Who Should Lead Wealth Management AI, and What Does That Person Actually Own?

An Advisor Just Emailed a Model-Written Portfolio Summary to a Client. Whose Problem Is That?

Right now it is nobody's, which is the problem. The copilot vendor points at the advisor. The advisor points at the tool. Compliance finds out during a routine review, three weeks after the email went out, and has no way to reconstruct what the model saw or what the advisor changed. A wealth management AI lead exists so that this moment has an owner before it happens rather than after.

The seat is narrower and more concrete than a general AI leadership job. It owns the boundary between drafting and advising: what the copilot may generate unsupervised, what a licensed human must approve before it reaches a client, and where the line moves as accuracy improves. It owns the data contract, meaning which systems the model may read from and whether a generated figure is reconciled against the book of record or merely plausible. It owns the evaluation set, which for this domain means a bank of real client situations with known-correct answers rather than a general helpfulness benchmark. And it owns the record: what version was running, what it produced, who reviewed it.

The reason it belongs inside the business line rather than in a central AI function is accountability. Advice given to a retail investor is a supervised act in most jurisdictions, and the person who answers for it has to be someone the business already trusts to make that call. Placement follows: JPMorganChase lists a vice president role titled CAO Artificial Intelligence Lead inside J.P. Morgan Wealth Management, along with a cluster of AI-focused vice president and executive director roles in wealth growth and sales optimization, and Goldman Sachs lists an applied AI and business intelligence vice president inside Wealth Management operations ¹. That is one job-board sweep on one September morning rather than a labor survey, and postings come down, so what it supports is a direction and not a market rate. The direction is consistent: two large firms, both putting the seat in the wealth org chart rather than borrowing it from the platform team.

One scoping note before writing the description. If the harder problem is model risk, controls and evidence rather than product decisions about client experience, the job you are describing is closer to an AI control and oversight researcher and should not be quietly merged with a product mandate. Regulatory expectations for AI in advice differ by jurisdiction and are moving quickly, so treat the supervisory boundary as a question for your compliance counsel rather than something the job description settles.

What Should You Listen For When a Candidate Explains a Generated Number?

Hand the candidate the email from the top of this piece and ask what should have happened before it left the building. The answer tells you most of what you need in ten minutes, because it forces a choice between talking about what the copilot can do and talking about what it is allowed to say, and only one of those is the job.

Listen for the order in which they take the problem apart. Strong candidates start with permission and arrive at capability later. The fluent-but-untested run the other way, opening with agentic workflows, personalization at scale and next-best-action, having never once had to explain to a supervisory principal why a generated sentence was fine to send.

Then ask how the client's year-to-date return should have reached that summary. The answer you want routes the figure out of the system of record and forbids the model from producing it at all. The answer that should worry you is a confidence claim about hallucination rates, which is a way of saying the number is probably right. Ask next what they would put in an evaluation set, and expect something boring: a few hundred annotated cases drawn from real advisor questions, scored by people who know the correct answer. Vague talk about benchmarks means nobody skeptical has ever made them prove a system was safe.

Ask what they killed. A candidate who has run AI inside a regulated business can name a thing they refused to ship and say what refusing cost them politically. Someone who has only prototyped cannot, because prototypes have no politics. Listen too for how advisors appear in their sentences. When the field shows up as a change-management obstacle rather than as users with a job, the likely outcome is a beautifully governed tool nobody opens.

One anti-tell to name explicitly. A candidate who proposes detecting whether a document or a client message was written by AI is offering a capability that does not reliably exist, and building supervision on top of it is a bad foundation. The durable version of the control is the opposite: assume AI is in the workflow, define what a human must verify, and keep the evidence of that verification.

Hire the Supervision Analyst Who Would Have Caught That Email

Somebody in your firm reads advisor communications all day and already knows which sentence in a client letter creates an obligation and which is decoration. That person is a supervision or surveillance analyst, they transfer into this seat unusually well, and almost nobody recruits them. The portfolio summary at the top of this piece is what they catch for a living, three weeks late, which is the timing this seat exists to fix.

That judgment is precisely what an evaluation set encodes, which is why the transfer works. The obvious feeders are wealth platform product managers and advisory practice leaders who have already run a technology rollout to a field force. The less obvious ones are better than their resumes suggest. Investment operations leads who have reconciled a book of record for years bring the second habit worth paying for, which is treating a number as unreconciled until it ties to something; a person with that reflex will not let a generated figure onto a client-facing surface without a source behind it. Financial planning specialists who have built the models advisors actually use with clients bring the other half, which is knowing what the client asks next.

How the good ones got good is unglamorous and easy to check. They used an assistant on their own work, at volume, and kept score. The specific answers that mean something: drafting a hundred client review summaries with a model and cataloguing the four ways it went wrong; asking it to compute a withdrawal-rate scenario and finding the arithmetic error by hand; building a small retrieval setup over the firm's own product documentation and discovering that the model answered confidently from the wrong document. Someone who has done that talks about model failure in categories rather than as a general caveat. The candidates who have never used these tools fear the wrong things, usually tone and phrasing, and miss the real one, which is a fluent answer resting on stale or mismatched data. The ones who trust the output fail more expensively. What you want sits in between and can describe the checking in detail without being asked, which is the same disposition that makes an AI finance strategist useful in an adjacent seat.

Searching for the title will not find any of them. It is roughly two years old and nowhere near settled, so the people who can do this job are currently carrying other names: wealth platform product lead, advisor experience product manager, head of advisory technology, applied AI lead, business intelligence and applied AI vice president. JPMorganChase's own posting for the adjacent product seat is titled AI Product Owner, Financial Insights, at senior associate level ¹, which is the same work at a smaller scope.

Supply is concentrated in four places. Large wirehouses and private banks are the only employers that have run advisor-facing AI at scale, which is why the visible postings cluster there. Custodial and advisor-platform technology firms have product managers who have shipped to thousands of independent advisors and know the compliance boundary cold. Wealth technology vendors building planning, proposal and portfolio software employ people who have argued about generated content with a hundred compliance teams. And the RIA aggregators have operations leaders who have standardized advice processes across dozens of acquired firms, which is the same problem in a different costume.

Look internally before you post. The person who has been informally running the copilot pilot alongside their day job usually already exists, knows where the data actually lives, and can be promoted faster than an outside search closes; backfill behind them. Screen on written artifacts rather than on interviews alone: a rollout plan, a supervisory procedure, a product requirements document for something client-facing. This job is largely writing that other people have to follow, and the samples are honest in a way a conversation about AI strategy is not. The shortcut to avoid is hiring a strong machine learning engineer and asking them to pick up the regulatory context. That version stalls at pilot, because the blocking constraints are not technical and the engineer has no standing to negotiate them.

Expect to Pay Against Your Wealth Product Bands, Not an AI Premium

Close on scope and on the reporting line, in that order. The candidates worth hiring have watched an AI initiative get reduced to a demo, so the offer conversation should name what they can decide alone, what needs the business head, and who they escalate to when compliance and the field disagree. A seat that can recommend but not decide will be vacant again within a year.

On pay, the honest answer is qualitative, because no government wage series covers this title and there is no credible published band for it yet. What the visible postings show is which existing band it hires against: at JPMorganChase and Goldman Sachs the leadership versions are posted at vice president and executive director level inside the wealth business, and the narrower product-owner variant at senior associate ¹. That is a handful of postings from one sweep, not a survey, so read it as placement evidence rather than as a rate. It means the seat pays against your existing wealth product and business leadership bands rather than against a separate AI premium, with the AI component argued as a scarcity adjustment rather than a new scale. There is macro support for some adjustment: PwC's 2026 AI Jobs Barometer, analyzing close to a billion job advertisements, reports an average wage premium of about 62 percent for roles requiring AI skills ². Treat that as a reason to expect upward pressure, not as a number to put in an offer.

What kills offers here is predictable. A reporting line into central technology, which tells the candidate the business does not own the outcome. A mandate that turns out to mean vendor management. No budget for evaluation work, which is the unglamorous half of the job and the half that makes the rest defensible.

On location, expect on-site or strongly hybrid, and expect that to be non-negotiable at a bank. Three parts of the work genuinely need proximity: sitting behind advisors while they use the tool, which is how a summary like the one at the top of this piece surfaces before a client receives it rather than three weeks after; the supervisory and legal conversations that decide the boundary; and access to client data, which at most firms is restricted to managed environments in named locations. The independent and vendor side of the market is more flexible, and a hybrid arrangement with regular time in a branch or on a desk is the version that works in both.

See the benchmarks >