A trustee who cannot decide may delegate, and a trustee who delegates remains responsible. Both halves of that sentence are older than any computer, and neither of them changes when the delegate is a model rather than a person. What changes is the evidence. When a professional investment manager makes a decision, there is a person who can be asked why; when a model does, there is a weights file, a training set nobody at the firm has read, and a compliance memorandum written by the vendor.
The question is being asked of us with increasing frequency, usually in this form: our trustee wishes to use an automated portfolio system, or an automated distribution-screening system, or a large language model to draft the minutes; is that permitted, and what happens if it goes wrong? The answer to the first part is almost always yes. The answer to the second part is where the practice lies.
Delegation is not the problem
English law has been comfortable with delegation for a century and a half. The Trustee Act 2000 put the matter beyond argument for asset management: a trustee may delegate any of its delegable functions to an agent, provided that it prepares a written policy statement, that the agent agrees to comply with it, and that the arrangement is kept under review. The statute is drafted around a human agent, but nothing in it requires one. A model is an agent in the ordinary sense that it acts on the trustee’s behalf, and the vendor that supplies and maintains it is an agent in every sense.
So the first move in most of these files is to stop asking whether the model may be used and start asking what the policy statement says. In our experience it says almost nothing. We have read policy statements which describe an investment objective in three lines and then attach forty pages of vendor terms limiting liability to twelve months’ fees.
The standard of care does not move
A trustee must exercise such care and skill as is reasonable in the circumstances, having regard to any special knowledge it has or holds itself out as having. A lay trustee is not expected to audit a model. A trust corporation that markets itself as running a quantitative allocation process is expected to understand the process it markets. That asymmetry is not new — it is the point Brightman J made about the specialist trust corporation in Bartlett — but automation widens it, because the corporation now claims expertise in something its own staff cannot inspect.
The comfortable middle ground, in which a trustee buys a system, describes it in the annual review as “industry standard”, and relies on the fact that everyone else uses the same one, is exactly the position Lindley LJ warned against when he distinguished prudence from following the crowd. A practice that is universal is not thereby prudent; it is merely universal.
Three things we now advise trustees to record
First, the boundary. Which decisions does the system make, and which does it merely inform? A screening tool that ranks distribution requests and a tool that declines them are different animals, and the file should say which one is installed. We have seen a system described in the board minute as advisory whose output was, in operation, never departed from in eleven months.
Second, the review. Section 22 requires the arrangement to be kept under review, and review of a model means something more specific than review of a manager: it means knowing when the model was last retrained, on what, and whether its behaviour has drifted. A trustee that cannot obtain that information from its vendor has learned something important about the vendor.
Third, the override. Somebody must be able to say no, that person must be identified, and the occasions on which they have said no must be recorded. An override that has never been used is not evidence of a good model. It is evidence of an unused override.
Where it goes wrong
The claims we expect to see are not claims that a model was used. They are claims that a trustee stopped thinking. The pleading writes itself: the trustee adopted a system it did not understand, adopted its output without consideration, and thereby failed to give real and genuine consideration to the exercise of its discretion. That is a claim under the general law, and it does not require the court to form any view about machine learning at all.
The related risk is at the other end: a trustee that used a model, obtained a result it did not intend, and now wishes to unwind it. The mistake jurisdiction is available in principle, but the distinction between a mistake as to the effect of a transaction and a mistake as to its consequences will be doing a great deal of work. A trustee that did not read the output is not obviously mistaken about it.
A note on the same problem elsewhere
None of this is peculiar to trusts. Our arbitration partners have written about the tribunal-appointed expert whose method is a proprietary model, and about the difficulty of testing an opinion that cannot be reproduced. The regulatory practice sees the same question in the guise of an automated advice suitability check. The common feature is that a duty framed in terms of a person’s reasoning is being discharged by a process that does not reason in a way anyone can examine.
The law is not, in our view, in need of reform for this. It needs to be applied with its usual indifference to novelty. A trustee that delegates to a model is doing what trustees have always done, and will be judged as trustees have always been judged: on what it decided to delegate, what it wrote down, and whether it kept looking.