Four technology areas touch this node's work, and each has a specific use case and a specific failure mode. Naming them separately stops the whole category being either dismissed or believed.
What technology is used in investment management?
Start with the analysis of very large and unstructured data sets, whose use case here is capital-market assumptions: alternative data (satellite imagery, card-spending panels, shipping records, job postings) can sharpen a growth or margin input that would otherwise be a judgement. The failure mode is that a series with three years of history and a survivorship-selected panel produces an input whose apparent precision is entirely spurious, and the input-sensitivity failure has already shown what a spuriously precise input does to a frontier.
Next come analytical tools including machine learning, whose use case is estimating a covariance matrix better than the sample estimate does, which is a real problem with a real literature: shrinkage estimators, factor structures and random-matrix filtering all reduce the estimation error that the input-sensitivity failure diagnoses. The failure mode is fitting a return forecast rather than a risk estimate. Covariance is estimable from a lot of data; expected returns are not, and no amount of model capacity changes that arithmetic. The quant-code region owns the implementations.
Third, automated trading and low-latency execution, which mostly does not touch strategic allocation; pretending otherwise is a category error. It matters at the execution stage of the portfolio management process, where implementation shortfall is measured, and it matters to the exchange-traded fund's authorised participants, whose arbitrage keeps the fund's price near net asset value.
Fourth, distributed ledger technology, whose live use cases are settlement, registry and fund administration rather than portfolio construction. AL1.03 owns the asset-class question; nothing in this section depends on it.
Advertisement
How does robo-advice work, and where does it fall short?
That leaves robo-advice, which is the one a client will actually ask about. The model is straightforward and worth stating without either the marketing or the sneer. A questionnaire elicits a risk profile, an algorithm maps that profile to one of a small number of model portfolios built from low-cost index funds, and the platform then rebalances to the model and, in taxable US accounts, harvests losses automatically. The advice is real, the portfolios are usually sensible, and for an investor whose alternative was an expensive closet-index fund or no plan at all, the improvement is large.
Four honest limits belong beside that.
The questionnaire measures willingness, badly, and ability, hardly at all. Twelve questions cannot establish a client's human capital, their job's correlation with the equity market, their liquidity needs over the next five years, or the concentrated stock position that dominates their balance sheet. PW1.01's economic balance sheet is exactly what the questionnaire omits, and it is the part that changes the answer most.
The rebalancing and tax-loss-harvesting claims are real but frequently over-quantified. Harvesting defers tax rather than eliminating it, its value depends on the gap between the client's short-term and long-term rates and on their eventual liquidation path, and the wash-sale rule constrains the mechanics. A platform advertising a fixed annual percentage of "tax alpha" is quoting an average over an unstated set of assumptions, not a number that applies to a particular client.
The fee comparison is genuine and is where most of the value sits. A typical robo platform charges 0.25% to 0.50% a year on top of underlying fund expenses of a few basis points, against 1% or more for a traditional advisory relationship and considerably more for a bundled distribution product in India. Over thirty years, a 0.75-point fee difference compounds to roughly a fifth of terminal wealth. That arithmetic, not the algorithm, is the product.
The regulatory treatment is that a robo-adviser is an adviser. In the US it registers with the Securities and Exchange Commission and owes the same fiduciary duty as a human adviser; in India it operates under the investment-adviser regulations, which since 2020 have separated advice from distribution. ES1.01 and ES1.02 own the standards themselves. The practical consequence for this node is that the algorithm's suitability logic is a regulated artefact, so the firm has to be able to explain why a given questionnaire answer produced a given allocation, which is an argument against the least explainable models rather than for them.
The place a robo-advice service is genuinely weak is the moment its client most needs it. An algorithm cannot talk somebody out of liquidating at a drawdown, and the behavioural gap between fund returns and investor returns is larger than any fee difference above. That is a service question rather than a technology question, and it is the reason hybrid models exist.