one $50k a year subscription and then get the whole team on that. So my quick answer is like buy all of them, let people try all of them and experiment.
And then when you’re looking at the space, I think there’s the main axis of decision-making I would say is how technical is the team that is going to be using this tool. If you have a more technical team, I think you want to lean towards the n8ns, Makes and Gum Loops of the world. If you have a less technical team, you want to lean towards the Zapiers, Relays, and Lindies of the world. That’s like the first highest order split. Then let me take our less technical world of Zapier Lindy Relay and I’ll I’m going to try to be really even-handed for everyone. Zapier, it’s the no one got fired for buying IBM product in this space. And I don’t mean that in a bad way. I mean, they have the most integrations. They’ve been around the long time. They’re executing really well. They have a strong team, but the product’s 14 years old. And so, you’re just kind of like limited in how much UI flexibility you can have to innovate on top of like their core mechanics. I don’t I mean that back when IBM maybe 20 years ago, like when that was still like a really good default choice.
Lindy and Relay I think are both like more intuitive, nicer UI, more modern versions. Lindy leans a little bit more towards agentic magic and Relay leans a little bit more towards define each of the steps of your workflow. Of course, you can also make workflows in Lindy and you can also make full scale agents in Relay, but that’s the main difference. And so when you’re thinking about your the ca the technical capability of your audience and the tools you need to integrate with that’s the way I would kind of make the choice and like but again the highest order bit is like let people choose their own tools let people experiment with all of them and then see which winner emerges.
Aakash: And if you had to make the pitch Relay versus Lindy what are the things you’re getting in Relay?
Jacob: So, Relay has fewer features than Lindy. We don’t have a built-in meeting recorder. We don’t have voice agents. We don’t have browser based actions. Relay has a more limited product scope. But if you want to build a basic agent that triggers based on your email or sends you meeting briefings or track your competitors or does the Reddit listener, we’ve built, I think, the simplest and most intuitive experience to get you started. It’s the easiest to create, easiest to test, and easiest to run. And so, so again, if you’re beginning your AI agent journey and you’re looking for simplicity, ease of use, and intuitiveness, like that’s where we see people succeed with Relay.
Aakash: How long does it take to build like an executive assistant like that 12 agents we saw there?
Jacob: Yeah. So, for me, if we were doing like a timed challenge and I had to build those 12 from scratch, I could probably do all 12 in an hour. But that’s not really a fair test because I’ve already built them all myself and I’m an expert in the tool because I use it all day every day. I think the better way to think about it is it’s going to depend on how experienced you are in ChatGPT and how experienced you are with workflow tools like Zapier. Let’s say you have a moderate level of experience with ChatGPT but you have very little experience with a workflow tool. I would say building a high quality meeting follow-up agent would take you about an hour and then subsequent agents will get faster.
But the key point here I think is I don’t think you should think about how long would it take to build one agent as sort of like a fixed quantity. I would really think about how can I invest in agent building as a skill so that I can get really good at building these things quickly. So for example, like if I wanted to build a quick meeting briefing generator for myself or competitive researcher for myself right now, I’ve now built up the skills that I could do that in five minutes and like blow the mind of my peers and my managers and I want everyone listening to to build up the skills that enables them to get to that point.
Aakash: Yeah, it’s like a skill set like anything. I think you need to put in I would say probably I’ve probably put in like 30 40 hours now across these various tools. So now I can build an agent in like 10 to 15 minutes and I feel pretty happy about it. So it it doesn’t take as much time as you think.
Jacob: Yeah. We’re not talking like a 10,000 hours like being a performance violinist. It’s way easier than that. Like if you spend like I I run live build with me sessions every week. If you come to like three of my build with me sessions, you’re going to be top 10% at using AI agents like among PMs already.
55-Agent Marketing Team Breakdown (1:00:37)
Aakash: So you mentioned that you are Relay’s marketing team. That’s actually mind-blowing. So you need to break down for us. What are the agents that are helping you? How are you a one-man marketing team for this 10%?
Jacob: Okay, I got to show you this because this was also my most viral LinkedIn post ever. It got 33,000 comments where I showed an org chart of my AI agent marketing team. Basically it’s a team of about 55 agents now that’s organized by marketing channel. So our marketing strategy is an organic content community education-based strategy. So, I have a set of agents that help me on LinkedIn. A set of agents that help me on X, a set of agents that help me on Reddit, a set of agents that help me on YouTube, a set of agents that help me with our email marketing and our newsletter, a set of agents that help me with our webinars, a set of agents that help me with competitive research, a set of agents that help me with our partner program and our affiliate program, and a set of agents that help me with our cohort-based learning. Those are sort of our main marketing channels. And in each of those buckets, there’s anywhere between three and 10 agents.
So, let me talk about webinars specifically because I’ll tell you the process of a webinar at a typical company and then I’ll tell you the process of a webinar at Relay app. In a typical company, there’s like a product marketer and then an email newsletter marketer and there’s like a webinar marketer and then there’s the PM and then there’s this whole process of setting up the calendar many months in advance and then a month before you do a pre-meeting and a pre-recording and all that. Here’s the process that I use with my AI agents.
I create an event on our relay.app events Google calendar and I put in the title of the event which is the topic of the webinar like Trigger Triggers 101 is the one I did yesterday. Then here’s what happens. One of my AI agents automatically adds that to our event signup page which was vibe-coded with Lovable so that people can sign up for that event. Another agent automatically writes a LinkedIn post for my personal account 4 days before with a poster that’s again AI created announcing the webinar. Another one of my agents writes an email to our newsletter saying, “Here’s this upcoming session. Here’s the topic.” When someone signs up on that Lovable vibe-coded signup page, another agent automatically adds them to the calendar event, which is like a standard Google calendar event with Google Meet. It automatically sends them a confirmation email that they’ve been enrolled. It automatically sends them a reminder 3 hours before the meeting. It automatically sends them the recording after the meeting. We have another agent that automatically tracks all the signups and publishes them into Slack. so we can see the LinkedIn profiles of everyone who’s attending. We have another agent that tracks the attendance afterwards and lets us know how many people are showing up to know how many people from our team should come to ask questions. We have another agent that automatically turns the transcript into a YouTube description and then a follow-up social post of the recording. So, those are like the eight or nine agents. And so our entire webinar process like I come up with a topic, I write it in a Google calendar event, I show up for the meeting, and then my AI agents do all of the other marketing around it. It’s pretty awesome.
Aakash: Whoa, that’s a lot of marketing channels. People say I’m productive, but you are really productive.
Jacob: Yeah. And it took a while to set up this whole system, and it didn’t happen overnight, but I was like, “Oh, it would be cool if we automatically promoted these to our newsletter.” Build that agent. Oh, I should also promote them on LinkedIn. Build that agent. Writing these YouTube descriptions is super annoying because I have to like re listen to the whole thing and make the chapters. Like, let me just transcribe the audio with Assembly AI and then automatically write the chapters and description. So, I built up this motion over time. But now, it’s truly a joy to do webinars because all I have to do is put an event on the calendar, show up, and like the rest of the value just happens. And that’s why I was saying, you know, earlier that I my mind has been blown by AI agents already. And yes, there are places where they’re overhyped and there’s places where they overpromise, but like already I feel like I’m living in the future relative to the old world of like going to Google Docs and typing up a script. And I’m just I’m excited for everyone to get there because it’s so much fun.
Aakash: It is so much fun. That’s insane. So, I want to shift because we’ve been talking a lot about how PMs use AI agents for productivity, but you also have a lot of expertise to help people with on building agents into their products. And I would assert that if you want to build your product for the future that we’re moving into, you need to be thinking about how to build AI agents into your products.
Jacob: Yep.
Aakash: Do you agree?
Jacob: Yes. I think and so if that’s true Oh, go ahead.
Aakash: No, no, no. I think it’s true and we’ll talk about why.
Building AI Into Your Product Strategy (1:06:43)
Jacob: Why? Why is it true? So the nature of everyone’s work is going to change and our expectations of product our expectations of productivity tools are going to go way up. A place where you go to just click boxes will no longer be a sufficient product in the market. You will be expected to have some sort of chatbot experience, some sort of co-pilot experience or some sort of agent experience. Now, I don’t think maybe I should caveat this and say I don’t think every product necessarily needs an agent experience in the product. I think every product needs a first-class AI experience that makes it dramatically easier to complete the user journeys within that product. So, for example, Figma may not be as well suited to an AI agent, but it’s very well suited to a co-pilot that automatically fixes the spacing on your mocks or automatically adopts the right color palette or automatically creates a nice prototype for it or automatically produces the right component. Like, Figma doesn’t necessarily need to have an agent that goes and works for you automatically in the background, but they definitely need some sort of AI co-pilot and they may also need some sort of AI chatbot.
So the way I would think about it if I were a product builder is if your product already has a very significant natural language interface for example it’s a search research or discovery product like almost certainly you should be trying to build a ChatGPT style chatbot to go from like a static search box with results to an interactive text-based sparring partner. If you have a product upon which there is a canvas of creation, a document is being created, a mock is being created, some code is being created, you better have a really good co-pilot and that co-pilot better have really good like in canvas autocomplete and it probably should also have some side panel natural language co-pilot experience. And then if you are a product that does workflows or automations, for example, like email marketing sequences or project management automations of marking things as done, like you better be adding some more AI and agentic capabilities into those. So, so maybe I was wrong to say that every product should necessarily have agents. Every product should be rethought through the lens of AI, but chatbot, co-pilot or agent could be the right modality depending on the product and the interaction pattern.
Aakash: Really helpful nuance there. And I think one other element of building an AI agent that people need to understand is you don’t need to just build an agent into your product. Actually, I think that maybe can you help explain MCP to folks because I think instead of building integrations to other platforms, what if you became the platform that AI agents integrate with?
Jacob: Yeah. So, let me touch on MCP briefly. I mentioned earlier that the definition of an AI agent is an AI system that is given a goal and a set of tools and then figures out how to combine those tools to achieve its goal. What is an example of a tool? An example of a tool could be asking a question of ChatGPT. An example of a tool could be running a web search. An example of a tool could be sending an email. An example of a tool could be looking up a record in Salesforce. And so the question became quickly is like I want to enable my AI agents with tools that touch lots of my SaaS apps. And there’s basically two ways to do that. One way to do that is what we do in Relay app primarily is you have API based integrations where every primary action that’s exposed in like for example the Gmail API you can encode as a workflow action in your flowchart. MCP is a slightly different technological approach of achieving the same thing of exposing the actions that can be taken in a product as tools for an AI agent. So if you connect to for example a monday.com MCP server, you could then write a prompt to your agent that’s like look up all the tasks that were involved in this project on monday.com and update all their status to done. And then the AI agent would have a tool that can then make these MCP calls to Monday for it to take action on its behalf. So that’s the promise of MCP is that you can go to ChatGPT or you can go to Claude or you can go to any agent building tool like Relay app and you can like magically give prompt based instructions to have real world actions taken in all of your tools.
In practice I have not seen many people succeed with MCP in real life workflows. I’m curious maybe you’ve seen something different. I think the promise is very real and very exciting. It is still something that only I think the technical visionaries have turned into like day-to-day value. Like I’m curious what percentage of your audience would you say has used?
Aakash: Very low. Again, that’s where like that’s why I want to build this content is to increase the percentage. I just feel like more people should become tools that these AI agent workflows are calling. You know, totally like so many people these SaaS apps, they’re often the way they end up building lock in is they become like a data source of some sort, you know, like an AI company out there.
Jacob: Yeah. Exactly. And so for all you all the product builders out there who have who are working on a system of record like a CRM or a work management tool, any kind of system of record or data store, you probably think a lot about your user interface for manipulating that data store and occasionally about your API. I would encourage you to think very hard about your API and very hard about your MCP server because in two years I bet many many more actions will happen on your data via AI agents and workflows than via humans coming into your UI. And so I expect we’re going to see APIs get way better and MCP servers get way better. And we’re very quickly going to see a usage pattern shift from primarily humans clicking buttons to primarily agents making API or MCP calls.
Aakash: 100%. The last thing I want to cover is a little bit about you because you have a fascinating story.
Why Jacob Left Google Director Role (1:11:47)
Aakash: First of all, you left your job as a director of product management at Google. That has some pretty heavy golden handcuffs associated with it. Why did you leave your job as a director of PM at Google?
Jacob: I left my job at Google for two reasons. So at the time I was the product lead for Gmail, for Google Calendar, and for several of other productivity and collaboration tools under the Google Workspace umbrella. I left for two reasons. One is I felt like I wanted to personally learn and grow in a different direction. I’d been at Google for six years. I was managing a team of teams of PMs. I knew how to do performance reviews. I knew how to recruit. I knew how to manage people out. I knew how to run OKR reviews. I knew how to do product strategy presentations. Like I felt like I knew that version of the job. And while it would have been I could kind of see what my life would be like 5 years from now. And I’m like, “Oh, shoot. It’s going to be the same job. I’m going to have like the same fundamental skills.” So, one big reason was I wanted to learn and grow. And I knew from my previous experience of starting a company that nothing compares to being a founder.
And then second was I knew I wanted to build something that was a cross product AI powered workflow tool and I did not feel like that vision could be successfully built at Google or at any other incumbent because at any incumbent you always look through the lens of like oh well everyone’s using our button clicking interfaces Gmail and calendar and docs and drive like let’s make them click buttons in our product better. I remember we’d have customer meetings where people would say like we’d say like, “Hey, look at this awesome new reply button we built in Gmail. Look, it does three AI replies. Isn’t that so cool?” And customer would be like, “Yeah, yeah, whatever.” Like, “Your reply button is fine. My real problem is I have this data in Salesforce and this data in HubSpot and this data in Zendesk and I want to connect it all.” We’d be like, “Oo, sorry that sounds hard. Good luck. Use the API.”
And so, for those were the two motivations for me. One is I wanted to learn and grow and only being a founder can really push you to learn and grow to the max. And then second this cross product AI powered workflow opportunity I think had to be tackled by a new startup rather than by any individual market leader in the productivity space.
Aakash: So you had that calling for this specific solution which is interesting. I think a lot of people just have that especially PMs especially PMs who have reached kind of the level you reached which was like very senior in your career to have that founder calling. And these days, I talked to a lot of PMs. I’d say probably like more than 40% of the PMs I talk to these days, they’re like they’re thinking, should I start a lean AI company like Jacob? Should I start a 10-person AI company? I could potentially have a huge outcome. What’s your recommendation for PMs who are considering becoming an AI founder?
Brutal Truth About PM-to-Founder Transition (1:14:43)
Jacob: Yes, do it. No. That’s the facetious answer. The physician facetious answer is like yes, obviously do it. There’s more nuance to it. You will learn and grow faster than in any previous role you’ve had in your career. You will build more skills faster than you have in every previous any previous role in your career. You will do your best work in a way that you haven’t been able to do anywhere. What will come with that is an intense feeling of being lost and confused and demoralized at least 75% of the time. And so that’s basically the trade-off you need to go into. Like you’re going to have this amazing experience and you’re going to come out really strong on the other side, but don’t let anyone convince you like it’s all glamour and fun the whole time.
I remember like I’ll give you a concrete example of that. Like when I was running Gmail, I could basically get a meeting with anyone I wanted for whatever reason because everyone wants to appear or integrate with Gmail or have their emails delivered in Gmail. It’s like it’s a position of authority in the technology world to be like the product lead for a top 10 app. The very next day when I was startup founder of unknown product like I could not beg a line marketer at a startup to spend half an hour giving me feedback on something like this like you you just have to be ready to like completely lose any ego that you have. You are going back to the beginning like no like no one cares about the title of your previous job. No one cares about how many users your previous product has. Like you are back at the zeroth rung. And so that’s like a humbling and fun experience.
The other big thing here here’s my other kind of hot take. I think being a PM at a large software company is the worst possible preparation for being a founder. I don’t think there’s any job that is worse preparation for being a founder. I think running a hot dog stand is better preparation for being a founder. I think, you know, being a dog walker is better preparation for being a founder. Being a contractor is definitely better preparation for being a founder. The reason being a PM at a large company is so dangerous is like you think it’s the same because you’re like, “Oh, I’m the CEO of the product.” That’s what the PM job description says. It says, “I’m the CEO of the product.” And you think you have these similar experiences of crafting product and doing user interviews, but it is completely different to manage the inertia of an already successful product than it is to create something from nothing.
And so that’s the other I hope people I hope people see that I’m saying this in a way like an uplifting way where it’s like an opportunity to learn and grow in new ways. But I think that it’s like it’s such a reality check for people where they think like, “Oh, I trained in this great PM culture at Facebook or at Google or at Amazon or at Stripe or like pick your favorite. Like I trained at this great PM factory. Surely I am now prepared to be an amazing founder.” Like immediately unlearn everything you ever thought you knew about product management or your career in general. And like go talk to a bunch of local small business owners and you’ll learn a lot more that way about like how do you get your first few customers? How do you set up basic marketing? How do you set up payroll? Like I’m not saying set up payroll is the hard part of starting a company. The hard part about starting a company is like there is literally no inertia. You have nothing. You have no product, no customers, no messaging, no marketing, no positioning. No one cares at all about you. Everyone already has way too many tools or everyone already has way too many sales people banging down their door. Like how to turn nothing into something is just so different. When like I can pat myself on the back for doing good job on Gmail all I want. When I inherited Gmail, we already had 800 million users. And yeah, we grew from 800 million to a couple billion. And I feel good about that. But like we already had the 800 million. So it is a totally different experience to go from like your zero with customer to your first customer.
Aakash: Yes, that’s a bold take, but I think a reality a lot of big tech PMs should hear so that they go do talk to those small business owners early on when they’re starting because so many of them do. We could talk for another 90 minutes. Perhaps we’ll have to in the future. Jacob, thank you so much for doing the podcast.
Jacob: Thanks for having me. This was super fun.
Aakash: All right, everybody. We’ll see you in the next one.
