The strongest near-term AI opportunity in wealth management is not replacing the relationship manager. It is giving the relationship manager better leverage.
Clients do not come to a private bank only for information. They bring context, uncertainty, family dynamics, preferences and decisions that do not fit neatly into a prompt. Trust depends on judgement and accountability.
At the same time, relationship managers spend significant time preparing, searching, documenting and coordinating. Those activities are necessary, but they compete with the work where a human creates the most value: understanding the client, framing trade-offs and maintaining the relationship.
AI can change that balance.
Start with capacity, not substitution
McKinsey’s “The looming advisor shortage in US wealth management” (10 February 2025) estimates that the US industry could face a shortage of roughly 100,000 advisors by 2034 at current productivity levels. The geography and market structure are different from Swiss private banking, so the number should not be transferred directly. The underlying pressure is still relevant: demand for advice is rising while experienced human capacity is scarce.
That makes productivity a strategic issue.
The wrong objective would be “automate the advisor.” The useful objective is “increase the amount of high-quality client work each advisor can perform without weakening trust or controls.”
This framing leads to a better use-case map:
- prepare for meetings;
- structure notes;
- draft follow-up actions;
- retrieve product and mandate knowledge;
- synthesise research;
- detect relevant client or market events;
- prepare communication for review;
- coordinate operational follow-through.
These tasks surround the relationship. They do not replace it.
Meeting preparation is a platform use case
Meeting preparation is often presented as a simple summarisation problem. It is more interesting than that.
A useful brief may need recent interactions, portfolio changes, open service items, investment context, relevant product information and client preferences. The sources have different owners and levels of sensitivity. Some are structured. Others are documents or notes. Some may be stale or incomplete.
The product therefore needs more than a good model. It needs identity, access, retrieval, source provenance, data-quality handling, a clear interface and user feedback. It also needs agreement about what should never be inferred.
That makes meeting preparation a strong early platform use case. It can create reusable components for client context, knowledge retrieval, citation and review while delivering visible user value.
A sensible first version is bounded. One relationship-manager group, a limited set of approved sources, a standard brief format and an explicit feedback step. The team should compare preparation time and usefulness with the current workflow, while reviewing errors and omissions.
The output should make its evidence visible. A relationship manager needs to know whether a statement came from the CRM, a portfolio system, an approved publication or model inference.
Responsiveness needs a human owner
AI can help draft a follow-up email, turn meeting notes into actions or prepare a response to a product question. Speed matters, but client communication has consequences.
The relationship manager should remain the owner of the final message. The product can reduce blank-page effort, maintain tone and surface relevant context. It should not create an illusion that a client-specific recommendation has been reviewed when it has not.
This is where interface design becomes a control. Source links, highlighted uncertainties, clear draft status and an easy editing experience make human review more effective. A generic disclaimer at the bottom does not.
For certain communication, an additional review step may be appropriate. The control should reflect content, client, channel and action rather than a single rule for every draft.
Knowledge access can reduce avoidable friction
Relationship managers navigate a large body of product, mandate, policy and process knowledge. Search quality varies, terminology differs and the right answer may depend on jurisdiction or client segment.
A governed knowledge assistant can provide leverage when it retrieves only approved sources, cites them clearly and handles ambiguity well. The product should say when it cannot find sufficient evidence. It should not improvise around missing policy.
The knowledge operating model matters as much as the model. Content needs an owner, an effective date and a route for correction. User feedback should reach the team responsible for the source, not disappear into a generic thumbs-down score.
McKinsey’s January 2024 report on US wealth management describes generative AI applications such as meeting-note synthesis, client briefs and virtual assistance, while also emphasising risk, compliance and change management. That combination is important. The technology opportunity and the operating-model work are inseparable.
Next best conversation is not next best action
Predictive and generative AI can help surface a relevant conversation: a liquidity event, a portfolio concentration, a change in client circumstances or a piece of research connected to known interests.
This becomes risky when a suggestion is treated as a recommendation without context. A detected event may be incorrect. A product opportunity may be unsuitable. A client may have expressed a preference that is not visible in the data.
I would design this category as decision support. The system surfaces evidence and explains why it may be relevant. The relationship manager decides whether and how to act.
The quality measure should include more than click-through or contact volume. False positives, user dismissal reasons, client relevance and control outcomes matter. An apparently productive alert system can create noise and damage trust.
Adoption depends on workflow fit
Relationship managers already work across several systems. Adding another destination is rarely a strong adoption strategy.
AI capabilities should appear at the right moment: a brief available before the meeting, suggested actions inside the follow-up path, knowledge retrieval within the existing workspace. The product should reduce switching and duplicate entry.
Training should use realistic scenarios. Users need to understand where information comes from, what the product can and cannot do, how to correct it and which tasks still require judgement.
Local champions can help, particularly across countries and business units. They translate the product into local workflows and give the central team evidence about differences that matter. This is not a substitute for product analytics and support. It is a bridge between them.
Deploying an AI tool is not the same as changing how work gets done.
Measure leverage without losing quality
Time saved is useful, but it is not sufficient.
For meeting preparation, I would examine active use, preparation time, completeness, source accuracy and user assessment. For follow-up, I would look at completion speed, edit patterns, missed actions and inappropriate suggestions. For a knowledge assistant, retrieval success, evidence quality, unanswered questions and content gaps matter.
Where possible, the team should also understand whether capacity moves toward more client activity or simply disappears into other administration. The desired outcome is better leverage, not a superficial productivity claim.
McKinsey’s 2026 view of wealth management in 2035 places AI alongside demographic change and evolving client trust. That is the right context. AI will influence the service model, but trust remains a competitive asset.
The near-term winner is likely to be the institution that combines strong human advice with a quietly excellent AI-enabled workflow. The client may never see the platform. They will experience a relationship manager who is better prepared, more responsive and able to spend more attention on the conversation.
Three concrete takeaways
- Frame relationship-manager AI around capacity and service quality, not replacement.
- Keep client-facing judgement with the accountable human and make evidence visible in the interface.
- Measure workflow adoption, quality and reclaimed client capacity—not availability alone.
The views expressed here are personal and do not represent those of my current or previous employers.