Listen to this episode on Spotify or Apple Podcasts
He says “check it out” the way you say it to a friend, not a customer.
We’re about half an hour in when Artem Koren tells me the SOC 2 story. His company, Sembly AI, is an AI meeting intelligence platform — Sembly, S-E-M-B-L-Y, not to be confused with AssemblyAI, which is a separate company that does speech-to-text infrastructure. He’s careful about the distinction. “We’re actually related, we partner on a few things, but completely independent. Evolutions of name just so happens, so it’s a little confusing.”
That out of the way, the story:
“I did something where, you know, we are a SOC 2 type two compliant organization. And so our auditor was on the call with us and I said, check it out. I’ll show you something funny.” The pitch is half its appeal. “I basically took all of our auditor sessions that we also had on Sembly, because Sembly attends all of our auditor sessions. And then based on all of our auditor sessions, I asked it to compile the SOC 2 type two report. Like what our auditor takes like two, three weeks to do by hand — I just asked Sembly to do it based on those sessions and they did it in a few minutes.”
A few minutes.
I want to react to that the way I’d react to seeing a magic trick — with the slight unease of not being sure how it was done. Then I remember: Artem and his co-founder, Gil Makleff, started building Sembly in 2019. Pre-ChatGPT. Pre everything. They didn’t fall sideways into AI when the wave broke. They were already underwater.
I realize the layout — the part most software founders won’t tell you — only after the conversation turns to what it cost.
“We started Sembly AI before AI was cool,” he says. “So 2019, not too many people were talking about AI. It was a very different AI landscape at that point. It was machine learning and data science.”
I’d somehow forgotten what “machine learning and data science” meant in 2019. Not a wrapper around an OpenAI API. Not a fine-tune. Custom models for everything. He runs through it without theatrics: “We had to build a bunch of custom models for detection of different items across meetings, for being able to interrogate meetings in natural language. So all of that we built before GPT was a thing. We also actually had to build our own transcription engine because at the time, the transcription engines, even from behemoths like Google and Amazon and Microsoft were just not up to par.”
Behemoths. His word, not mine — and the framing carries: the giants existed, they just weren’t good enough.
“They were just not as good as they needed to be for the quality of results that we needed to achieve. So we had to build that. And we got a very sharp engine that was rivaling the best transcription offerings in the market, but only for the English language at that time.”
Then a smaller line, almost an afterthought, that I can’t shake later:
“We also had to build infrastructure and the whole platform to attend meetings.”
He means: Sembly had to invent a fake meeting participant. Software that joins your call the way a person joins your call. “We have our own proprietary platform that basically simulates a user who will join a meeting with you. That user simulated for all the meetings that Sembly is present on.”
This is the part that matters. The thing under the surface. Sembly’s whole 2019-to-2022 buildout was a tunnel — a long, expensive, mostly invisible tunnel — to the same place that, by late 2022, you could parachute into for free.
“Then the end of 2022, early 2023 revolution happened with GPT and we had to step back and rethink the ways in which we approach some of our architecture and technologies and kind of revamp Sembly to take advantage of all the new AI stuff that’s coming up.”
That’s the mild way of describing a moat becoming a floor.
I ask the version of the question I came to ask. “What was the hard question you had to ask yourselves to reposition the company? You mentioned that you guys had to step back. Like what changed?”
Artem doesn’t get cute about it. He doesn’t say “we had a dark night of the soul” or “we re-platformed.” He gives the answer that I think tells you most about him:
“With the introduction of modern LLMs, what it did was it kind of raised the base watermark for what’s easy or what’s very achievable to do in technology. So pre-LLM, the kinds of features we had on our platform, basically no one had and would be very challenged to put together. There was really a lot of hard innovation baked in.”
The watermark rose.
I want to think about that line for a second. There’s a default founder pose around moments like this — wounded, defiant, prove-the-doubters-wrong. Artem doesn’t do it. He frames it as a tide condition, not a war. The water came up. It came up for everyone.
“With LLMs coming to the fore, suddenly you can create a basic version of what we’re doing fairly easily. And so there was a proliferation of these AI note taking apps of all different kinds of flavors.”
He runs through what they don’t do — privacy, SOC 2 type 2, geography, accuracy, mid-track language switching, speaker disambiguation — almost as muscle memory. Not bragging. More like a checklist someone goes through when they’ve been the only one in the room who has to think about these things.
Then the actual answer.
“Note taking in and of itself was not a big challenge. And so we had to step back and say, okay, note taking was the low hanging fruit for post-meeting value. What would be kind of the next leap? And we thought that giving AI the ability to work with your entire meeting set would create a lot of power. And that’s basically the direction we went into, and continue to go into.”
I keep pushing the metaphor — partly because the metaphor is doing real work for both of us. “Just because the tide has risen, you still have to keep your head above water. Ideally, it’s like you’re doing it through a boat. And like, hopefully the boat’s rising with the tide. And I’d imagine like when this new disruptive technology came up, it started creating some leaks in the boat. We’re like, oh shoot. We have to fix the direction of the boat, fix the leak, and then how do we build and improve the boat now that this technology is freeing up a lot of capabilities.”
He doesn’t disagree, exactly. He restates the problem in his own register, which is calmer than mine: the meeting note isn’t going away. They’ll keep maintaining a best-in-class meeting intelligence product for their existing customers. But the new product — the one that’s going to matter — is going to be built on top of an idea about what kind of organization actually needs Sembly.
Professional services.
“It’s, we’re kind of midway in realizing that strategy change because once you reach a certain level of momentum as a company, you can’t just pivot on a dime.”
Midway. That’s the most honest word a founder can say about a transition. Most pretend they’re done before they’ve started. Artem says they’re midway and you can hear that he means midway.
The professional services pivot makes sense in the way good positioning makes sense — quietly, after a few sentences. Consultants live in meetings. Meetings produce information that goes nowhere unless someone manually extracts it. Consultants also have clients, and clients have projects, and that two-level hierarchy is the difference between a tool and a system that understands your work.
Artem describes the agentic version: “Whereas when we started it was open-ended — it was very ChatGPT-like, where you can just go in and ask for anything you want. We are starting to create very agentic centered experiences where the agents are very specialized in creating certain kinds of materials and output.”
Then the example that reframes the whole product category for me.
“Imagine being able to on-demand generate case studies that are specific to a particular client situation. So you’re a professional service organization. Maybe you’re selling marketing or you’re selling technical services or maybe selling HR services, whatever it is. You have clients and then the client comes to you and says, okay, we’re considering you as a vendor, we’re also considering these other vendors. How cool would it be if you had some discovery calls with that potential client and then you can generate a case study that showcases that you’ve done something similar to what the client is asking for another customer.”
He keeps going. “That’s something that most agencies can’t afford to do. Like that’s a long and difficult process. Only the biggest agencies can really invest in doing something like that.”
The case study — the real one, the kind a Bain or a McKinsey would produce — on demand, for the conversation you’re already in. Not a generic one-pager. The artifact that wins the deal.
I get a question back that he says he’s been getting all the time lately: but can’t they just go into Copilot or Claude and ask it to generate a case study? Upload some documents. Press the button.
“This is super interesting because there’s kind of like a completeness factor to the output.”
Then the line I write down twice:
“If you’re using Claude, it will help you, but it will take you like 15 to 20% of the way, and you’re gonna need to spend another 80% on making that presentation real, usable, client-shareable. Like it’s way far from client shareable.”
Fifteen to twenty percent.
I think about what I do every week. I draft something with Claude. I think it’s done. I then spend the rest of the week massaging it into something I’d put my name on. The fifteen percent feels like more in the moment, because the page was blank and now it isn’t. But the work I’m doing after is the actual work.
“The general platforms, including Copilot, they have a last mile problem, which is they can only get you so far, but there’s specialized logic and mechanisms you need to get it to a client presentable level.”
I ask if he thinks Copilot will close that gap. He doesn’t think so. He thinks specialized companies — Gamma, in his example — own the last mile, because the last mile isn’t about model capability. It’s about completeness within a specific output format. Brand. Layout. Context. Voice. The eighty percent.
“And the reason that’s not true is because we’re so used to going down the checklist of features — does it have this feature? Does it have that feature? But it’s no longer about features. It’s about achieving goals or achieving certain kinds of experiences, and those things have more qualified metrics than just yes or no.”
This is the bigger frame underneath everything. Software is shifting from features to goals. The deliverable is no longer “does the system have this capability.” The deliverable is “did your customer attain the outcome.” And nobody’s scoreboard for that exists yet.
I push him on something else, because the language is starting to remind me of consulting more than of software. I ask if there are partners at consulting firms who are bad at delegating.
“I don’t think there’s a single partner that I’ve met that’s bad at delegating. Because once you’re a partner you’re only delegating — you’re actually not doing any of the work. And that’s part of, I think, what helps you to rise through the ranks in management consulting. Is effective delegation.”
Then the wrinkle.
“But I wouldn’t say it like delegation. I guess delegation is one way of talking about it. I think it’s more about pure communication.”
He hedges around the word communication because it’s fluffy — and says so. So he restates it sharper.
“It’s about relating to someone else in a very sharp way the idea that you have so that that person has the exact same idea.”
He keeps going. “And so this is where the metal is tested because first of all, it has to be sharp in your mind first. And I think that’s a lot. You know, sometimes that’s where things fall apart. Like if you don’t really know what you are trying to do, you’re gonna have a very difficult time trying to get someone to figure out the thing that you haven’t figured out yet that you want done, right?”
Then the phrase I cannot stop turning over in my head:
“Communicating that effectively to someone else so that they can get a twin representation of that idea in their mind.”
Twin representation. Not a copy. Not a brief. Not a spec. A twin — an idea that exists in two minds, fully dimensional in both, sharable from there.
“That includes things like providing the right kinds of context, bringing in context that would be important to understand in the given situation, providing counters like — this is, you know, we don’t want it like that, we want it more like this. And the end audience is like that, and it should have these performance characteristics and criteria.”
Context, counters, audience, performance characteristics, criteria. He’s describing what a senior consultant does to brief a junior associate on a Tuesday. He’s also describing what a good prompt looks like. He’s not pretending these are different skills.
“But ultimately you have to convey that idea in all of its dimensions to that other mind. But that other mind could be human, but today it could also be AI.”
The reframe lands sideways, the way good reframes do. I tell him I think it never was delegating. It was directing. The senior person never wanted to make the slide. They wanted the slide made to a standard they owned. The whole point of the junior analyst is that the analyst is responsible for executing to that standard. The senior’s job is communicating the standard, not lowering it.
“That’s a really nice reframe,” I say back, because what I really mean is: I wonder how many of the people complaining about AI are actually complaining about a directing skill they never built.
We go a little further. The interoperable agents future. The MCP interface Sembly is about to ship, where you can sit in Cursor and say: “That bug fix that we discussed on our call with Max, take the details of that bug fix and fix the bug that we discussed in my code.” Cursor pulls the meeting context from Sembly. It just happens. Vendor lock starts to look antique. Best-of-vertical agents talking to each other across companies. A different shape of software entirely.
But the part of the conversation I keep returning to, days later, isn’t the demo or the architecture or the pivot. It’s the geometry of what Artem is describing.
When the moat becomes the floor, you don’t try to dig the moat deeper. You don’t insist that what you built was hard, that the new arrivals are cheating. You climb up on the new floor and look at where you are. If you’re early, you’re standing on infrastructure no one else has — speaker disambiguation, mid-track language switching, the simulated meeting participant, four years of compliance and edge cases. If you’re late, you’re standing where everyone else stands.
Then you ask the question Artem asked: what’s the next leap?
The answer he gave is harder than it sounds. The leap isn’t features. The leap is goals. The leap is the eighty percent. The leap is the artifact at the end — beautiful, branded, factual, referenced, client-shareable — and the difference between a company doing real work and a company that’s added a button.
The leap, I think, is also the directing skill. The twin representation. The sharpness of the idea in your head before it leaves your head.
That part doesn’t have a moat at all. It has a floor — the same floor everyone else is standing on. The difference is what you can describe from there.
Guest Bio: Artem Koren
Artem Koren is the Co-founder and Chief Product Officer at Sembly AI, a meeting intelligence platform that transcribes, analyzes, and synthesizes professional service meetings into structured work products. Rising to prominence in the early 2020s, he became widely known for building enterprise-grade AI transcription and natural language processing systems before large language models made such capabilities broadly accessible — engineering custom models from scratch for a category that had no clear precedent. The company, which he co-founded in 2019 alongside Gil Makleff, operates across 35 languages, holds SOC 2 Type 2 certification, has raised $4.64 million in total funding, and was named in the 2025 Gartner Innovation Guide for Generative AI Technologies.
Previously, as Senior Manager and Director in the IT Capital Markets Services practice at EY, Koren spent more than a decade advising Fortune 500 clients across financial services, auto insurance, energy, and professional services sectors in North America and Europe. He was recognized as a top 1 percent performer and became known for deploying enterprise-scale work management applications — hands-on experience in the client service delivery cycle that would later define Sembly’s product thesis.
Earlier in his career, Koren served as CEO and CTO of Visual Trading Systems, where he built and distributed technology solutions for the capital and commodity markets, and co-founded Neusana, applying deep learning to digital biopsy image analysis. He holds a BS in Computer Science and Economics from Columbia University and an MBA from NYU Stern School of Business.
Hey,
Thanks for reading this. I mean that. There's a lot of content out there competing for your attention, and you spent some of it here. I hope it was worth it. Even better, I hope it prompted you to think about something differently enough that you'd share it with someone who'd get something out of it too.
I started this podcast because tactics never stuck with me. What stuck were stories — business biographies, autobiographies, the decisions people made and why they made them. The principle only clicks once you know the story behind it.
So I built the thing I wanted to read. Every week I have two conversations with people who build in technology and product. Then I write the essay I wish I could find — one that puts you inside the conversation, through my eyes. What caught me off guard. What I kept thinking about after we hung up. Where the principle actually lives once you strip away the jargon.
I make this for myself first. If you read the way I do, you’ll want it too.
PS — If you want to pitch coming on the show, or you know someone I should talk to, shoot me an email at caden@hey.com with "January752" in the subject line so it gets past my filters. I'm not optimizing for famous guests. I'm optimizing for interesting conversations, even from people who aren't LinkedIn influencers.
