Anton Lobintsev has spent more than twenty years building companies at the intersection of infrastructure, regulation, and product innovation, from enterprise server integration and legal tech to today’s stablecoin‑powered payments rails. As co-founder and CPO of SquareFi, he is now helping fintechs and global platforms move money across 150+ countries and 25+ currencies on a unified infrastructure that ties together business accounts, cards, wallets, and crypto‑fiat conversion into a single settlement layer.
But Lobintsev’s most provocative ideas are not about any single product launch; they’re about what happens inside the team when AI starts writing most of the code. At SquareFi, the core engineering group has shrunk while throughput has grown, with roughly 95% of production code now authored with AI assistance—freeing human engineers to think in systems, clarify specifications, and reserve their attention for the moments that truly require judgment. In parallel, a non‑technical co‑founder has assembled a live internal CRM in weeks, and the design team now uses AI‑powered Figma workflows to ship prototypes before engineering ever touches the queue, expanding who counts as a “builder” inside the company.
Across this conversation, Lobintsev argues that the real constraint in AI‑heavy startups is no longer coding capacity but specification quality: the ability to describe problems precisely and to validate outputs without hand‑waving. He draws a hard red line around large, irreversible financial decisions—payment authorizations, compliance sign‑offs, AML determinations—that must remain human‑gated even as agents prepare the analysis, draft the orders, and watch the logs. And he lays out a people philosophy for the AI age in which curiosity, domain depth, and comfort with ambiguity matter more than traditional credentials, offering a roadmap for founders who want to build tiny, high‑leverage teams without abandoning accountability.
Many founders hear numbers like “95% of code is AI‑assisted” and assume it’s a cost‑cutting story. How do you explain the real resource reallocation that happens when you shrink a team yet increase throughput?
The cost-cutting framing misses what’s actually happening. When AI handles repetitive logic, it fundamentally changes where skilled people spend their attention. Engineers are thinking more carefully about architecture and spending more time on problems that truly require human judgment. The increase in throughput comes from a dramatic shift in the ratio of meaningful work, not because engineers are working longer hours. We’ve also seen code quality improve, not decline, because people have more bandwidth to think clearly about what they’re building, rather than just grinding through implementation.
ALSO READ: What “High-Risk AI” Actually Means for the Teams Running HR, Finance and Customer Ops
If another fintech founder told you they want to go from a 20‑person engineering team to a 6‑person AI‑native team, what are the first two or three disciplines you’d insist they master before touching headcount?
First, system design. You need people who can think in systems, who understand how components interact, where failure propagates. Second, specification writing. This is underrated and genuinely hard. The quality of what you get from an AI model is almost entirely determined by how precisely you can describe what you want. That’s a skill, and most engineers haven’t had to develop it deliberately because historically they could figure things out iteratively with a team. Third, validation. Knowing whether a result is correct and having the domain depth to read an output and immediately sense whether something is off. The cost of shipping something subtly wrong is high, so you need people who can catch that before it becomes a problem.
You described a non‑technical co‑founder at SquareFi building an internal CRM in roughly two weeks that combined deals, accounting, fundraising, agents, and partner workflows, with parts of it live in production while it was still being built. What did that experiment reveal to you about who gets to be a “builder” in an AI‑first company?
Our co-founder didn’t suddenly learn to code, and he doesn’t need to for that matter. It was that he understood the business logic better than anyone, could describe exactly what he needed, and could immediately tell when something wasn’t working the way it should. That’s the core of building. What that experiment told us is that in an AI-first company, domain expertise and the ability to think clearly about systems can be as valuable as technical credentials. That opens the building to a much broader group of people than most companies currently assume.
ALSO READ: DXC’s LabX is Beating AI Theatre
You’ve argued that large financial transactions should remain human‑gated: agents can prepare a payment order, but a human must sign it; similarly, final compliance decisions carry legal responsibility that sits with the officer, not the agent. Where do you draw the hard red line on what should never be fully automated in fintech?
The red line follows legal and financial liability. If a decision carries a signature, say a payment authorisation, a compliance sign-off, an AML determination, a human has to own it. Not because the agent can’t prepare it accurately, but because accountability has to sit somewhere real. An agent can certainly do the analysis, surface the relevant information, flag the risks, and draft the order. But the moment of commitment has to be human.
Beyond liability, I’d draw the line at anything where the cost of a wrong decision is asymmetric and hard to reverse. A large payment sent to the wrong counterparty, or a compliance decision that exposes the company to regulatory action, cannot be undone easily. The more irreversible the consequence, the more important it is that a human is in the loop at the decision point and not just reviewing logs after the fact.
You’ve written that in many startups today “specification quality is the constraint” and that the most valuable people are those who deeply understand the domain, can describe systems precisely, and can validate results without hand‑waving. How do you identify those people in interviews?
We only work with engineers who are already leveraging AI, so that filter alone brings you a pool of professionals who are already deep into the technology and genuinely passionate about its capabilities. From there, you can provide a real problem and watch how they describe the solution before they start building anything. The quality of their specification (how precisely they decompose the problem, what edge cases they surface unprompted, what assumptions they name explicitly) tells you almost everything. You can also ask them to validate something by giving them an output and asking whether they believe it’s correct and why.
ALSO READ: ‘Do Not Fall in Love with the Prompt’: Ellis Crosby on Building Outcome‑First AI
What kinds of people struggle the most in an AI‑heavy environment like SquareFi, and what have you found that helps them make the transition rather than be left behind?
Our team embraced AI early on, so it has mostly become about learning, testing, and understanding where it delivers and where it falls flat. We also don’t impose specific tools; we allow team members to find what works best for them and their workflow. That made the transition far less challenging than it might sound. And it’s not just engineering; our design team uses AI heavily too, including Figma plugins that convert designs directly to HTML, which then lets them build small prototypes for first-level testing before anything reaches the development queue. Overall, many ideas get tested earlier, without waiting for engineering capacity.
That said, the people who’ve struggled, even if only initially, tend to be those focused on the overlap between where their role ends and the AI begins, and where their skills still matter. That’s an understandable reaction, but what helps most is giving low-stakes opportunities to work alongside AI tools on real problems. The transition happens through experience rather than through being told the world has changed, and they need to catch up. They get curious about what they can now build that they couldn’t before.
ALSO READ: When Money Moves Itself: Why Agent‑Readable Banks Still Need Human Guardians
