Most marketing ideas that sound impossible aren't limited by imagination. They're limited by cost. That limit just moved.
A founder I know sells an AI ad optimiser. Her company has three people: a CEO, a CTO and one marketer. By her numbers, they do about $5 million in revenue and around $4 million in profit. They produce thousands of pieces of content, their cost per lead is a few cents, and they have almost no SaaS subscriptions.
They aren't smarter than you. They noticed earlier that nearly everything in marketing can be automated, and that the ideas you drop as "too expensive" are mostly just priced for last year.
This month the price dropped again.
The cost of a decision is going to zero
On 15 September a lab called TypeSafe released a model called Jev. It doesn't write. It decides. You give it a messy situation and a narrow question — is this lead real, which of these twelve topics fits this founder, is this quote accurate — and it returns a structured answer with a confidence score in 70 to 500 milliseconds. Input costs about four cents per million tokens, and output is free. On TypeSafe's own benchmarks it's roughly 190 times faster and 440 times cheaper than frontier models on these tasks. Those are their numbers on their tests, but the direction is what matters.
A chatbot made writing cheap. Jev makes judgement cheap, and marketing is mostly judgement: which lead, which angle, which person, which channel, which version.
Lead agencies are in trouble
Here's what that means for growth right now.
Tools like Monid work like an app store with one checkout for AI agents. An agent describes what it needs in plain language, finds a tool, checks its price and runs it, paying per call from a single balance. The catalogue covers thousands of tools, including web search and scraping, people and company enrichment, social platforms, and image, video and voice generation. Many calls cost a fraction of a cent.
Put a fast decision model on top and your agent becomes a researcher who never sleeps. It finds a company, enriches the contact, verifies they still work there, decides whether they fit, and drafts the first message. In my experience, you can get that down to about a cent a lead, verification included.
So think about what you currently pay for per-seat subscriptions and list-selling agencies. My bet is that most of that category goes away. Your competitors will run the same pipeline, which is exactly why you can't skip it.
A thought experiment: FOMO at 100 stories a day
Let's make this concrete with the publication you're reading.
FOMO published 150 stories in its first six months, which is good for a small newsroom. Now imagine that 130,000 Estonians (10% of the nation) have a story another founder could learn from. Revisit each of them every four or five years, and you need about 30,000 conversations a year. That's roughly 100 a working day. You wouldn't publish them all. Maybe the best ten.
Everyone deserves a chance. Not everyone gets published.
The ranking shouldn't come from an editor's mood on a Monday, either. A story's real audience often arrives months later, when an investor asks an AI assistant about a founder before a meeting, and your interview turns out to be the best source. You rank on the story's potential over time, not on who happened to open LinkedIn that morning.
A human newsroom can't do 100 interviews a day. A pipeline can. Here's how I'd build it.
1. Research the person before anyone talks to them. A coding agent with search and enrichment tools pulls everything public about the founder: past interviews, their company's history, what they post, what they avoid. Then it proposes three or four topics that only this person can speak to, and prepares them. Being an expert in someone else's industry used to take days. Now it takes minutes.
2. Let AI do the interview. This part makes people nervous, so here's the evidence. Economists at the London School of Economics found that AI-led interviews were rated roughly on par with an average human expert interviewer by researchers who judged the transcripts blind. On sensitive topics, many participants preferred talking to an AI because it didn't judge them. A University of Chicago study of more than 70,000 job applicants found that an AI voice interviewer drew out richer, more interactive conversations than human recruiters. The honest caveat is that AI interviewers follow the script unless you design them to chase the interesting thread. That isn't a reason to avoid them. It's a design problem, and design problems have solutions.
It also helps that this is asynchronous. You're not on live television. The founder answers when it suits them, at 11pm, in their own words.
3. Turn one conversation into everything. From one transcript: the article in Estonian and English, quote cards, short video clips, an audio version, and a clean profile page. Run a quick decision check on each output: does this quote match the tape, is the company name spelt consistently, does the headline claim anything the person didn't say?
4. The editor-in-chief decides. They get a finished piece, not raw material. They kill it, publish it, or send it back with follow-up questions, and the AI goes back to the founder to ask. People read FOMO because they trust the people behind it and what those people choose to publish and cut. You're paying for that judgement. The machine just gives it far more to judge.
5. Learn who actually cared. After publishing, an agent reads who reacted to the post on LinkedIn, researches them and tracks what happens in their world. Over time you build a small map of your real readers: who they are, what they want and how that changes. Give your agents access to that map, and the next round of topic choices gets sharper.
Every junction in that pipeline is a cheap decision: rank, route, verify, approve, escalate. That's where Jev earns its place. A thousand small judgments a day used to be unaffordable. Now they cost about the same as a coffee.
Build it like a software factory
The best template for this doesn't come from marketing. It comes from software.
Earlier this year, the security company StrongDM described what it calls a software factory. Humans define what the system should do, the scenarios it must handle, and the constraints that matter. Agents do the rest: they build, test against real-world behaviour and iterate until the work passes. The key trick is a set of test scenarios the agents never see, so they can't game them.
Apply that to content, and you get the most important rule in this piece: a question can fail; the system shouldn't.
When the AI misses the real story in an interview, don't just fix the article. Ask why the system missed it. Was a skill missing, like how to push for a number? Was a tool missing, like access to the founder's old podcast appearances? Was the brief wrong? Fix that, then add the failure as a test so it never happens again. Every correction your editor makes becomes a permanent improvement to the machine.
StrongDM's team has a mantra for this that works for marketers too. For every task you're doing, ask: why am I doing this? The model should probably be doing it instead.
Vertical AI wrappers are in trouble
This also explains why I'd be nervous selling a vertical AI product like a legal assistant or an AI sales rep. Open one up, and it's usually a model, some skills and a few tools. A capable team can now rebuild that to a good-enough standard in weeks.
The only lasting moat is data nobody else has. For a publication, that's the interview tapes, the reader map and the editor's taste turned into rules. For your business, it's whatever you know about your customers that isn't on the open web. Build the pipeline yourself, and that data compounds for you, not your vendor.
Practical rules for setting it up
Learn to code, and do it in a real tool. Coding agents like Cursor or Claude Code are far better for this than no-code app builders, which in my experience fall apart the moment you need something custom. The interface rarely matters much. The pipeline behind it does. It takes less skill than you think, and it changes how you see every problem.
Spend your own time on System 2. Jev is System 1: fast, cheap, instinctive decisions. Your job is the slow thinking — what your brand stands for, what insight only you have, what you refuse to publish. If you hand that to a general-purpose chatbot, you'll get the median answer, and so will everyone else who asked.
Write for agents, and repeat yourself. Clicks from search keep falling, and agents now read your content many times more often than people do. They're the ones summarising you to your next customer. Say your core message clearly, in the same words, everywhere you appear.
Pay growth people two to three times the market rate. One person running a pipeline like this does the work of a department. Expectations are far higher, so pay should be too. Everyone on the team will code.
Why this matters for Estonia
AI is excellent in English and noticeably weaker in Estonian. That's partly because the language is hard, which we can fix. Mostly it's because there isn't enough Estonian data to learn from. We can't make more Estonians. We will never have as many speakers as English or Chinese.
But we can make far more Estonian stories. We've punched above our weight before; Bolt has spent a decade out-executing Uber from a country the size of a large city.
If we want our language, culture and stories to still be heard in ten years, most of what we publish will be written by machines. That's fine, as long as Estonians are the ones judging, editing and deciding what's worth keeping. Nobody is stopping the AI revolution. We might as well teach it to speak Estonian properly.
--
Karl-Gustav Kallasmaa is an AI maximalist and founder of Attensira, an Estonian platform helping brands dominate AI search visibility.
Comments ()