Where do your fractional shares actually trade when an AI agent rebalances?

A quiet Tuesday this week made the point for me. On June 23, with Big Tech under pressure and Alphabet sliding on news that a senior Google DeepMind scientist was leaving for Anthropic, an investor running an AI rebalancing agent would have generated dozens of small orders to keep target weights intact. Most of those orders were for amounts like 0.41 of a share or $250 of an index sleeve. I went looking for where, exactly, those fractional pieces trade. The answer is not the exchange. It is the broker’s own book, and almost nobody who runs one of these agents knows it.

This is not a scandal. It is plumbing. But it is plumbing worth understanding before we let software place a hundred of these orders a month on our behalf.

The split inside a single dollar-based order

Here is what happens when an investor, or an agent acting for one, buys 1.41 shares of something.

The order has two parts. The whole-share part, the 1.00 shares, is executed on what brokers call an agency basis. Agency means the broker acts as agent and routes the order out to a venue to find a counterparty, the same way a normal stock order works. The fractional part, the 0.41 shares, is executed on a principal basis. Principal means the broker takes the other side itself, filling the customer from its own inventory account rather than sending the sliver to an exchange. A fraction of a share cannot settle through the normal clearing pipes anyway, so the broker holds the whole share and parcels out the fractional interest internally.

Read one broker’s own disclosure and the mechanics are stated plainly. DriveWealth, which provides the fractional-trading rails behind a long list of consumer apps, says it executes the fractional component “against its principal facilitation account,” and that any leftover fractional interest in the whole share “may be allocated to other customers or retained by DriveWealth as principal.” In other words, the broker nets one customer’s 0.41 against another’s 0.59, keeps the rounding, and trades the resulting whole lots for its own account. The same two-part structure appears, almost word for word, in National Financial Services disclosures behind Fidelity fractional shares: full-share leg agency, fractional leg principal.

How is the fractional fill priced? The disclosures say the broker will, on a best-efforts basis, execute the fractional component at the best price visible across exchanges at the moment of execution, the number most people know as the national best bid or offer. That sounds protective, and often it is fine. But notice what it is: a snapshot the broker reads at a moment the broker chooses, not a fill that competed against other venues for that slice. And the broker says, in the same breath, that “when trading as principal for its own account” it “may make a profit or incur a loss on each trade.” On that leg, the customer is trading against the broker, by design.

Why an AI agent makes this leg the main event

For a human clicking buttons, fractional trading is an occasional thing. You round up a position, you put a spare $100 to work. The principal leg is a footnote.

An AI rebalancing agent inverts that. When Public launched what it called the first agentic brokerage on March 31, 2026, the pitch was that an investor describes intent and the agent executes, with example instructions like “if my checking account balance exceeds $20,000, sweep the excess into my direct index.” A direct-index sleeve is a basket of individual names held at exact target weights. To hold exact weights on a real-world balance, the agent has to trade fractional amounts of almost every name, almost every time cash moves or prices drift. Interactive Brokers, whose June 22 and 23 webinars this week included one titled “The Future of Wealth Management: How AI Is Reshaping the Advisor’s Playbook,” markets smart routing on its fractional trades, but the fractional residual is still a principal leg there too. The routing applies to the whole-share part.

So the rebalancing agent that feels precise and tidy is, underneath, a machine for manufacturing principal-leg fills at the broker. That is the part regulators noticed.

That date is the tell. The audit trail regulators use to reconstruct every order, the Consolidated Audit Trail, only started capturing fractional share quantities on February 23, 2026, with testing from November 2025. For years, the fractional legs that brokers filled as principal did not show up as their own data points in the record regulators inspect. The same Norton Rose review confirms the standard duties apply to these trades: best execution under the rule that requires brokers to seek the most favorable terms reasonably available, and the order-routing disclosures under the national market system rules. The duties were always there. The visibility arrived this year, which is a strange thing to be able to say about a feature this many people already use.

What this is worth checking before handing an agent the keys

The decision lens here is execution quality, and it is narrow and concrete. On the whole-share part of any trade, an investor can read the broker’s routing disclosure and form a view. On the fractional part, what is being accepted is a principal fill at a broker-chosen reference price, and the cost of that leg does not surface in any fill report we would normally check. For a once-a-month round-up, that is a rounding error on a rounding error. For an agent rebalancing a direct-index sleeve of 200 names every time the wind changes, the fractional legs are most of the trades, and the small structural cost rides on all of them.

It also informs tool fit. An agent that rebalances continuously to perfect weights is optimizing for a precision the principal leg quietly charges for. An agent that rebalances with a tolerance band, only when a weight drifts past a threshold, generates far fewer fractional fills for the same risk control. The setting that looks like laziness is often the cheaper one.

I am not against fractional shares or against rebalancing agents. I use fractional positions, and the access they give a normal-sized account to expensive names and to direct indexing is real. The point is smaller and more useful than alarm. The fractional leg is the one part of a trade that does not get the benefit of competition, and an AI agent is the thing most likely to multiply how often we touch it.

The question I keep turning over is this. We spent years learning to read our brokers’ routing disclosures so we could see how the agency leg of a trade gets handled. The fractional leg, the one our agents now generate by the hundred, has only had its own line in the regulators’ record since February. What will we find once a few quarters of that data exists to read?