Vertical AI is the product layer that survives when the underlying model becomes cheap. Artem Koren’s team at Sembly AI spent three years building a transcription engine. They deleted it once transcription fell from two dollars an hour to cents. The defensible layer was above it: client context, brand systems, and the same finished deliverable every time. Grounded research and repeatability are what founders in commoditizing categories still sell.
How do you keep an AI product defensible as foundation models continue to improve?
Sembly AI’s answer is to build the layer a general model has no reason to build. Asked what Sembly’s moat was, Artem Koren named consistency over proprietary data. A chatbot takes a different route on every run, and every colleague arrives with different prompts. A product returning the same result each time survives a model upgrade.
About Artem Koren
Artem Koren is co-founder and Chief Product Officer at Sembly AI. His team built a transcription engine good enough to beat what Google, Amazon, and Microsoft were shipping in 2019. They retired it three years later, once transcription fell to cents an hour. Founders across every AI-adjacent category now face the same problem. The capability they spent years engineering is available to anyone who can call an API. Working out what is still worth selling is harder than building the thing was. He first appeared on the show three years ago, when Sembly was still building that engine. Koren’s work rebuilding Sembly above its own commoditized layer shows what holds up in vertical AI.
Watch the Episode
Table of Contents
Vertical AI Means Owning the Layer the Model Cannot Reach
Artem Koren does not think ChatGPT is a product. He calls it a platform service, useful for a phenomenal number of things, yet delivering none of them reliably. A refrigerator is a product, because you knew what it would do before you bought it.
Koren describes the stack in aviation terms. Carbon fiber is the raw material, and some companies make nothing else. Others specialize in design, others in certification, and a few build the aircraft.
The foundation model is the carbon fiber. Sembly AI sits at the top layer, where a stated goal becomes a finished deliverable. Koren calls that the target user experience.
The pages currently ranking for this term define the category differently. SymphonyAI’s glossary defines vertical AI as artificial intelligence purpose-built for a single industry and trained on domain-specific data. IBM defines the same agents by their specialized algorithms and industry-specific data.
Bessemer’s founder playbook frames the category as competing for labor budgets rather than IT budgets.
All three definitions describe what goes into the system. Koren defines the category by what comes out of it. He notes that Microsoft spans everything from Azure compute to Office without fully owning any single layer.
In Koren’s framing, which layer a company occupies matters more than which industry it serves.
Vertical AI, in Artem Koren’s framing, is the layer where a goal becomes a finished deliverable. The foundation model is raw material, like carbon fiber in aviation. Companies that sell the material compete on price. Companies that build the aircraft sell an outcome the buyer can name in advance.
What It Costs When Your Own Moat Goes to Zero
Sembly never wanted to be a transcription company, and vertical AI was not the plan either. Koren describes building the engine as an unfortunate quest the team was forced into. What came out of Google, Amazon, and Microsoft in 2019 was not up to par.
The moment he knew it would end came at the end of 2021. Koren sat down with GPT-2, handed it a paragraph, and asked what it was about. The answer came back well written and correct.
Other models had attempted that job before without doing it well. Koren told his team that day that everything was going to be different.
Good transcription cost $1.50 to $2.00 an hour when Sembly started. Today Koren puts it at a few cents, or 15 to 20 at a premium model.
The transcription price fell more than tenfold, on the capability his engineers had spent three years perfecting.
Koren told the team to take a moment to appreciate what they had built. Then they were to let it all go and switch to the new thing. The developers were melancholy about it, he says, and the company moved on.
Koren is unsentimental about why that was the right call. Good transcription meant Sembly could stop transcribing and return to its actual business.
When I asked what the delay cost, Koren estimated millions in lost revenue and customer productivity. He then undercuts his own figure.
The models that make the current product possible did not exist a year earlier. The counterfactual company may never have been buildable at all.
Artem Koren’s team retired a transcription engine it had spent three years building. The market price had fallen from around two dollars an hour to cents. Sembly AI treated the collapse as a release rather than a loss, because transcription had never been the business. The lesson Koren draws is that a capability worth deleting is rarely what customers were paying for.
Consistency Is the Moat, Not the Data
I asked Koren directly whether the moat was the document layer or the meeting data. Koren picked neither and named consistency instead. The data helps, he says, but consistency is the bigger moat.
Koren’s answer puts him at odds with the pages ranking above him. Greylock’s 2023 thesis names proprietary data generated through product usage as the long-term moat. NEA argues that specialization drives defensibility, built on proprietary data and ownership of systems of action.
Both are investor positions about what should hold up over time. Koren is describing what actually happened inside his own company.
Koren’s reasoning is operational rather than theoretical. A chatbot decides its own path on every run, so the second deck differs from the first. A colleague who has never seen your prompts starts from nothing.
Koren’s customers do not want a different good answer each time. Koren’s phrase for the standard is that it produces the same awesome every time.
The hardest thing Sembly built was not any single generation step. Koren names the agentic framework itself: runtime, sub-agent handoff, coordination across models, and context that grows without breaking.
Deploying an agent and asking it to build a deck is easy. The engineering problem is getting that same standard on the hundredth run. It has to hold for a colleague who was not there for the first.
Consistency is also Koren’s answer to the LLM wrapper question that investors have asked him about for two years. Every piece of software wraps an operating system, and every operating system wraps a CPU.
Every database wraps a hard disk. Koren’s conclusion is that everything is an LLM wrapper, and the label decides nothing.
| What the ranking pages say the moat is | What the operator says |
|---|---|
| Greylock: proprietary data generated through product usage | Koren: consistency, with data as a supporting input |
| NEA: data moats plus ownership of systems of action | Koren: a framework that returns the same standard every run |
| Bessemer: domain focus and labor-budget positioning | Koren: repeatability across a whole team, not one user |
Artem Koren names consistency, not proprietary data, as the durable moat in vertical AI. Greylock and NEA both argue the opposite in their published theses. Koren’s case rests on repeatability rather than data. A general model produces a different result on each run, while a purpose-built product returns the same standard every time.
How Much of a Client-Ready Deck Can AI Actually Produce?
Koren estimates that a general tool gets you 15 to 20 percent of a client-ready deliverable. The figure is his own, drawn from watching customers try. The remaining work is where he argues the product lives.
He describes the six-step chain as large firms actually run it:
- A VP sets the deck strategy, the core message, and the key examples.
- An associate researches the client’s domain, languages, register, and whether they buy on cost savings or revenue growth.
- The associate maps that research into a slide sequence with a specific ask at the end.
- A design team applies the branding, whether the agency’s own or the client’s.
- The associate reviews, rejects the slides that are not visually strong, and builds any model the argument needs.
- The VP signs off, and the deck goes to the client.
Koren says every large house works this way and smaller agencies run a slimmed-down version.
Against that chain, a chatbot covers a narrow slice. Koren allows that it might get you some research and a certain flow, after enough prompting. It will not get you the design, and PowerPoint still takes a long time.
Microsoft closed part of that gap in April 2026. Copilot agent mode reached general availability in Word, Excel, and PowerPoint, honoring corporate templates. Koren’s answer is that Copilot holds the universe of your information without any concept of a specific client.
Sembly’s mechanism is a knowledge deck. Before the generation runs, the system shows the facts and the claims it surfaced through its research, each with its grounding. The user accepts or rejects them, reviews the outline, and only then gets the deck.
Artem Koren estimates that general AI tools produce 15 to 20 percent of a client-ready deliverable. The rest is client context, brand systems, slide design, grounded research, and formatting. Sembly AI has users approve a grounded ledger of facts before generation starts. Reviewing a finished document afterward is the step he designed around.
Why Do Humans in the Loop Miss Fabricated Work?
Reviewers miss fabricated work because it looks correct. Deloitte and PwC both had review processes, and both shipped fabricated client work anyway.
Deloitte produced an A$440,000 report for Australia’s Department of Employment and Workplace Relations, published in July 2025. It contained invented academic references and paper titles that do not exist. A quote from the Amato v Commonwealth case was misattributed.
Deloitte partially refunded the contract in October 2025. PwC’s 2025 Transforming Governance report scored 84 percent AI-generated on GPTZero’s detector. The report was built around a Citizen Pulse framework that does not exist.
The PwC report claimed Denmark, Saudi Arabia, the United States, and Australia as users of that framework. GPTZero flagged four PwC reports in total.
I put both failures to him. His diagnosis runs to six words. The humans in the loop were looping themselves.
His point is that the failure mode was never an error. AI is very good at making things look right, he says. A reviewer facing a plausible document has nothing to catch.
Critical thinking and professional experience are the only defenses he names for a person. For a product, he names grounding. Every fact and claim Sembly surfaces carries its source, whether a meeting, a document, or a web reference.
Koren does not stop at reassurance. He expects the review relationship to invert within a year or two, as models pass the people checking them.
Senior professionals become more valuable in his account, because they retain the standing to say that something is wrong. He is direct about who loses that standing. The middle guys and girls, and certainly the juniors, will struggle to push back on anything AI gives them.
His question for them is the uncomfortable one. What ground are you standing on to push from?
Deloitte refunded part of an A$440,000 Australian government report in October 2025 after AI-fabricated citations were found in it. PwC’s 2025 Transforming Governance report scored 84 percent AI-generated on GPTZero. It invented a framework and attributed it to four national governments. Artem Koren argues both failed because reviewers read work that looked correct rather than work that was grounded.
Where Vertical AI Wins, and Where a Chatbot Is Enough
Koren does not argue that everyone needs his product. Asked when a general tool is the right choice, he answers: one-off work. If you want a single deck, spend some time with Claude or ChatGPT, and you will have one.
The economics change on the second run. Koren describes doing the same task again ten minutes later, for a different client. The result has to be completely different each time.
Repetition at that frequency is where a purpose-built product earns its place. A team doing it 50 times a month cannot reinvent the prompt fifty times.
The clearest fit is multi-client work. Agencies, consultancies, and any firm holding separate context for separate accounts need that context to persist between jobs.
Meeting-heavy organizations are the second fit, because the source material already exists. Koren notes that Sembly captures across Zoom, Teams, and Google calls, while Copilot only sees Microsoft Teams.
Bessemer’s founder playbook frames the same buyer differently. Vertical AI competes for labor budgets rather than IT budgets. That puts the purchase next to the team doing the work.
Koren is equally clear about where it loses. Hiring, promotion, and exit decisions stay with people, and so does any conversation the participants have not agreed to.
There is a third category worth naming: the pilot that never becomes a habit. The pattern behind AI pilots that deliver no return is that nobody opens a tool twice. A one-off deck generator is exactly that shape.
Repeatability is what separates a bought tool from an adopted one.
Koren’s test for whether a vertical product is worth buying is repetition rather than capability. One-off work belongs in a general chatbot, where prompting cost is paid once. Multi-client, high-frequency work belongs in a purpose-built product. The context, the brand system, and the standard have to survive from job to job and across a team.
Where a Vertical AI Product Should Refuse to Operate
Koren draws the first line for vertical AI between presence and authority. There are more domains where AI should not decide anything than where it should not be present.
His public position is that AI should never decide who gets hired, promoted or managed out. He extends it to legal matters related to HR and to anything involving a violent incident.
For everything else, he refuses to write a universal rule. HR practice is cultural, and organizations differ in what they will tolerate.
The reframe he offers is the useful part. When AI is privy to your conversation, it means AI has a seat in the room. Recording is not the question, because participation is what people are actually consenting to.
Koren’s answer to the question of who decides is straightforward. The people in the discussion do.
Koren is candid that the product enforces almost none of this. Sembly emails participants before the call, joins as a visible participant, and announces that it is recording. The duty still sits with whoever brought the AI into the room.
Sensitive data gets a harder rule. Personally identifiable information requires a product built for it, with defined storage, retention, and access controls. Koren’s rule is that pointing a general tool at that content is the mistake.
He rates Europe as further ahead of the United States on privacy, and calls the extra work worthwhile.
His sharpest objection is reserved for silence. Koren names Granola directly for promoting local recording, which lets someone capture a conversation without telling the other person.
Frequently Asked Questions
What are verticals in AI?
Verticals in AI are the specific industries a system is purpose-built to serve. Examples include healthcare, legal work, and construction. SymphonyAI’s glossary defines vertical AI as artificial intelligence trained on domain-specific data for a single sector. Artem Koren adds a second test. A vertical product delivers a named outcome, where a general platform only makes one possible.
What is the difference between vertical AI and a general AI platform?
A general AI platform is capable of many things without committing to any one of them. Artem Koren compares ChatGPT to a service rather than a product. You cannot say in advance what you will get. A vertical AI product commits to one. A stated goal produces a finished, on-brand deliverable, and the result repeats on every run.
Is every AI company just an LLM wrapper?
Every AI company wraps a model, and the label settles nothing. Artem Koren’s response to two years of investor questions was simple. Every piece of software wraps an operating system. Every operating system wraps a CPU, and every database wraps a hard disk. Koren’s test is whether the wrapper delivers something the model underneath cannot.
How do you keep an AI product defensible when foundation models improve?
Build the layer the model has no reason to build. Koren names consistency as the moat. It rests on client context libraries, a brand system, and grounded research. The finished standard is the same for every colleague. Greylock and NEA both name proprietary data instead. Koren’s counter is that repeatability is what a model upgrade cannot erase.
Should you let an AI notetaker into a client meeting?
The people in the conversation should decide, because an AI that hears the discussion is participating in it. Artem Koren argues that recording is the wrong question and that a seat in the room is the real one. He rates disclosed bots as safer than local recording tools that capture a conversation without telling anyone.
What to Do Before Your Category Commoditizes
Artem Koren’s team spent three years building a transcription engine and was better off for retiring it early. That pattern applies to any B2B tech founder whose product sits close to a fast-moving model layer. In Koren’s account, what survives a price collapse is the accumulated client context and the brand system applied to it. The product must also meet the same standard on the hundredth run as on the first.
Koren’s two years of investor questions suggest buyers do not work this out on their own. Founders in commoditizing categories have to make the argument publicly and repeatedly. Sproutworth’s digital PR and authority work is built for founders in exactly that position. The goal is to make the case for what your company now sells visible to the people evaluating it. The question for founders in this position is simple. Which part of your product would you still be proud of if the underlying model became free next quarter?
Related Resources
- Why AI Projects Fail: 95% of Pilots Deliver No Return: what separates a bought tool from an adopted one
- Enterprise AI Adoption: 1,000 Seats, One Daily Login: the adoption gap behind most AI spend
- How to Sell AI When Buyers Aren’t Ready and Still Win: positioning a category the buyer has not defined yet
- AEO for B2B SaaS: The Founder’s Playbook for Getting Cited by AI: how to become the answer AI engines quote
- Enhancing Productivity With AI: Artem Koren’s first appearance, recorded three years earlier
Related Links
- Check out Sembly AI
- Get a copy of Artem’s book Untethered: Unraveling human nature through instinct, culture, and reason
- Connect with Artem on LinkedIn
Some topics we explore in this episode include:
- AI’s shift from premium technology to a commodity is reshaping business value.
- How Sembly AI’s pivot from transcription to client-ready documents brought operational challenges and new lessons.
- The impact of widespread access to core AI on competition and innovation.
- What happens when cheap, universally available AI services like transcription disrupt business strategies.
- Why true product value now depends on branding, context, and completeness, not just the underlying AI.
- The challenge of building high-quality client deliverables with AI and overcoming platform shortcomings.
- How smaller players are carving out a niche against AI giants like Microsoft and OpenAI.
- Why consistency, repeatability, and data traceability are critical for professional AI use.
- The debate between usage-based versus subscription pricing models for AI products.
- The importance of ethical boundaries, privacy, and consent when deploying AI in sensitive business decisions.
Listen to the episode
Subscribe to & Review the Predictable B2B Success Podcast
Thanks for tuning into this week’s Predictable B2B Podcast episode! If the information from our interviews has helped your business journey, please visit Apple Podcasts, subscribe to the show, and leave us an honest review.
Your reviews and feedback will not only help me continue to deliver great, helpful content but also help me reach even more amazing founders and executives like you!
