The signal

On Tuesday, May 26, Coinbase released a tool that lets an AI assistant act inside your crypto wallet. Two trading days later, on Thursday, May 28, Robinhood announced Agentic Trading. The feature lets retail customers connect third-party AI agents to their brokerage accounts. Those agents can monitor markets, rebalance portfolios, and execute trades inside guardrails the user sets.

The two announcements happened five days apart on opposite sides of the regulated financial system. One is a crypto exchange. The other is a traditional brokerage. Both did the same thing in the same week. They handed AI a seat at the trading desk that used to be reserved for the human holding the account.

Set those two stories against a number from the International Monetary Fund and the major global exchanges. AI-driven algorithms are now estimated to handle close to 89 percent of global trading volume, up from roughly 60 percent in the early 2020s. That number describes the institutional layer. It describes what already happens every trading day on the New York Stock Exchange, Nasdaq, the London Stock Exchange, and the major commodity venues. AI is not arriving in finance. It has been embedded inside it for years. What changed this week is that the same logic reached retail.

This column is about what that means, and why it matters that it is happening at the same moment the assets themselves are being rebuilt on blockchain infrastructure.

What AI already does, quietly

The deepest deployment of AI in the financial system is not visible to most people, because it lives inside three professions. Trading desk execution, asset manager credit analysis, and quantitative portfolio construction.

On the trading desk, execution algorithms decide how to break a large order into smaller pieces to minimize market impact. A pension fund that needs to sell $400 million of a single stock cannot drop that order on the market without crashing the price. An execution algorithm slices it across hours or days, routes pieces to different venues, and adjusts in real time as it watches order book depth and volatility. The newest generation of these algorithms uses machine learning, a form of AI that learns from historical data, to anticipate which venues will have liquidity at which moments. The buy-side trading desk is no longer a row of humans yelling at screens. It is one or two humans monitoring outputs from machines that have already made the decisions.

At the asset manager, large language models, the same kind of artificial intelligence behind ChatGPT and Claude, are now being used to read financial statements, pre-screen public filings, and automate parts of credit risk evaluation. A junior analyst at an investment bank used to spend a week pulling out the relevant ratios from a 200-page annual filing. A language model does that in minutes. The model does not replace the analyst. It hands the analyst a finished draft that the analyst then audits. Work that used to take days now takes hours. A team that used to cover thirty companies now covers ninety.

A recent academic benchmark called FinanceQA tested current language models on the kind of multi-step financial analysis tasks junior analysts at hedge funds and private equity firms actually do. The models failed about 60 percent of them. That is the honest ceiling on this technology as it stands today. Not perfect. Not even close. But on the 40 percent the models get right, they are doing work in seconds that used to take a person hours.

In portfolio construction, the third pillar, machine learning models build diversified holdings by analyzing correlation patterns and tail-risk exposures across asset classes. The math behind mean-variance optimization, factor analysis, and similar frameworks has been quantitative for decades. The new piece is that the models now adapt in real time to new data rather than being recalibrated quarterly. A model that detects a regime change in correlations on Tuesday adjusts the portfolio on Wednesday, not at the next rebalance window.

These three layers, execution, credit, and construction, are what produce the 89 percent figure. Almost none of it touches a retail account.

What just changed

This week, the kind of AI that has been running quietly inside institutional finance arrived at retail.

Robinhood's Agentic Trading launched on May 28. The platform now lets customers connect third-party AI agents to their accounts. Those agents can read the user's portfolio, monitor markets, rebalance positions, and execute stock trades, all inside guardrails the user sets. The company has built in spending limits, separated trading accounts, notifications, and an instant shutoff. The structure is conservative, but the precedent is significant. A major American retail brokerage is allowing software to make trading decisions on a customer's behalf.

Coinbase made a structurally similar move on May 26 on the crypto side. Through what is called a Model Context Protocol server, Coinbase wallets can now be controlled by an AI assistant like Claude or ChatGPT. A user can ask their AI to swap tokens, deposit into yield positions, or pay merchants that accept stablecoin payments, and the AI executes the actions inside the user's wallet. The keys remain with the user. The AI never holds custody. But the AI initiates real economic activity in a real account.

Two announcements, five days apart, from two prominent platforms in their respective markets, doing functionally the same thing. The agentic AI layer reached retail.

The combined effect is straightforward. The same kind of decision-making that used to be reserved for trained professionals at trading desks is now available to a twenty-three-year-old with a Robinhood account and an AI subscription. The quality of the AI is uneven and the FinanceQA limits apply. But the access has flipped permanently. Retail with AI is not the same retail it was a year ago.

Why this matters now

AI's expansion into retail trading on its own would be a story. AI's expansion into retail trading at the same time the underlying assets are being rebuilt on blockchain infrastructure is a different story altogether.

Our recent column on the SEC's innovation exemption covered what is happening on the asset side. Blockchain-based versions of American stocks are moving from offshore experiments to a regulated path inside the United States. The DTCC, which is the entity that settles American securities, is going hybrid this summer. The largest asset managers in the world, BlackRock and Franklin Templeton among them, have tokenized Treasury funds already operating with billions in assets on-chain.

Now put the two trajectories next to each other.

On the asset side, equities, Treasuries, commodities, and private credit are moving onto blockchain infrastructure. They settle instantly, trade around the clock, and can be programmed.

On the decision side, AI is taking over more of the work of choosing what to buy, when to buy it, and how to execute, at every layer from institutional desks down to retail accounts.

When tokenized assets settle instantly and AI makes decisions in milliseconds, the gap between a thesis and a position closes to the speed of the slowest piece. The slowest piece used to be the human. That piece is being removed. What remains is a market in which an AI agent can read a fresh data release, evaluate the implications against a user's portfolio, place a trade, and have it settled on-chain before the human user has even looked at their phone.

That is a different financial system than the one that existed when retail meant putting a buy order in at the open and waiting two days for it to settle.

The structural skepticism

A few things should temper this picture before anyone makes it grander than it is.

AI in finance still gets a lot wrong. The 60 percent failure rate on real analytical tasks in the FinanceQA benchmark is a real ceiling, not a marketing number. Production deployments at banks and asset managers exist alongside heavy human oversight precisely because the models are not yet reliable enough to be left alone. The buy-side trading desk that uses an execution algorithm still has a human monitoring the algorithm's output. The credit analyst using a language model to draft a memo still reads the memo. The portfolio construction model still has a portfolio manager who decides whether to apply its recommendations. These systems are advisory, not autonomous, almost everywhere they matter.

AI at retail has the same problem with thinner guardrails. Robinhood's Agentic Trading and Coinbase's wallet AI both put guardrails in place, but those guardrails depend on the user setting them correctly. A retail customer who hands an AI agent broad authority over their portfolio without understanding what the agent can actually do is exposed in ways that an institutional desk with compliance officers and risk managers is not. The infrastructure is here. The literacy to use it safely is not.

Regulatory direction can shift. The current administration is broadly favorable to AI deployment and to tokenization. A change in direction could mean stricter limits on AI-initiated trades for retail customers, mandatory disclosure of AI involvement in execution, or restrictions on how third-party agents connect to brokerage accounts. None of those would stop the underlying trajectory. All of them would slow it.

And the model layer itself is still a fast-moving target. The agent that works today may produce different outputs tomorrow as the underlying model is retrained. The audit trail question, who is responsible when an AI executes a bad trade, is unresolved in both crypto and traditional finance. None of this is reason to dismiss what is happening. It is reason to track it without overstating it.

The signals ahead

A few things over the coming months will indicate how seriously to take this trajectory.

Watch how many retail brokerages follow Robinhood. If Charles Schwab, Fidelity, and the major bank brokerages roll out comparable AI agent connectivity in the next six months, the agentic retail model becomes standard rather than experimental. Watch how many crypto wallets follow Coinbase. If MetaMask, Phantom, and the other major wallets adopt comparable AI integrations, the same model becomes default in crypto as well.

Watch what regulators say about retail AI deployment. The SEC's recent posture has been receptive. A speech, a no-action letter, or a proposed rule from the agency in the next quarter would clarify the regulatory perimeter.

Watch the FinanceQA-style benchmarks. If models go from 40 percent task accuracy to 70 percent over the next year, the case for AI as an autonomous decision-maker in finance gets materially stronger. If accuracy stalls, the human-in-the-loop model stays in place for the foreseeable future.

The frame

The financial system has been integrating AI for over a decade. What changed this week is not the technology. It is the layer. The same logic that has been running on institutional desks for years just reached the retail account. And it reached it at the same moment the assets themselves are being rebuilt on blockchain infrastructure.

The two stories are not parallel. They are intersecting. An AI agent making trades on tokenized assets that settle instantly is a fundamentally different market participant than a human investor making trades on Apple stock that settles in two days. Both will exist for a long time. But the second is now possible in a way it was not six months ago, and that possibility is going to compound.

For anyone holding crypto, the practical implication is that the AI agent layer just became part of the demand for the infrastructure layer. Wallets, oracle networks, payment protocols, and the chains that host them now have an entire new class of automated user.

For anyone holding stocks, the practical implication is closer to home. The retail brokerage you use is now competing in a market where AI agents on the other side may have a real-time information advantage. Whether you opt in to that layer or not, you are now trading against it.

The convergence this publication has been mapping has been about crypto, blockchain, finance, and AI meeting. This is the week the AI piece met the other three at the same time.

The system you trade in has changed. Trade in it carefully.