Where AI earns its place in an advice firm, and where it does not

Most conversations about AI in advice firms start with which tool to buy, and end in disappointment. Anthony Marris proposes a different filter: how cheaply can a human check the output. Where a person can verify a draft in seconds because they already hold the source, the tool saves real time. Where verifying the answer means redoing the work, it saves none and adds risk. The piece applies that test to eight jobs in a typical planning business, and explains why the structured, boring integration work still matters more than the model.

The question I am asked most often about AI in an advice firm is which tool to buy. It is a reasonable question and it produces poor answers, because it starts in the wrong place. The tools are converging and most of them are competent. What varies is the job you point them at.

The filter I have ended up using is a single question: how cheaply can a human check the output.

Cheap to check, expensive to check

Some outputs come with their verification attached. An adviser who has just spent an hour with a client can read a drafted meeting note in ninety seconds and know whether it is right, because the source is in their head. A paraplanner comparing a summary against the factsheet it was drawn from can check it line by line in the time it takes to read.

Other outputs have no such shortcut. Ask a model to infer a client's attitude to risk from six years of correspondence and you get a confident paragraph. To trust it you would have to read the six years, which is precisely the work you were trying to avoid. The output looks like a saving and is not one.

That distinction does more sorting than any feature comparison.

Applying it

Here is the same test run across eight jobs that come up in most planning businesses.

Job	Cheap to check	Why
Meeting note from a recording	Yes	The adviser was in the room
Summary of a factsheet or report	Yes	The source sits beside the summary
First draft of a client letter	Yes	A human reads it before it is sent, as they always did
Tidying a fact find into prose	Yes	The fields are in front of you
Explaining a technical point in plain English	Yes	Somebody in the firm already knows the answer
Inferring risk tolerance from past correspondence	No	Verifying means reading everything
Picking holdings, or ranking them	No	Verifying means doing the research, and this is regulated ground
Extracting figures from a scanned statement	Partly	Checkable, but only if every figure is checked, and nobody does

The last row is the one that catches firms out. Extraction feels like the obvious win, and it is genuinely useful, but the error rate is not zero and the failure is silent. A wrong number in a valuation does not announce itself. If the process does not include somebody reconciling the total against the statement, the time saved has been converted into risk rather than capacity.

The unglamorous work still matters more

The single biggest drag on a paraplanner's week is not the absence of a clever model. It is that the same client data is typed into three or four systems by hand, because those systems do not talk to each other. The 2023 Scottish Widows paraplanner survey found 83% of UK paraplanners thought their technology could be improved, up from 64% in the survey before, and rekeying is the friction they name most.

No model fixes that. An integration does. Getting the back office, the research tool, the cashflow tool and the report writer to agree on one client record is dull, unfashionable work, and it returns more hours per week than anything else on the list above.

There is also an order of operations here. A model working on data somebody typed in three times inherits all three versions. Fixing the plumbing first makes everything downstream more reliable, including the AI.

What I would actually do

If a firm asked me where to start, I would say: pick one job where a human already holds the answer, put a tool on it, and measure the minutes. Meeting notes are usually the right first choice, because the check is free and the saving is immediate.

Then leave the regulated judgements alone. Not because the technology cannot form a view, but because a firm cannot evidence a view it did not form, and evidence is the thing being asked for.

The firms getting the most out of this are not the ones with the longest tool list. They are the ones who worked out which of their own jobs were cheap to check.

Every Wealth Analytica article is fact-checked against primary sources where applicable. Read our editorial policy for our sourcing and review standards.