Part three of a series. Part one covered why you ask an LLM for the analytical framework rather than the answer. Part two built a qualified prospect list out of primary source filings with every number traceable. This post is about what you do with that list, on repeat, without ending up in a compliance file or a screenshot on social media.
Stop running it on a calendar
The first instinct with any recurring workflow is to pick a cadence. Run it every two weeks. Run it monthly.
For this population that is the wrong architecture, and the reason is population math. There are only a few hundred thousand households worldwide above the ultra tier, and one estimate puts UHNW households in the United States at roughly 250,000. That universe barely changes in two weeks. Re-running discovery on a static population on a fixed schedule burns money to produce the same list.
What changes in two weeks is events. So the loop should be event-triggered, with the calendar reserved for re-scoring what you already have.
The triggers that matter are the same primary sources from part two, watched rather than queried:
- A Form 4 showing a large insider sale, which is liquidity that just landed
- An S-1 filing, which puts a lockup expiry on the calendar 180 days out
- An 8-K on a completed acquisition, which tells you whether the consideration was cash
- A Form D, which shows private placement activity
- A property transfer in county records
- An executive or board appointment, a company sale, a retirement announcement
Each of those is a document with a date. A watcher on the sources fires the loop for one person, at the moment the event makes them relevant, rather than putting everybody through the same cycle every fourteen days.
The right cadence for the calendar part is quarterly, re-running the wealth and liquidity computation over the existing list to catch drift. That is a re-scoring job, not a discovery job, and it is cheap.
The personalization problem, stated honestly
Here is where most write-ups on this topic go wrong, including my own first pass at it.
The obvious idea is that you learn things about a prospect from their public posts and use those things to personalize the approach. They posted a photo from a vacation in Aruba, so you send a spa gift card. That feels thoughtful.
It is not thoughtful. It is the tell.
Put yourself on the receiving end. A firm you have never done business with sends you a gift that maps precisely onto something you posted last week. You will do that arithmetic in about two seconds, and the thought that arrives is not “how considerate,” it is “these people are reading my vacation photos.” You have converted an interest signal into a surveillance signal, and there is no recovering the relationship from there.
So here is the rule I would put at the center of the design:
An observed signal may inform the category and the timing of an outreach. It must never be specific enough to trace back to a single post.
And the test that operationalizes it: if they asked how you knew that, would your answer be comfortable?
“You mentioned it when we met” is comfortable. “It’s in your firm bio” is comfortable. “You were quoted in the trade press about it” is comfortable. “We saw the photo” is not, and no amount of softening the copy fixes it, because the specificity itself is the disclosure.
This is a place where an LLM is genuinely well suited, as long as you use it in the right mode. The task is deliberate abstraction. Given a set of observed signals, produce a broad interest category with the underlying artifacts stripped out and not passed downstream. Wellness and travel, not the Aruba photo. Classic cars, not the auction the person attended in March. Then the downstream system only ever sees the category, so it cannot leak the source even by accident.
That is a judgment call with a fuzzy boundary, which makes it exactly the kind of step that should carry a human review gate before anything is sent.
Observed signals
Public bio, press quotes, conference talks, board seats, published interviews.
Abstraction
The model generalizes those signals to a category and discards the source artifact. Wellness and travel, not the photo. Classic cars, not the auction.
Human review
Apply the test: if they asked how you knew that, would the answer be comfortable?
Match
Match the category to an event you already host.
Compliance gate
Gift policy, marketing rule, suppression list, and the archive. An invitation, not a gift.
Send
Only what passed the gate and the review.
Invitations scale, gifts don’t
Once you accept the rule above, the whole gifting idea reorganizes itself, and honestly it gets better.
A physical gift to someone you have never met is a hard problem. It is either generic, in which case it does nothing, or it is specific, in which case it is the surveillance signal. There is not much room in between.
An invitation does not have this problem, because an invitation is something you were doing anyway and the prospect is one of the people you thought of. Nobody finds it strange to be invited to something.
The versions that work, roughly in order of how well they hold up:
- A small dinner with a genuinely interesting speaker, ten or twelve people
- A year-end tax planning briefing, timed to when the question is actually on their mind
- A family governance or next-generation workshop, which is the single most under-served need in this band and almost never sold well
- A liquidity-event planning session, triggered by a filing, framed around the decision they are actually facing
- A wine tasting, an art market briefing, a collector’s evening, anything where the interest category matched the theme and the theme is broad enough that the match is invisible
- A curated introduction to someone useful, which costs nothing and outperforms everything else on this list
Notice what all of these have in common. The signal you learned changed which list you put someone on. It did not appear in the message.
The gift rules, since they are real
For direct gifts to prospective clients, the constraint is mostly your own firm’s policy and the fiduciary standard rather than a specific dollar figure, because the SEC does not set a prescribed limit for advisers. Most firms adopt one anyway as part of their code of ethics, typically with pre-clearance and reporting.
For the referral channel, which is the next section, FINRA Rule 3220 applies directly, and the limit moved from $100 to $300 per person per year effective March 30, 2026, the first increase since 1992. That rule is about gifts related to the business of the recipient’s employer, which is precisely what a referral relationship with a CPA firm is.
One consequence worth noticing: the compliant, cheaper channel is the one almost nobody automates.
Target the CPAs, not the millionaires
The uncomfortable truth about this population is that acquisition is overwhelmingly relationship-driven. The households in this band already have advisors. They do not respond to outbound. They move when someone they trust tells them to move, and that someone is usually their CPA, their estate attorney, or the banker who ran their liquidity event.
Which means the highest-yield application of everything in this series is aimed one step to the side of where you thought.
Build the same pipeline around the professional network surrounding your target households. Who filed as counsel on the transaction. Who is the accountant of record on the entity. Which attorneys appear repeatedly around the estates and trusts in your geography. That work is B2B, which is the easy case: Lusha, Apollo and Clay have this data, the FINRA gift limit applies cleanly, and there is no wealth screening problem at all because you are researching professionals in their professional capacity.
Then cultivate it with the things those professionals actually want, which are not fruit baskets. Continuing education content that counts toward their requirements. Co-hosted client seminars where they get the podium. Technical briefings on the parts of your world they get asked about and can’t answer. Reciprocal referrals, sent first.
An agentic system runs this well because the work is genuinely tedious and genuinely repetitive: watching filings for which professionals appear around which transactions, keeping the relationship map current, noticing that a firm you have a relationship with just picked up a client who fits your band, and surfacing the moment where an introduction would be natural.
Measuring it, which is where most of these die
I am not going to hand you a target number, because anyone who does is selling you something. But the choice of metric is a design decision and it belongs in the design conversation, not the quarterly review.
The trade-offs, honestly:
| Metric | What it is worth |
|---|---|
| Response rate | Cheap, and nearly worthless. Volume inflates it, and it says nothing about whether the responses were good. |
| Meetings booked per hundred qualified prospects | Better, because it sits downstream of quality. It can still look healthy while nothing closes. |
| Qualified opportunities | A household that fits the band and has a reason to move. The best leading indicator here, and the easiest definition to let drift. |
| Assets in pipeline and assets won | What the business cares about. Nearly useless for steering: the cycle runs twelve to twenty-four months, and attribution is genuinely hard. |
| Cost per qualified conversation | The number I would put on the wall. It says whether the automation is earning its cost. |
| Complaint and suppression rate | How many people asked not to be contacted, reacted badly, or flagged something as inappropriate. In a referral business this can end the program, and it moves before the revenue number does. |
Instrument the guardrail metric from day one. Retrofitting it after an incident is not the same thing.
The loop, end to end
Watch
Filings, records, and appointments. An event fires for one person.
Recompute
Wealth and liquidity, from the documents.
Abstract
Signals become a broad category. The artifacts are discarded.
Compliance gate
If it fails, hold the record and log the reason.
Human review
If the case is unsure, a person decides.
Route
An invitation, a call, or an introduction.
Archive
Every message is retained.
Measure
Include the complaint and suppression rate. That result feeds back into scoring.
Quarterly re-scoring on the existing list. Event triggers for everything else. A compliance gate that nothing gets past. A human on the ambiguous cases only. And every outbound message archived, because an agentic system that drafts and sends communications is generating records subject to retention and supervision, and building that in afterwards is much harder than building it in first.
That is the build. Event-triggered, not calendar-driven. Personalization that never traces to a single post. Invitations over gifts. Then the machine around the professionals those households already trust, with the complaint rate on the wall from day one.
This is the last post in the case study.