Is AI replacing human traders, or just the equity desk and not the credit desk?

This week Barclays put a number on something many of us suspected from our own statements. Its strategists Zornitsa Todorova and Andrea Diaz Lafuente surveyed 410 institutional investors across North America, Europe, the Middle East and Asia, and the headline, reported by Bloomberg on June 4 and Hedgeweek the next day, was that AI has quietly become routine. Nearly three-quarters of the hedge funds in the sample now use it every day. Among long-only managers the figure was 49%, and among asset owners 38%. That is not a pilot phase. That is infrastructure.

Then comes the part worth sitting with. When the same respondents were asked whether AI plays a real role in actually executing a trade, the answer was a near-unanimous no. Depending on the group, somewhere between 77% and 88% said the technology plays a very limited role at the point of execution. So the machine reads everything and decides almost nothing about the fill. And the gap is widest in credit, the corner of the market where the survey was focused.

AI saturates research but stalls at trade execution

This is not a Barclays quirk. The Mercer 2026 survey of asset managers, published in late May, found 55% have AI embedded in at least one investment process, 91% plan to use more of it within a year, and yet only 5% let it make autonomous or semi-autonomous investment decisions. Every serious data point lands in the same place: AI has saturated the research layer and stalled at the decision layer.

The easy reading is that everyone is being cautious, that the lawyers and risk officers are holding the line until the models earn trust. That reading is comforting and mostly wrong. The line is not being held by caution. It is being held by plumbing. And the plumbing is different depending on what we trade, which is why the same investor can feel half-replaced as an equity screener and barely touched as a credit picker.

The mechanism: why the machine can read credit but cannot run it

Here is the part that took me a while to see clearly. The reason AI eats the equity research desk faster than the credit desk is a market-structure fact, not a talent fact.

A share of a large listed company trades on a continuous central exchange that broadcasts a live, executable, consolidated price thousands of times a second. That is a dense, time-stamped, machine-readable river. Point a model at it and the screening and pattern work automates quickly, because the data is exactly the shape a model wants.

A specific corporate bond is the opposite. It does not sit on a continuous exchange. To buy it, a desk sends a request-for-quote, meaning it asks a handful of dealers to each name a price, rather than hitting a public order book. The trade then prints to TRACE, the post-trade reporting tape that FINRA (the brokerage industry’s self-regulator) runs, only after the deal is already done. A single company can have dozens of separate bonds outstanding, and any one of them might trade twice in a day or not at all.

So the data a model would train on in credit is sparse, late, and missing the one input that actually moves a block: which dealer is sitting on an inventory position they need to clear, and where real size can be done. That information lives in dealer relationships and chat color, off any feed a model can subscribe to. The human credit trader’s edge is not that they are smarter than the model. It is that they can see a market the model is structurally blind to.

Where the durable human edge in credit trading sits

A few things get easier to hold once the gap is named as structure rather than sentiment.

First, it tells us where the durable human edge actually sits, and it is not where the headlines point. The replaceable work is the dense, continuous, lit-market screening. The defensible work is the relationship-priced, intermittent, off-feed negotiation. If we hold credit managers, the question for them is no longer “do you use AI” but “what part of your edge depends on data a model will eventually get, and what part depends on data it structurally cannot.”

Second, it reframes the survey’s own anxiety. The respondents named data security and privacy as their biggest barrier, not model quality. As Hedgeweek summarized the finding:

The machine reads everything and decides almost nothing about the fill. In credit, that is not a stage we are passing through. It is the shape of the market.

Third, it is a quiet permission to do less. If we run credit ourselves and feel behind because we have not wired a model into our execution, the survey says the largest, best-resourced desks in the world have not either, and not because they are slow. The honest move is to use AI exactly where the data supports it, in reading and screening, and to keep the human hand on the part of the trade that no feed describes.

What electronifying corporate bond markets would change

I want to steelman the other side, because the structure is moving. The electronic share of investment-grade corporate bond trading has climbed from roughly 25% in 2019 to somewhere near 45% to 50% today. Portfolio trading and all-to-all venues, where buyers and sellers meet without a dealer in the middle, are growing. Dealer axes, the inventory signals I called off-feed, are increasingly published as structured data. Every one of those shifts hands the model a slightly richer river.

If credit keeps electronifying, the moat narrows. The day a meaningful slice of bond liquidity carries a continuous, executable, machine-readable price, the credit desk starts to look more like the equity desk, and the displacement curve steepens. The mechanism is real today. What is not guaranteed is that it stays real, because the thing protecting the human trader is a data gap, and data gaps are exactly what this industry spends its money closing.

So the question I keep turning over is not whether AI will run a credit desk. It is whether we would even notice the handover, given that it would not arrive as a smarter model. It would arrive as a better feed.