Going AI-Native: When the Firm Is the Product
AI-native went from rare to everywhere in under a year. It appears in product launches, hiring posts, and the research notes that reach allocators, usually without any detail on what it changes inside the organization. That gap, between the word and the operating model under it, is where credibility in this category gets decided over the next two years.
Our answer is one line: we started without the layers, and we run on shared context and shared intelligence instead of management ones. The rest of this is what that means and how we got there.
Ascend operates market infrastructure for regulated securities. That makes the question specific. AI-native means something different for a firm meant to hold onchain securities under stress than it does for a consumer app or a research lab.
Two versions of the same word
The first version is cosmetic. A firm announces a strategy, introduces a few internal tools, circulates a note. Six months later the team works the same way, the product looks the same, and the risk surface is unchanged. The shift is real to the people who wrote the announcement and invisible to anyone who depends on what the firm produces.
The second is structural. It changes how the firm decides, documents, stores institutional memory, replaces its own components, and draws the line between deterministic work and human-owned work. This change shows up over months as fewer surprises, tighter feedback loops, and a shrinking distance between what the firm says and what it does.
Both use the same word. Only one is load-bearing. Cosmetic AI-native is worse than not using AI at all: it signals a firm that took on the language without changing its substrate, and counterparties read that accurately. The cost compounds for the whole category, not just the firm that sent the signal.
What the market is doing, and where we sit
The clearest thinking has come from firms cutting. Cloudflare split its workforce into builders, sellers, and measurers: builders compound, sellers gain, and the measurers (middle management, much of finance, legal, internal audit) are largely replaced by AI, except the best, who are augmented. Coinbase flattened its org and set a hard ceiling, never more than five layers between the CEO and the newest contributor. Jack Dorsey argued that pure managers are finished, and that what survives is individual contributors, directly responsible individuals, and player-coaches.
We think those are basically true. Jamie Dimon reflects the same truths from the regulated end. JPMorgan redeployed staff rather than firing at scale and holds a consistent line: AI can process information, but accountability stays human, because compliance, risk, and trust do not compress the way a marketing function does.
The difference for us is that we never had to cut. Those firms are subtracting layers they already built. As a new firm, we reimagined the org before we went past a single team. We started without the layers, and add shared context and shared intelligence in their place, so a small team carries the capability that once required a large one, without the coordination tax.
A three-part operating model
The AI work at Ascend is run by three people with three distinct jobs, and the structure of the firm reflects it.
It started with two people and a sequence. Dennis O'Connell, who led the protocol build, stepped up to CEO. He is the engineer of the group, with two decades in derivatives and capital markets behind the code, and he runs the ERC-3643 securities standard that Ascend's compliance model is built on. He brought in Marcelo Fleitas as Chief Strategy Officer. Marcelo is the strategist and operator, with 27 years across emerging technology and large-scale financial infrastructure, including modernizing the listings and regulatory-technology stack at the NYSE. He was already at maximum conviction on what AI plus people could do, so as the two of them designed the org, they designed it AI-native. Besides identifying the AI bet early, Marcelo shaped Ascend's team around it, bringing in the people who could turn an AI-native intent into a firm that runs that way.
To co-design and implement AI Ops, Marcelo brought in Manny E. Reimi as Chief Product Officer: a zero-to-one product and ProductOps specialist at the intersection of DeFi and regulated finance, whose studio Escape Velocity Labs was early to shipping custom AI Ops builds and to running an AI-native product, engineering, and growth studio.
The three jobs stay separate on purpose. Strategy decides where AI belongs and where it must never touch a regulated decision. The build proves every such decision can be inspected later, and holds the line between deterministic work and human-owned work. Product turns the operating model into systems that actually run. No one of the three absorbs the others, because the moment strategy, build, and accountability collapse into one voice is the moment the audit trail stops meaning anything.
In their own words:
Marcelo Fleitas, Chief Strategy Officer. "The question was never whether to use AI. It was where AI belongs and where it must never touch a regulated decision. Drawing that line is the strategy."
Dennis O'Connell, Chief Executive Officer. "AI raised our capacity to keep the audit trail, not our permission to skip it. Accountability stays with a named human on every regulated decision. That order is the whole design."
Manny E. Reimi, Chief Product Officer. "An operating model that only lives on a whiteboard is a liability. My job is to turn it into systems a counterparty can inspect and a small team can actually run."
How we operate
These are operating principles, stated as how we work. Beneath them sit the values that hold a small team together: radical candour, maximal autonomy, and the highest leverage we can put behind each contributor.
We separate deterministic control from probabilistic control. Deterministic processes run by policy: parameters onchain, lifecycle actions on runbooks. Probabilistic work can use AI evaluation. We use optimistic controls where we can, on the probabilistic, non-regulated side, never on the path that touches a regulated decision.
A decision we cannot inspect later is one we will not make. The audit trail is built before the feature, across code, vault parameters, oracle policy, and internal calls. AI raises our capacity to maintain it. It does not relax the requirement.
Ownership is a clear human chain, and the work is split by topology. Separation of concerns, then context engineering against each concern, with deliberate depth-versus-width tradeoffs in how we run agentic operations. Every regulated decision traces to a named person.
We run in continuous loops, and we try to run them faster and at higher amplitude. Shared systems of record and shared systems of intelligence keep effort aligned and coordination cost low, so discovery, delivery, and operations compound instead of colliding.
We skillify our standard operating procedures. As we build our own harnesses, the procedures and best practices that work get encoded into them, so the floor under routine work keeps rising and the knowledge does not leave when a person does.
Every component is replaceable, including the ones we built. Custody is documented so another team could run it. Tooling is composable, not monolithic. We assume drift, vendor failure, and personnel change, and design for them.
We trade short-term excitement for long-term trust. We would rather ship a smaller piece of infrastructure a counterparty can verify than a larger one taken on faith.
Why the firm is the unit of trust
When a firm builds a consumer product, the product is what users assess. Bugs get fixed, versions ship, trust accrues over time. When a firm operates market infrastructure, the firm itself is what counterparties assess. Allocators cannot redeem a position because they liked last quarter's release notes. They evaluate the operator the way they evaluate a custodian: the operating model, the recovery paths, the audit trail, the cadence of disclosure, and what happens when something breaks. The protocol is the artifact. The firm operating it is the unit of trust.
The three-part model and the principles above are how we translate that into a firm-level operating model: not a category we are claiming, but a description of how we have chosen to work, open to inspection by anyone whose decisions depend on us being accurate about it.
The institutions that allocate to onchain market infrastructure over the next decade will look past the language of the moment and assess the operating model underneath. It's a long bet, and a quiet one. For a team building something this consequential for onchain finance, it is the only one worth making.
Ascend is the standards-first protocol building the credit market structure for onchain, regulated securities; and is built by the team at PSG Digital Labs, part of PSG Digital.