By Stuart Breyer, CEO, mallowstreet
The FCA’s Mills Review, published last week, has been read across the investment industry as an efficiency story. If AI can not only recommend but act – reallocating, rebalancing and switching within agreed limits – then discretionary and advisory propositions can run at a scale and cost not previously possible. That it can do this is true, but it is also the least interesting thing about the report. The efficiency is not the story. The accountability is.
Mills is explicit that AI is moving from assistance to action. For anyone running a model portfolio service or an advisory proposition, that sharpens one question above the rest: when a decision is part-automated, who owns the outcome, and can you evidence the logic behind it?
A machine can execute a switch in milliseconds. It cannot, by itself, explain to a client or a regulator why that switch was right for that person at that moment. That explanation is where the liability now sits.
The industry has spent two decades driving cost and friction out of the transaction. Execution is cheap, platforms are fast, and rebalancing is a button. But how is accountability factored into these transactions?
Accountability lives behind the client’s circumstances, and the objective and trade-off that made an allocation suitable rather than merely permitted. As AI takes over the action, the trade record tells you less and less about whether the outcome was right. Firms that assume their transaction data is their audit trail will find out that it is not.
See also: Artificial intelligence: The environment’s ally
This is where AI accountability collides with consumer duty. The duty asks firms to evidence good outcomes and fair value continuously, not to show that a process was followed once. When the process is automated, following the model is not evidence of a good outcome; it is only evidence that the model ran.
What a firm needs is the reasoning: a record of the client understanding that sat behind the decision, still legible after the event to a compliance officer, a client, or the regulator. Transaction data cannot reconstruct that. Capturing the decision itself can.
It is a solvable problem, and some firms are already doing it. The reasoning that matters are mostly spoken, in the client meeting, the investment committee, and the suitability discussion. Capture it, structure it and check it, and you have the evidence layer the automated part of the proposition cannot produce for itself.
See also: FCA announces review into AI impact on retail markets
This will separate the market. Propositions that pair automated action with a captured, auditable record of the reasoning will scale safely, because every decision can be explained after the event. Propositions that automate the action and leave the reasoning to chance will scale their liability exactly as fast as their assets. The difference will not show in a good year. But it will show the first time an automated decision goes wrong, and someone asks the firm to show its working.
So, what does an investment business do now, before the agentic tools arrive in earnest? Three things. Decide where human judgement sits, and document it rather than assume it. Capture the reasoning behind decisions at the point it happens, not reconstructed months later for a file review. And treat the tools you adopt as part of your regulatory perimeter, because general AI trained on the open web is not built to the evidential standard your permissions demand – and the gap will show exactly where it costs most.
The Mills Review asks whether AI can run more of the investment process. It plainly can. The harder question it puts to your firm is narrower: when the machine acts and the outcome is challenged, can you still show why? The propositions that can answer that will own the next decade. The ones that cannot will spend it explaining themselves.
