A client sits down and says they have already run the numbers. Not in a spreadsheet. In a chatbot.
They arrive with a savings rate, a rough allocation, and a view on when they can stop working. The natural response is to evaluate the answer. A better first move is simpler: Show me the prompt.
An AI-generated conclusion is an artifact with a hidden history. It reflects the client's circumstances, the details they chose to include, the details they did not know to include, the assumptions made by the system, and the particular model used at that moment.
The answer may be reasonable and still be incomplete.
What recent research tells us
A 2026 working paper by Taha Choukhmane, Tim de Silva, Weidong Lin, and Matthew Akuzawa offers a useful way to think about the problem.
The researchers collected prompts from a demographically representative sample of U.S. adults, used those prompts to obtain spending and investment guidance from two large language models, translated the guidance into quantitative decisions, and simulated the results over a financial life cycle.
The advice was often directionally sound. In the simulations, following it moved many respondents closer to the prescriptions of the researchers' life-cycle model, including broader participation in diversified equities, equity exposure that declined with age, and larger savings buffers.
The study also found recurring weaknesses:
- Recommended saving and withdrawal rates often relied on round-number heuristics.
- Recommended spending responded too sharply to income shocks, even when liquid savings were available to help smooth consumption.
- Portfolio allocations tended to drift with market returns instead of being actively rebalanced.
- Outcomes varied with the information and language in the prompts.
These results do not establish that a model's recommendation is suitable for a particular client. They show something more useful for an advisor: an answer can sound sensible while still depending heavily on incomplete inputs and simplified assumptions.
Better inputs changed the advice
The most practical finding concerns the prompts.
When the researchers replaced participant-written prompts with structured prompts containing relevant financial variables and explicit modeling assumptions, the same models relied less on simple rules of thumb and produced advice that more closely matched the study's life-cycle benchmark.
For example, the share of recommended saving rates that were multiples of 10 fell from 31% to 14.6%. The structured prompts also produced better consumption smoothing after income shocks, although they did not resolve the portfolio-rebalancing weakness.
The study found meaningful differences across groups as well. Advice generated from prompts written by respondents with lower financial literacy led to simulated wealth at age 60 that was $45,878, or 4.11%, lower than advice generated from prompts written by respondents with higher financial literacy. The paper also found differences associated with prior AI experience and gender.
Those findings require care. They come from a simulation, not observed client outcomes, and they do not show that any individual prompt or recommendation is poor. They do show that the quality of an AI-generated answer depends partly on whether the person asking the question knew which facts and assumptions mattered.
That is familiar territory for an advisor. The professional value is not merely producing an answer. It is discovering the information the first question left out.
Five questions worth asking
An advisor does not need to dismiss the technology or endorse its output. Treat the answer as you would any other artifact of uncertain provenance: understand its source, test its assumptions, and decide what weight it deserves.
1. Show me the prompt
A useful first response is simple: Show me the prompt.
Ask the client to share the prompt and relevant follow-up exchanges, if they are comfortable doing so. If the conversation cannot be shared, ask them to describe the information they provided.
The purpose is not to grade the client's question. It is to recover the inputs behind the conclusion.
2. What could it not have known?
Consider information that may change the analysis: held-away accounts, a pension, concentrated stock, an aging parent, a business interest, a tax carryforward, state tax treatment, an estate-planning decision, or a health circumstance that changes the time horizon.
The system may not have ignored those facts. It may never have received them.
3. Is the number calculated or recalled?
A savings rate of exactly 15% or a withdrawal rate of exactly 4% may be a useful starting point. It may also be a familiar convention presented as though it were derived from this client's circumstances.
Ask which inputs drive the number, which assumptions sit behind it, and how the answer changes when those assumptions change.
4. What does the answer assume happens next?
Many recommendations depend on future behavior: the household continues saving after a job change, the portfolio is rebalanced, spending adjusts as expected, or a fixed withdrawal rule remains appropriate.
An ongoing advisory relationship can revisit those assumptions as circumstances change. A one-time answer cannot do that on its own.
5. What would need to be true for this to be appropriate?
This moves the discussion away from a contest between the advisor and the tool. The question becomes conditional: under what facts and assumptions would the conclusion hold, and which of them have been verified?
That is a more productive conversation than deciding whether the answer is simply right or wrong.
Preserve the reasoning, not the novelty
If the client's AI-generated conclusion materially affects the discussion, preserve the reasoning that follows in the firm's approved recordkeeping process.
The important point is not that AI appeared in the meeting. It is what the client brought, which relevant facts were verified, which assumptions were challenged, what was set aside, and how the advisor reached the next conclusion.
That record supports continuity. Another authorized professional should be able to understand the basis for the work without reconstructing the conversation from memory.
Keep client information within firm-approved workflows
Reviewing a client's prompt can expose a dense collection of financial, family, health, tax, and planning information. Use only systems and workflows approved by the relevant firm for receiving, reviewing, or storing it.
Removing a client's name does not automatically make the remaining information appropriate for an unapproved tool. The surrounding facts may still be sensitive, identifiable, or subject to firm policies.
Questions about records, privacy, supervision, suitability, or the use of a particular system are firm- and deployment-specific. Advisors should follow their firm's policies and involve the appropriate compliance, privacy, legal, or security professionals when needed.
Avoid the two easy extremes
The first mistake is dismissal. Telling a client that the tool does not know what it is talking about may be inaccurate and is unlikely to invite a useful conversation. The research found advice that was often broadly sensible.
The second mistake is adoption. Treating the output as a ready-made starting point imports assumptions that nobody has examined, including facts the client may not have known to mention.
The better posture sits between the two: take the answer seriously, and take its provenance more seriously.
The working paper has important limits. It evaluates advice from specific model versions at a particular point in time. Its outcomes are simulated under a defined economic model, and it does not measure whether people would follow the advice in practice. It is also a working paper rather than peer-reviewed research.
Those limits narrow what we should claim from the study. They do not weaken the practical lesson.
The client may arrive with an answer. The advisor's work begins with what the question left out.