We use the word AI to describe all sorts of different things, from talking to ChatGPT in a browser to running entire businesses where operations would simply cease to exist in their current form without language models. Where is the line between using AI automation and AI being at the core of a business? At what point does a fundamental shift occur? And most importantly, should companies even strive to put AI center stage?
When Is a Business AI-Native?
Today, even if management has never announced an AI transformation, employees are still using Gemini to search for information, ChatGPT to write an email, Grok to prepare a report, Claude to write their code, and so on. Access to LLM tools has become so widespread that simply using them no longer tells you much about the company itself.
For me, there is a much simpler test: what happens if you switch AI off?
Take an accounting software company. Its developers may use AI to write features or maintain the product. Turn those tools off tomorrow, and the engineers will probably work more slowly. But the software will still exist, accountants will still use it, and customers will still pay for it. AI improves how the company operates, but the business itself does not depend on it.
Now imagine a bank call center that might once have employed ten thousand people. The first generation of AI in that environment worked almost like a prompter sitting next to the operator: it listened to the conversation, understood the problem, and suggested an answer. In many cases, that prompter has now become the agent itself. A customer says they were charged the wrong amount; the system checks the records, reviews the relevant data, identifies the issue, and sometimes resolves it without involving a person. A human steps in only when the case needs escalation.
That is a very different relationship with the technology. The accounting company becomes less efficient without AI. The call center may need to rebuild a large part of its operating model. That difference is much closer to what I mean by AI-native than how many AI subscriptions a company has or whether “AI” appears on its homepage.
Where Is This Most Visible?
Some companies have built their business around helping others make that transition.Cohereis a good example: the Canadian company grew into a major business by helping enterprises move from traditional software and workflows toward AI. In other industries, though, AI is already so deeply embedded in the product that it becomes almost impossible to separate the two.
Self-driving is probably the clearest example. If a car is expected to see the road, recognize objects, understand what is happening around it, and make decisions on its own, then you cannot remove AI without removing the product’s basic idea.
Medicine and pharmaceuticals make the same point in a different way, especially in drug discovery. One of the best examples of what this can mean at scale is DeepMind’s work on protein structures.
The Protein Data Bank was created in 1971, and it took decades of work by scientists around the world to build up a large collection of experimentally determined structures. Thedatabasepassed 1,000 structures in 1993, 10,000 around 2000, and 100,000 in 2014. By January 2023, it had passed200,000experimentally determined biomolecular structures. In other words, after roughly six decades of structural biology, researchers had accumulated on the order of two hundred thousand experimentally determined structures.
Then came AlphaFold, DeepMind’s system for predicting the three-dimensional structure of a protein from its amino acid sequence. In 2022, AlphaFold made more than200 millionpredicted protein structures available, covering nearly every cataloged protein known to science at the time. That does not mean AI “discovered every protein that exists.” These are predicted structures based on already known amino acid sequences. But the difference in scale is still enormous.
The importance of that shift became even clearer in 2024, when Demis Hassabis and John Jumper of Google DeepMind received half of the Nobel Prize in Chemistry “for protein structure prediction”. The other half went to David Baker for computational protein design. Something that only a few years earlier looked like an impressive AI breakthrough had very quickly become recognized as a fundamental advance in modern biology.
DeepMind is the global, headline-grabbing example, but AI-first medical startups are no longer unusual. During my years at Keymakr,our team worked on projectsranging from dentistry and ultrasound imaging to interventional cardiology, in which models were trained to detect heart valves, the aorta, and other structures during procedures. We also worked on neurosurgery projects where models learned to distinguish brain tissue, blood vessels, surgical instruments, and the operating area. In all of these cases, AI was not an extra feature added to the product; it was the product itself.
From Protein Structures to a Cold Email
The same thing is happening in much more ordinary parts of business. Sales and marketing are probably among the fastest-moving examples: work that used to take hours can now be repeated hundreds or thousands of times with very little extra effort.
CRM, campaigns, prospecting, and lead research are built around language models. We are already far beyond the point of asking ChatGPT to write a cold email. A system can look through someone’s LinkedIn profile, see what they do, what they care about, and what they post about, then use that information to write something that feels very personal.
I once got a message along the lines of: “Michael, it must be incredible to fly at 4,000 feet and look out through the glass…” It referred to my aeone had actually gone through my LinkedIn profile and read it carefully. Of course, nobody had done that. The system had found the information and automatically generated the outreach
That is where the economics of personalization start to change. Research that once took a person a long time can now be assisted by tools that scan public information and identify relevant details. A human can then decide what is actually useful, appropriate, and worth including.
I saw the same shift atKeymakr, though for us, the goal was never to automate human judgment or relationships. I kept encouraging people to look at the repetitive parts of their daily work and ask where agents could save time, bring more consistency, or help them work at a larger scale.
AtScryon, we started with that logic already built into the company. Much of the marketing work is handled by several agents across articles, the blog, LinkedIn, and the website.
The interesting part is that their biggest advantage is often not writing. Say you add twenty new pages to a website. A person writing a new article will probably add a few internal links. Almost nobody is going to go back through 80 or 800 older pieces and check where those new pages should now be linked. An agent can. In that kind of work, the agent does not need to be smarter than a person. It just does not get bored, and the difference between 80 pages and 800 pages is much less important to it than it is to us.
But this is where automation reaches its limits for me. Good PR can’t simply be handed over to an agent and forgotten. The same applies to an opinion piece, where positioning, audience understanding, and the ability to decide what to say are crucial.
So even when you build a fairly complex system of agents, you still need a person running it. Someone has to judge the output, decide what matters, and set the direction.
Can You Automate a Founder?
In an early-stage startup, the founder is often technical, operational, and executive all at once. So the thought comes naturally: if agents can take a lot of work off the team, why not give them part of the founder’s job as well?
I have a separate agent connected to Scryon Business that knows a lot about the company. I can ask it something like, “We want to get 5,000 users next month.” Is that realistic, or is it just a nice number I came up with?
It can pull data, look at the market, check competitors and connected systems, and come back saying that 5,000 is probably too much and even 2,000 would be aggressive at this stage. That helps me test my own thinking. But the agent is still giving advice.
The decision is still mine. After all the data and calculations, sometimes I still need to go outside, walk through a park, and decide where I want to take the company.
For me, real offloading starts when I can give someone ownership, say, “This is yours,” and come back a month later to see that it has kept working without me. I do not see that with the founder role today.
The same applies to C-level roles. Their job is to set direction: six months from now, we want to enter the automotive market, or we want people in the industry to know the brand. VPs, heads, and team leads then turn that into specific goals and work.
AI Toolkit: Personal and Company Costs
I do not really believe in one “best” model. I care much more about which one works best for a particular task and what it costs me to use it at scale. I often use Codex for coding and rotate models inside it, depending on my needs at the time. I might try Claude for design one day, and switch to ChatGPT for doing more language-related tasks, but these choices change month-to-month, and sometimes in the middle of a task.
Ultimately, everyone builds their own stack over time, judging which model is optimal for a specific goal. If you are a technical founder who codes a lot yourself, I would expect you to spend around $1,000 a month on your own stack. But I know people who spend $10,000–15,000 a month because they automate much more aggressively.
Once AI becomes part of the core business, the numbers change again. With Scryon, I am comfortable with the idea that 30–40% of revenue may go directly into the infrastructure and services the product runs on, before you even count the people building it.
First the Business Goal, Then the Technology
In the race for automation and the label “AI companies,” I think the very question of “what kind of AI should we implement?” is dangerous. It’s starting from the wrong end.
We recently set ourselves a fairly standard business goal – to reduce costs. To do this, we created an agent, gave it access to information about where the company was spending money, including Amazon, Google, and other services, and asked it to regularly look for anomalies and raise red flags if anything looked odd.
In just the first week, we reduced costs by approximately 30% because the system quickly showed where money was being spent on things we’d practically forgotten about. The agent itself can cost as little as $50 per month, but the savings can be in the tens of thousands.
At Keymakr, AI primarily strengthened people and individual processes. At Scryon, it’s already built into the company’s architecture. In self-driving, without it, the product disappears, and in drug discovery, it changes the scope of exploration.
For me, the question of underlying goals is a much healthier starting point than the desire to become an AI-native company. First, a business goal must emerge: to earn more, scale faster, reduce costs, enable a new product, or radically change an existing process. Only then does it make sense to ask which technology will get you there.
