The editors are pleased to share these insights; the customary editorial disclaimers apply to views of guest writers. To comment, email tom.burroughes@wealthbriefing.com and amanda.cheesley@clearviewpublishing.com

Dr Gleb Tsipursky

Asia’s wealth industry is moving quickly from AI assistants that summarise information to systems that can recommend, prepare, and increasingly act across a workflow.

The change is visible in Singapore, where HSBC said in late July that it will build a Global AI Centre of Excellence with more than 100 AI specialists, initially focused in part on customer conversations about wealth.

That is a useful signal for the wider industry. The next competitive question is not whether a wealth manager can give advisors more AI. It is whether the firm can draw a clear decision boundary around what the technology may prepare, what a professional must validate, and what must be escalated before a client is affected.

Artificial intelligence and digitalisation of the wealth-management value chain are major focus areas. The highest-value uses of AI now sit inside a chain of judgment: research becomes an insight, an insight becomes a recommendation, a recommendation becomes a client conversation, and that conversation may become a transaction.

Each handoff changes the risk.

A model can summarise a company announcement with relatively low consequence if the output stays inside an analyst’s research notes. The same summary carries more weight when it appears in a client briefing. It carries more again when it influences an asset-allocation recommendation. The technology may be identical; the decision context is not.

Singapore investors already seem to understand this intuitively. An HSBC survey of more than 600 mass-affluent and high net worth investors found that 76 per cent use AI for finance and investment, while still preferring advisor validation at the point of decision. That combination should shape how firms design AI-enabled advice. Clients may welcome faster analysis and more personalised preparation without wanting accountability transferred to a machine.

A practical response is to create an AI decision boundary for every consequential workflow.

The boundary has three zones.

The first is the preparation zone. AI can gather permitted information, summarise source material, compare scenarios, draft meeting notes, surface questions, or prepare a first version of a client explanation. The output is useful, but it is not yet a decision.

The second is the validation zone. A named professional checks the facts, source quality, assumptions, suitability context, and whether the output answers the client’s circumstances. This is where the advisor’s judgment matters most. Validation should not mean glancing at a polished paragraph and clicking approve. It should mean checking the evidence behind any claim that could change a recommendation.

The third is the escalation zone. Certain outputs should trigger a higher level of review because the potential cost of error is larger. Examples include an unusual concentration risk, a conflict between client objectives and a proposed action, a recommendation based on incomplete data, a material tax or legal implication, or an automated action that could move money. The value of these zones is that they convert a vague instruction to “keep a human in the loop” into an operating rule.

That matters as agentic systems arrive. A traditional assistant waits for a prompt and returns an answer. An agent may chain several steps together: retrieve information, decide which source to use, draft an analysis, select a next action, and hand the result to another system. The more steps are connected, the easier it becomes for a small error early in the chain to look authoritative by the end.

A decision boundary interrupts that compounding effect.

Firms should also record why a human changed or stopped an AI-generated recommendation. The point is not to create a compliance diary for every prompt. It is to identify recurring categories of intervention: stale data, missing client context, unsupported inference, unsuitable product, policy conflict, or a judgment call that cannot be reduced to the available information.

Those patterns are operationally valuable. If advisors repeatedly correct stale product information, the problem may be the knowledge source rather than the model. If they repeatedly reject recommendations because a client’s real priorities are missing from structured data, the firm has learned something about its CRM and discovery process. If the same type of escalation occurs across teams, the firm can tighten the boundary before expanding the system.

There are a variety of pressures: AI adoption, automation, data governance, operational resilience, regulatory complexity, and scalable operating models. Wealth firms should treat those as one connected design problem. AI cannot be scaled safely by discussing the model separately from the operating process around it.

This is also a talent question.

When advisors know exactly where their judgment is expected, AI can strengthen professional identity rather than threaten it. The advisor is no longer competing with software on who can produce a faster summary. The advisor owns the interpretation, the client context, the challenge to weak evidence, and the decision to proceed or stop.

That is a better division of labour than asking people to supervise an opaque system without telling them what supervision requires.

Leaders can start with one workflow rather than attempting to govern every use at once. Take investment-review preparation, for example. Let AI assemble approved materials and draft the comparison. Require the advisor to validate every claim that changes the recommendation. Define a short list of conditions that force escalation. Then review two weeks of human corrections before deciding whether the workflow should scale.

This approach produces a more meaningful adoption metric as well. Instead of counting licences, prompts, or documents generated, firms can ask how often the workflow stayed inside the preparation zone, how much advisor correction was required, which issues triggered escalation, and whether the final client outcome improved.

The wealth industry has good reasons to adopt AI quickly. Clients expect responsiveness. Advisors face heavy information loads. Firms need productivity without losing the trust on which advice depends.

But speed is useful only when responsibility remains legible.

As Asian wealth managers move from copilots to agents, the most important control may be a simple one: before the system scales, everyone should know where the machine’s work ends and professional judgment begins.
 

About the author
Gleb Tsipursky, PhD, is a behavioural scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026).