The hardest question about AI in wealth management is not whether it works. It is how you would know. Most AI tools in this industry produce activity (summaries, transcripts, notifications) and activity is easy to demo and impossible to manage by. If reading the firm's communications is going to justify a line in next year's budget, it has to move numbers a desk head already reports: retention, share of wallet, cost of compliance. This post is a practical frame for getting from one to the other, and for setting the measurement up honestly before any tool is switched on.
Start from where the value actually sits
Roughly 80% of what a firm knows about its clients sits in unstructured communications; the structured 20% is what the systems of record hold. Three outcome streams run through that unstructured layer:
Retention. Departures are sequences, legible in the mailbox long before they are visible in AUM: engagement declining, replies shortening, a review declined twice, a competitor named. In one documented sequence, 141 days separated the first legible signal from the transfer request. Every day of that window that becomes visible is retention capacity.
Revenue. Liquidity events, successions and inheritances get said in passing and lost. Surfacing them while they are actionable is the growth case; by our published analysis, acting on relationship signals earlier carries a potential AUM uplift in the 4 to 6% range. Treat that as the size of the prize to be tested, not a promise.
Compliance cost. Suitability evidence gets reconstructed after the fact, and by our analysis 40 to 50% of the compliance task load per relationship manager is automatable documentation work. Evidence written while the work happens converts that time back into client time.
The KPI ladder: from signal to number
The way to make any of this measurable is to insist that every layer resolves into the one above it:
Layer | What it is | The KPI it feeds |
|---|---|---|
Source | The firm's own communications, read with consent, identifiers removed | Coverage: share of the book actually being read |
Signal | A ranked recommendation with evidence attached | Signal precision: acted-on rate, false-positive rate |
Decision | What the relationship manager did, recorded | Time to action on at-risk and opportunity Signals |
Outcome | What moved | Retention of flagged relationships, wallet growth on surfaced opportunities, hours of documentation removed |
Two disciplines keep the ladder honest. Every number must trace downward: a retention figure that cannot be traced to specific Signals and decisions is a coincidence wearing a KPI's clothes. And false positives are a first-class metric, not an embarrassment: in a Luscent pilot the false-positive rate is triaged weekly with the compliance function, because a signal feed the desk stops trusting produces no outcomes at all.
Set the metrics before day one, not after
The single most important implementation decision costs nothing: agree the success metrics in writing before anything is connected. A 90-day evaluation has a natural shape. Connect read-only, on synthetic data first, so the firm sees what the system does with invented clients before any real correspondence enters. Rank the book and test the ranking against what the desk already believes: the first conversation should be about where the system agrees and disagrees with experienced judgement. Then run the loop on live Signals: one Signal, one decision, one record, with the false-positive triage running weekly. At day 90, the firm holds a decision on the record rather than an impression.
What this deliberately does not include is a revenue promise. Any vendor quoting a euro figure your firm will earn per relationship manager, before your book has been read, is quoting someone else's number or no one's. The honest sequence is the reverse: agree the metrics, run the 90 days, and let your own book produce the figure that goes to the budget committee.
What compounds after the pilot
The KPI frame above is deliberately conservative because month one is the system's weakest month. A relationship representation built from a firm's own communications accumulates: month six is worth considerably more than month one, and no amount of capital shortens that for anyone. Which is exactly why the measurement discipline matters at the start: the firms that set honest baselines in the first 90 days are the ones that can prove the compounding later, to their boards and, increasingly, to their regulators.
Luscent is the system of intelligence for wealth management: an EU-native AI platform that reads client communications to surface relationship insights and generate compliance evidence, with the relationship manager deciding. The pilot programme is 90 days, free, for three to five relationship managers, starting on synthetic data; the Q4 2026 cohort closes on 30 September 2026. This post describes a measurement approach and published capabilities; it is not advice, and figures described as potential are sizes of opportunity to be tested, not promised outcomes.
