The New Baseline: Managing Money After AI

For most of investing history, the alternative to hiring a professional was doing nothing, and doing nothing was a weak competitor. Cash under the mattress loses to inflation. A savings account loses more slowly. Even the index fund, the first great democratisation, required the investor to choose it, hold it through a crash, and resist every headline in between. A manager's pitch, for a century, quietly rested on that comparison: whatever our flaws, we are better than the mattress.

That comparison is over. An investor today can open an account and have software build a diversified portfolio, allocate across markets, rebalance on schedule, apply factor tilts that were institutional products a generation ago, and harvest tax losses lot by lot, for a fee that is now a small fraction of what managed portfolios traditionally cost. Artificial intelligence is now being layered through those tools: better interfaces, better explanations, faster customisation. None of this requires talent on the investor's side, or time, or discipline beyond the initial decision not to interfere. The baseline an external manager must beat is no longer inertia. It is a competent machine, running patiently, nearly free.

We state this as a manager because it disciplines everything else we say. The burden of proof on professional management is now permanent. It does not reset after a good year, and it cannot be met by being slightly better at the things the machine does, because the machine does them at a price no firm can match. Any manager whose value proposition is allocation, rebalancing, and a factor tilt is competing with a product that costs almost nothing, and should expect to lose. The honest question, and it is the only question this essay asks, is what remains.

The strongest counterargument deserves its full weight first. It says: the same tools are available to the professionals, so the baseline has risen for everyone and nothing structural has changed; worse for our thesis, the retail tools are shallow, built for the average case, and the first serious crisis will reveal the difference between a portfolio and a product. There is truth in both halves. Professionals do use the same machinery, and we use it ourselves, extensively, at the periphery of our process. But the conclusion does not follow. When a capability becomes universal and nearly free, it does not stay part of what a client pays a premium for, whether or not the premium firms also have it. Typesetting became free and publishers stopped charging for it. The gap did not close; it moved. And as for shallowness: the average case is most of life. A tool that handles the ordinary ninety percent well, and cheaply, has not left the professional ninety percent of the fee. It has left the ten percent where things break, and repriced everything else to zero. That repricing is the event. Managers are not being replaced. They are being confined, correctly, to the residue where judgment is worth more than its cost.

Where does the residue live? First, in the integrity of the data the decisions rest on. Automated tools consume the past as they are given it: revised series presented as known at the time, universes with failures removed, histories that no investor could actually have traded on. Garbage at this level is not a detail; it propagates through every allocation the machine makes, silently and at scale. Building point-in-time data, preserving corrections instead of overwriting them, and knowing exactly what a dataset claims versus what it can support is slow, unglamorous, human-accountable work, and it decides the value of everything downstream. The machine will optimise beautifully over a history that never happened if you let it. Somebody has to not let it.

Second, in validation under regime change. The baseline portfolio is calibrated, explicitly or by inheritance, to the world that produced its data. Systematic strategies earn or lose their keep when the regime turns: when correlations that were low for a decade converge in a month, when the relationship a tilt rests on inverts, when liquidity that was a background assumption becomes the binding constraint. Testing a strategy only on the world that made it look good is the standard failure of automated products and of most human ones. Knowing which parts of a process are robust across regimes, and refusing the parts that are merely fitted to one, is a judgment discipline, not a feature a platform ships.

Third, in crowding and capacity. The machine gives everyone a similar portfolio, and similarity is a risk the individual user cannot see from inside their own account. When a factor, a theme, or a whole model becomes the default setting of millions of portfolios, its returns carry the seeds of their own reversal, and the exit, when it comes, is shared. A professional process that measures its own footprint, watches what the crowd holds, and treats size as a risk rather than a trophy is doing something the baseline structurally cannot: the baseline is the crowd.

Fourth, and oldest, in the investor. There is a gap between what a fund earns and what the average dollar invested in it earns, and it exists because money arrives after strength and leaves after weakness. The arithmetic is simple enough to do on the back of an envelope. Imagine a fund that returns the same amount over two years either way, falling first and recovering second. An investor who holds throughout earns the fund's return. An investor who adds after the recovery and withdraws after the next fall earns something worse, though the fund itself did nothing different. Repeat that across many investors and many cycles and the investors' collective return falls short of the fund's, by an amount that depends only on their timing. A machine-built portfolio does not close that gap; if anything, a frictionless app makes the costly trades easier to execute at the worst moments. One durable part of a manager's job, unfashionable as it sounds, is standing between an investor and their own worst month: through terms that encourage patience, communication that says what is happening and why, and a process that does not itself panic. The behaviour gap can be larger than a fee. It is the one inefficiency that better technology has so far made easier to commit.

There is a reason the careful use of artificial intelligence in this industry points the same way as our argument. A firm that puts a model at the core of its decisions, unexplained, inherits a review problem it cannot solve: when the model is wrong, nobody can say whether its reasoning failed or its world changed, and those two failures demand opposite responses. A firm that keeps the model at the edges, where its work can be checked, and keeps the core decisions in a form a person can state and defend, inherits a slower process and a reviewable one. The choice is not nostalgia. It is a conclusion about where the risk sits: an unexplainable decision cannot be reviewed, and a process that cannot be reviewed cannot be trusted with the decade in which its assumptions break.

A principle is only useful if it changes what happens on an ordinary Tuesday, so we have turned this one into a habit. Every material decision in our process is meant to be tagged, at the time it is made, with the part of the residue it belongs to: data integrity, validation under regime change, crowding and capacity, risk under convergence, or investor behaviour. A decision that cannot be tagged is a decision the baseline could have made, and it is a candidate for automation or for removal. Over time the tags form a ledger of where human judgment was actually spent, and the ledger can be audited against outcomes. If a category turns out to absorb effort without changing decisions, that is evidence the residue is smaller than we thought, and we would rather learn that from our own records than from our investors.

So the new baseline, stated plainly. Assume the reader can obtain a competent, diversified, tax-aware portfolio for almost nothing, because they can. Assume every routine task in portfolio management is now priced at nearly zero, because it is. What a manager may charge for is the residue: data that does not lie about the past, validation that survives the regime it was not fitted to, awareness of the crowd and of one's own footprint, risk that is watched when correlations converge rather than when reports are written, and a structure that protects investors from the timing instincts the data says they have. That residue is smaller than the industry's fees imply and harder than the industry's marketing admits. It is also, we believe, real. Our own standard follows from it: if we cannot show, in a specific case, which part of the residue a decision belongs to, the machine should have made it, and nearly free.