This is the third interview of the series, where we explore Birch’s transformation into an AI-powered platform. Find the first one, where we focus on building the agent from idea to beta, here, and the second one, about the go-to-market strategy in 2026, here.
There’s one more person essential to this conversation, Milan, Birch’s Head of Product. Milan joined the team when they were already working on the new vision. The first experiments were already done, confidence had grown, but nothing had been released yet. We’ll take it from here to explore the long-term goals and whys behind Birch AI.
Joining mid-transformation: ambition vs. focus
You joined the team a few months ago, when they already had an AI vision. Do you remember your initial reaction to it?
To put it simply: I saw that the team recognized there are many opportunities ahead, but they needed to be paired with focus.
And I liked the direction we had, which was always to build an AI with a performance-marketing mindset. Which I think is the most correct approach to AI that an ad tech partner can have today, and a truly differentiating one. Not ad creation, video or image editing, not a copy generator, not everything all at the same time, compromising on quality.
So for me, when I saw that there is a clear theme, performance marketing, it got me super excited. And then the only thing it needed was an even tighter focus, so that we could actually start delivering.
The agent launched with concrete use cases, so some focus happened. What helped achieve it?
The biggest use case: how can you perpetually be more efficient
I can give you an example of the biggest use case for me, where I was like, «yes, this is the direction we need to double down on», and it’s rule creation. When you think about it, the reason why that was exciting is extremely simple. If you talk to any performance marketer — or marketer in general — they can very easily explain to you the complexity of their campaign: what they’re trying to do, where they’re active, why it’s structured in such a way. All of that complexity is maybe thirty seconds or a minute of conversation. Obviously, not for all of their ad accounts or campaigns, but if they had to describe a single one, it’s fairly easy.
But then, if you have to translate that into a system that helps with performance or helps with navigating that campaign, it becomes really difficult. Because it’s also not really something that is native to many performance marketers: you have to think in Boolean logic used in most of the performance tools, including Birch.
You have to think: okay, what is the logical flow of what might happen, how do I prevent it, and what would that prevention snowball into? And when I saw that the team was already trying to figure out how to get this very messy problem of contextual explanation into a rule, that simply excited me a lot.
To put it succinctly: verbally explaining what you need from a performance tool to guide your campaigns was, I think, the biggest first use case we’re still tackling.
Finding the causality of actions and synthesizing it out of a bunch of spreadsheets.
You can put it that way. I’m also thinking of it in terms of: how can you just perpetually be more efficient? So it doesn’t really matter where that data lives, as long as you can explain it in a way that you can also very easily keep track of: is it helping you be more efficient?
Secret question: LinkedIn contradiction score
I’ve been asking everyone in this series the same question that they didn’t prepare to answer: what’s your favorite AI use case from your personal life, if you had to name one?
Recently, I haven’t been on LinkedIn much, but people keep sending me articles posted on it. So I pair them with the author’s website publication and get a contradiction score. If the score is high, I know I should not read their content. I’ve noticed that a lot of them try to push certain narratives, and they sometimes switch them because of partnerships and promotions. I have this hypothesis that if you do that often, probably whatever you’re writing is not really of value, you’re just selling something.
Can’t they just contradict themselves innocently? LinkedIn is fast, short-form, you might just be thinking out loud.
It’s a matrix-calculated score. I have it from one to ten. If it’s below four, then it’s fine; it’s most likely them sharing opinions and trying to gauge what the community also thinks and wants to read. But if it’s above that, then thank you for sharing it with me; they’re not really an author for me.
Agent vs. MCP: the why, not just the what
Birch MCP just launched. Before we get to what a marketer can do through it, is there a difference between the MCP and the agent?
This depends on the setup marketers are using. Because Birch AI is built with a performance marketer mindset, it is one of the differentiators between it and models like Claude or Gemini. Birch AI is really designed with a sense of: just give it what you want to achieve, what the outcome is, and it will already have enough context for you to achieve it. It will explain the steps, what it’s doing, and why it’s doing it that way.
Whereas, with let’s say, Claude and the MCP plugged in, it doesn’t understand why it’s doing it or what the outcome is. I would say that’s the biggest differentiator.
And obviously, when we’re thinking about functionality, there will always be a slight difference between what Birch agent can do and what Birch MCP can do. Sometimes, we’ll allow something to be done with Birch MCP first, sometimes with the agent first. It usually depends on the amount of context that the specific job needs.
So it’s not about locking people inside the Birch ecosystem, keeping them in the agent, so they never leave for Claude or their favorite model?
Absolutely not. My thought process is that the surface — the interface — doesn’t matter. Our main focus is to make the performance marketing job way easier. Be it faster, more efficient, or enabling workflows that were impossible before. And when we’re deciding whether to do it first in the MCP or in the UI through Birch agent, I don’t really mind whether they do it from one tool or the other. The focus is: where does it actually work better? Or where would we get more learning so we can improve the tool and expand it to other surfaces?
What marketers are already doing with the MCP
They’re analyzing what the rules are doing, they’re analyzing overlaps, and they’re getting suggestions: how do I actually improve the existing system here? Do I add something? Do I remove something? Essentially, how can I make it smarter?
And to be honest, I expect this trend to continue for some time, even as we’re adding write tools or actions an agent can take on its own. Because I think that is the most powerful approach for most existing customers. And obviously, one of the trends I’m hoping to see is that once we add write tools, we will see more and more new customers actually creating systems using the MCP.
Birch MCP vs. Meta MCP: a platform in a box
What’s the difference between Birch MCP and Meta Ads MCP? How do they work with each other?
You can use both at the same time; they’re designed to accomplish complementary workflows. The Meta MCP intrinsically has access to the Meta platform, and I like to explain it this way: imagine it as someone putting the Ads Manager interface into a box you can talk to. And for Birch MCP, it’s a very similar mindset, but it’s not the ads manager in a box, it’s literally your favorite performance platform. You can use it for managing rules, adding rules, and figuring out what’s going on and why.
We’re also extending that to Stage. Stage is one of our new products, focused on bulk uploading ads and creatives. We’re currently working on adding more Stage tools to the MCP. This also ties into the example of some workflows being available in the MCP first: currently, the AI agent doesn’t include Stage, but the MCP will.
To put it simply: the Meta MCP has access to Meta’s native tools. Birch MCP has access to the native Birch tools. They’re mainly focused on completing the workflows you would do in the interface, but from the agent’s perspective.
So Birch MCP plus Meta MCP should be your favorite combo for launching and managing campaigns.
Yes, especially with AI. And I would say that in the very near future, we will also see that it’s not just the combination of Meta MCP and Birch MCP, but also Snapchat MCP, potentially TikTok MCP. We’re really looking into which MCP combinations complement each other the best. Birch MCP will, by default, be multiplatform, simply because it supports working with different platforms — on rules and on ad uploads. But then, this combination of adding different platform MCPs will help us unlock workflows that were simply impossible in Birch. A very simple example, going back to the Ads Manager illustration: instead of creating an entire interface within Birch that replicates the Ads Manager platform, you can now do it through the MCP, pretty straightforward. So there will be certain use cases that we unlock extremely quickly, simply because we’re focusing on AI and MCP.
Building and not building: AI, MCP, and Stage
You already led into my next question: the future of what you’re building, and especially what you’re not building and why.
For this year, at least, and the start of next, we’re concentrating mostly on high-context tasks that are difficult for any one person to accomplish or were simply impossible to accomplish before. One of these would be, for instance, the migration of rules. If you give one person the task of switching all the rules from one workspace to another within Birch, it is possible, but it’s just very context-dependent. You have to do it with rigor, you have to double-check everything by yourself. Or wait for a colleague to, let’s say, come back from vacation to do that for you. So this is one of the things we’re really focused on doing. We’re going all in on the AI workflows and the MCP.
Our same level of focus and determination is on Stage because we see significant potential to help customers distribute their ads across different platforms. And then combining those three solutions to add a layer of intelligence. So it’s not just that with Stage, you launch ads in bulk, you can create systems that support the performance of those ads: use the logic to determine if there’s ad fatigue, to prompt you to do another distribution of ads, to create new rules that will help you maintain the performance you had.
To put it succinctly: a huge focus on AI, MCP, and Stage right now. And we’re obviously always looking to improve the underlying systems that made Birch what it is, which is rules. But those three are our top priority.
The magic there is system prompts
You said Birch AI is built with performance marketing skills and a mindset that knows how to help you better. But as far as I understand, Birch AI is built on Gemini, and it doesn’t have any internal expertise as such, it only has access to your data, accounts, and rules. How does that make it better than just asking Claude?
The magic there is system prompts. We develop our own prompts for Gemini, which serve as the initial baseline instructions the model receives. And obviously, someone could develop, let’s say, skills in Claude or any other LLM to try and essentially mimic that. But I think at that point you’re kind of offloading your own subjective thinking as to what a good system prompt should be or what good skills should be. What we’re able to do is take the learnings from many other advertisers, as well as our own, and create a system prompt that acts not just as a guardrail but as a way to distill the knowledge you need to accomplish things. And also to explain what it’s doing and why.
Build vs. buy: month two or three
This connects to what you and CTO Mike wrote on Slack. I captured those takes for the first interview of the series. Want to elaborate on why developing your own custom solution might not beat Birch AI?
The way I see it: if you have a team that has a budget — meaning they have, let’s say, two or three engineers, one to three performance marketers guiding the decisions the engineers need to take, and the engineers have enough domain knowledge to actually work with models, tweak them, customize solutions — they will most likely develop a niche solution that will work better and be available quicker.
However, the biggest issues would start becoming very noticeable in, let’s say, the second or third month. Simply because the nature of the advertiser’s business is to put one dollar into an ad platform and get five dollars out. That is their main focus. And if they’re now maintaining an AI model that is supposed to help them in their work, their focus is no longer just on how to get the most out of our ad spend. It now also becomes: how do we maintain this model, how do we continue to train it, how do we make sure it’s working well? How do we react — and do we react — when there are API outages, when we hit token limits, essentially all of these other things? Which creates an interesting problem: okay, do you then continue paying your own developers to be essentially dedicated to solving all of these nuances? Do you always have a performance marketer on standby to help them understand what’s not working well and how it should be tweaked? You’re again taking their focus away from what they should be doing — turning one dollar into five — toward something that might help them accomplish that job indirectly.
Versus Birch AI, which might not solve that specific niche use case for which you already informed the team, but solves a wide variety of problems, and completely removes that maintenance headache: the struggles with different APIs, integrating different MCPs, and making sure the entire workflow works across different platforms instead of just one.
So I would say, at the moment at least, it is still very, very difficult to build your own system and almost forget about it, without needing to maintain it. However, that might become a reality in, like, two to three years. Depending on how people set up their agents, their routines, and how they divide and conquer the tasks that come with having this intelligence that helps you advertise. But we’re also going to do the same. So it’s going to be an interesting competition, if we can call it that.
We even had quite a few customers decide not to continue with Birch and instead build their own tools using AI agents. And then quite a few of them actually reached out, which is why I’m so confident about the problems others will experience, since they shared they were having those problems. And that they realized that maybe it is still, today, a bit too early to just give AI agents the job of developing a platform that does everything Birch does.
Conventional things, unconventionally well
We should do conventional things unconventionally well. When I think about what a performance marketer’s job essentially is, it’s not just about following the newest Meta campaign types or ad formats, or the newest AI and machine learning technologies that can help. It’s actually finding ways to do these things in an unconventional way, one that isn’t really tracked or understood that well by the majority, but provides unprecedented results.
One example I remember from earlier in my career: there were a lot of advertisers panicking over the learning phase — what it actually means for Meta, how it will impact their campaigns, what to do. Everyone was trying to tweak their campaigns on the go. And I met one person who just literally increased the budget and the bid for the first few days to an absurd amount, but he had dayparting, which meant the campaign would not actually be able to spend that amount. It kinda contradicts itself, but what it enabled him to do was calculate how much he was actually allowed to spend, and with dayparting, he controlled that amount. But because the signal indicated he could spend more, the algorithm actually pushed more customers to him, both low-value and high-value. He made no other changes, no other tweaks. And somehow, his campaigns exited the learning phase extremely quickly and continued to perform very well.
That is the essence of doing conventional things unconventionally well.
There is no sexy way to do it
This might be a silly question, but how do you plan to actually do things the unconventional way?
So, unfortunately, there is no sexy way to do it or an easy way to do it. The most straightforward way is to be very curious and to work with customers extremely closely. For instance, I found out how that particular advertiser was handling the learning phase by simply following different forums and being very engaged with other advertisers. And when I noticed someone sharing something interesting, I just reached out and asked: hey, can we talk for ten to fifteen minutes?
I think that today this will still be true, with the only difference that if you have too many conversations, you will most likely get the context of them using AI. You might also set up different surveys that will be parsed by AI instead of you going through them. Unfortunately, today you have a lot of shortcuts you can take to get as much information as possible and shorten the decision time you need.
You said «unfortunately»
I do say unfortunately, because I think it is a thing that helps you make a decision quicker, but over the long run, if you lose this connection with customers, where you’re doing it yourself, you also lose the understanding that comes with it. That’s one unfortunate part of being a product manager: you have to make these decisions very fast, get as much data as you can very quickly, and be extremely close with the customer.
If you kind of anonymize the customer, and only have some of their problems contextually explained to you by an AI, you might also forget that you’re actually building products for people.
.jpg){kind=spacer}
