Finding Wealth Management Prospects Without Making Anything Up

This is part two of a series. Part one made the case that you should not ask an LLM for the analytical recommendation, you should ask it to find real data and write the code that analyzes it. This post runs that all the way through on a hard problem.

The problem: a wealth management firm wants new clients in a defined wealth band. Find them, qualify them, and do it in a way that survives a compliance review.

The naive version, and why it fails

“Find me ten high net worth prospects for my wealth management firm in the tri-state area.”

You will get ten names in four seconds. Some of them will be real people. The net worth figures attached to them will be invented or half-remembered. There will be no source for anything. And you have now created a document, inside a regulated firm, that asserts financial facts about named private individuals with no evidence behind any of it.

That is not a lead list. That is a discovery risk with formatting.

Worse, it fails on its own terms, because it never asks the two questions that determine whether any of these people are prospects at all.

Question one: which band, actually

The industry does not segment on “high net worth.” It segments on tiers, and the tiers are commercial conventions rather than legal definitions, which means you have to say which one you mean.

The common cuts are commercial conventions, not legal definitions. Capgemini’s World Wealth Report has anchored the $1 million line since 1997, measured on investable assets excluding the primary residence. Wealth-X starts its ultra-high-net-worth census at $30 million.

The service model changes at each line, and so does who will even talk to you.

Separately the SEC ignores the marketing labels entirely and gates on its own tests, so accredited investor status under Rule 501(a) and qualified purchaser status at $5 million in investments determine what you are permitted to offer, regardless of what tier your brochure calls them.

Now apply this. A firm that says it wants clients between $25 million and $250 million has described two different businesses. At $25 million you are pitching a private bank or VHNW advisory relationship to someone who is probably still running the company that made them wealthy. At $250 million the household already has a family office, a chief investment officer, and an RFP process, and will never respond to an approach from outside its existing network.

The first useful thing an analytical system does is tell you your band is wrong and show you the population math behind that conclusion.

Question two: net worth or investable assets

This is the one that separates people who know the business from people who don’t.

Net worth and investable assets are not the same number and the gap between them is where deals die. A founder with $40 million of net worth held almost entirely in illiquid private company stock is not a $40 million wealth management prospect. There is nothing to manage. A retired executive with $12 million in a diversified brokerage account is a live prospect today.

So the real target is not a wealth number. It is liquid or soon-to-be-liquid assets, and the date they become liquid.

An agent can genuinely help here, and here is how, because this is where the framework earns its keep. Rather than asking a model to estimate someone’s wealth, you have it construct a per-person balance sheet where every line is computed from a primary source document, along with the date each line becomes investable.

• Insider holdings in public companies are computable. Form 3, 4 and 5 filings give share counts and transaction prices. Schedule 13D and 13G give larger positions. Multiply by the current price and you have a real number with a filing behind it.
• Liquidity timing is computable. An S-1 tells you a company is going public and the lockup terms tell you when insiders can sell, typically 180 days. An 8-K on a completed acquisition tells you whether consideration was cash or stock and when it closed. A Form D tells you about private placements.
• Real property is computable from county records. Aircraft from the FAA registry, vessels from the Coast Guard registry, business ownership from state registrations.

Every one of those is a public filing with a URL and a date. The number is not the model’s opinion. It is arithmetic over documents, and anyone can recompute it.

Where the model does the creative work is in deciding where to look, what to include, and how to handle the ambiguous structures. Where the code does the work is the arithmetic. Where the human does the work is the cases the system flagged as uncertain.

Getting the vendor stack right

A common mistake, and one I made in my own first sketch of this, is reaching for the tools you already know. None of the B2B contact tools carry individual net worth, because that is a different industry.

Use the B2B tools for the referral network, which I will come back to in part three, and the wealth screening tools for the households.

The paths you cannot go down

This is the section most write-ups skip, and it is the section your compliance officer will read first.

Credit data is a wall, not a hurdle. Marketing is not a permissible purpose under the Fair Credit Reporting Act. The FTC established this against TransUnion over the sale of target marketing lists built from credit information, on the basis that those lists were consumer reports and marketing is not among the permitted uses. Wealth screening vendors stay clear of this by not sourcing from bureaus. Your pipeline has to stay clear the same way. The risk is not just buying the wrong data. It is that if your system computes something that functions as an eligibility or capacity score out of covered data, you may have walked into the statute without buying anything at all.

Scraping is a contractual and legal question, not a technical one. Whether you can extract content from a platform is governed by that platform’s terms and by an active line of litigation, and the answer is not the same for public pages and for content behind a login. A screenshot taken by an automated browser session is still collection. Using a vision model to read the screenshot does not change what happened, it just changes who wrote it down.

Once someone is a customer, a different regime applies. Gramm-Leach-Bliley and Regulation S-P govern nonpublic personal information, and the amended Reg S-P adds service provider oversight and incident response obligations that reach every vendor and contractor touching that data.

State privacy law applies before that. Roughly twenty states now have comprehensive privacy statutes with their own definitions, notice requirements, and opt-out rights.

The legal pathways that remain are wide enough to build on: public regulatory filings, public government records, licensed non-FCRA wealth data under its permitted use terms, professional information the person published about themselves, and first-party data the person gave you directly.

Where AI helps most here is not in finding clever ways around the rules. It is in encoding the rules as a gate that every record passes through before a human ever sees it, with the reason for every hold written down. Which source did this fact come from, is that source on the permitted list, does this record’s jurisdiction change the answer, is this person on a suppression list, has the retention window expired. Those are exactly the deterministic checks the model should never be improvising, and exactly the kind of tedious universal screening that is too expensive to do by hand and trivial to do in code.

What comes out the other end

A qualified list where every record carries the wealth estimate, the liquidity date, the source URL for each component, the retrieval date, the confidence band, and the compliance gate result.

You can hand that to a regulator. You cannot hand them ten names an LLM produced in four seconds.

Part three covers what to do with the list: event triggers instead of calendar cadence, outreach that doesn’t read as surveillance, the referral channel that is legally easier and works better, and how to measure whether any of it is working.