The same clients, two years apart
In 2007, a large Italian bank surveyed 1,686 clients about their attitude to risk. In June 2009, after the worst financial crisis in eighty years, three economists went back to the same people and asked the same questions, reaching 666 of them. Asked to choose between four combinations of risk and return, 15.9% had picked the most conservative option, "low return and no risk," in 2007. Two years later, 42.6% of the same people picked it. Across the sample, 46% had become more risk averse and 10% less. Among those who had chosen the most aggressive option in 2007, 83% moved to a more conservative one.
The obvious explanation, that these clients had lost money and could no longer afford the risk, found little support: changes in wealth and in consumption habits had no measurable effect on changes in risk aversion. The authors investigate fear as a possible explanation: a frightening experience may raise risk aversion directly. In a laboratory experiment alongside the survey, students who had watched a horror film had a certainty equivalent 27% lower than those who had not. A questionnaire filed in 2007 had no means of anticipating that change.
For a wealth manager, the corollary is uncomfortable. A risk questionnaire measures something real, but it measures it at one moment, usually a quiet one, since onboarding and periodic reviews are rarely scheduled for the week of a crash. The number then sits in the file as if it were a stable trait, and later recommendations are checked against it. One important dimension of risk tolerance, however, is how the client responds when the portfolio is falling, and that is precisely the moment the questionnaire rarely captures.
In its 2011 review, the UK's Financial Services Authority found that, in half of the unsuitable files it assessed from the March 2008–September 2010 period, the investment selection failed to match the risk the customer was willing and able to take. It also examined eleven risk-profiling tools and found that nine had weaknesses that could, in certain circumstances, produce flawed results. The European Securities and Markets Authority took the point further in its 2023 suitability guidelines: firms should not rely unduly on clients' self-assessment, "self-assessment for the risk tolerance should be avoided," and the information collected should be "reliable and consistent," with firms alert to "any relevant contradictions between different pieces of information collected."
Many firms already hold a substantial second body of information about how each client responds to risk. It sits in their order books, their recorded calls and their meeting notes. It is rarely read for what it says about client behaviour during periods of market stress.
Three records of one client's tolerance
A firm that has served a client for years holds at least three separate records of how that client relates to risk, each produced by a different part of the firm for a different reason.
The first is the stated record: the risk questionnaire, the resulting score or category, the answer to the question about acceptable losses, and the assessment of capacity for loss. It is produced at onboarding, refreshed on a schedule set by the firm's policy, and owned by the onboarding or compliance function. In many firms, it is the principal record the suitability process formally consults. ESMA's guidelines push firms to make it concrete: instead of asking whether a client is comfortable with risk, "the firm should ask what level of loss over a given time period the client would be willing to accept." That improves the question, but the answer is still given in advance, about a loss the client is imagining rather than living through.
The second is the enacted record: what the client actually instructed the firm to do. It includes sales into falling markets, requests to move money to cash, refusals of a proposed rebalance, demands to override the model portfolio, and insistence on adding to a position that has just doubled. Each carries a timestamp that can be matched to the state of the portfolio at that moment. That makes the enacted record the most directly observable account of what the client instructed the firm to do. It shows the conditions under which he acted, measured in his own portfolio rather than in a hypothetical. It does not, by itself, establish why he acted or where his tolerance for loss actually lies.
The third is the spoken record: what the client said, and when. Under Article 16(7) of MiFID II, investment firms must record telephone conversations and electronic communications relating to the reception, transmission and execution of client orders, including those intended to result in a transaction even when none follows; a 2025 ESMA Q&A confirms that this includes discussions of investment options that are intended to result in a transaction, even when the order is subsequently placed through another channel. The records must be kept for five years, a period the competent authority can extend to seven. The obligation does not reach every conversation between advisor and client, so the rest of the spoken record depends on what the firm keeps: minutes of meetings, the advisor's notes and the client's emails. Together, these often hold years of the client's own words about risk, many spoken on the day of a sell-off.
The three records rarely sit in the same place. The questionnaire lives in the onboarding or suitability module, the instructions in the order management and portfolio systems, the recordings in a surveillance archive, and the notes in the CRM. Each system has its own purpose, access rules and review process. No single process routinely brings the three records together to assess how the client has responded to risk.
Where the records disagree
Consider a client we will call Paolo. He, his firm and his portfolio figures are illustrative; the market episodes are real. Paolo is in his late sixties and has sold his stake in a family-owned manufacturing business. He opened an advisory mandate of €6.8 million in 2019 with an independent wealth manager in northern Italy that looks after about 600 families; at its February 2025 peak, the portfolio was worth about €8.4 million. He also owns property and draws a pension, so his capacity for loss is high. His questionnaire, completed at onboarding in 2019 and refreshed in March 2023, places him in category 5 of 7, "growth," and records that he would accept a loss of up to 25% over a year. Applied to the February 2025 peak, 25% is €2.1 million.
His enacted and spoken records tell a different story. In the last week of February 2020, with his portfolio 8% below its high, he called requesting that half his equity exposure be sold; the advisor persuaded him to limit the sale to €900,000, which was moved to cash and reinvested in November, after most of the recovery. In May 2022, with the portfolio 9% below its previous peak, he asked in a review meeting to cut equities to 50%; the advisor's note records that he "would rather sleep at night than chase returns," and the allocation was reduced from 70% to 60%. On 4 April 2025, two days after the United States announced its new tariffs and with the portfolio 8% below its February peak, he called three times in four days and moved €1.2 million into a money-market fund. With the portfolio's previous peak at about €8.4 million at that time, an 8% drawdown represents a loss of roughly €670,000.
Three times between 2020 and 2025, Paolo acted when his portfolio was 8% or 9% below its previous peak, substantially below the 25% annual loss recorded in his file. The figures are not strictly equivalent: one is a portfolio drawdown from a previous high, the other a stated loss over a year. But the divergence is large enough to warrant examination. His questionnaire is not a lie. It is an honest answer to a hypothetical, and his behaviour is an equally honest answer to the real thing. The suitability file holds only the first. The second exists, in pieces, in three other systems. By March 2023, the first two episodes were already sufficient to raise questions about the refreshed questionnaire.
Paolo's is a common form of divergence, but not the only one; comparing the three records across a client book tends to surface four patterns.
| Pattern | What the records show | Why it matters |
|---|---|---|
| Overstated | Stated growth; acts at drawdowns well short of the stated loss | Portfolio riskier than the client can live with; losses crystallised at lows |
| Risk-seeking divergence | Stated cautious; repeatedly requests concentration, leverage or exposure beyond the recorded profile | Pressure to move beyond the documented profile; needs a suitability assessment, not automatic reclassification |
| Drifted | Older records agree; recent ones diverge | Profile out of date after a life event or a shock |
| Asset-specific | Calm about equities; alarmed by illiquids or currency | One score hides the constraint that actually binds |
An overstated profile leaves the client in a portfolio he abandons at the worst moment, which damages his returns and, eventually, the relationship. A risk-seeking divergence is a conduct risk in the other direction: a client who keeps demanding more risk than his file supports will eventually get some of it, with a thin record of why. The drifted profile is the one ESMA has in mind when it asks firms to review client information so that it does not become "manifestly out of date." The asset-specific pattern is the subtlest. A client may sit calmly through a 20% fall in listed equities and lose sleep over a private-equity fund that has not reported a valuation in nine months; a single risk score cannot record the difference.
None of this makes the enacted and spoken records automatically right. A sale into a falling market can reflect fear, but also a need for liquidity, a change in conviction, new information or a rebalance planned anyway, and a client who calls during a sell-off may simply want reassurance. The records show observed responses to risk; a conclusion about the client's tolerance is an inference that the firm has to test, ideally with the client. When the records disagree, the disagreement is evidence that the suitability assessment should weigh, and at present it is evidence that often fails to reach the suitability assessment.
Why the evidence stays in the archive
The evidence is already in the firm's possession, yet it rarely reaches the suitability file, because each record was built for a purpose other than understanding the client's tolerance. Call recordings exist for surveillance and disputes, so they sit in an archive searched after a complaint, for a specific call. Meeting notes are free text, written quickly so that the firm can show what was agreed, and read mostly by their author. Order histories exist to settle trades and calculate performance. None of these systems was designed to show how a client behaves when losing money, or that the evidence sits in the other two.
The review cycle compounds the problem. Many firms update the risk profile on a calendar: at onboarding, then at fixed intervals or when the client reports a change in circumstances. None of Paolo's three episodes coincided with those dates, and nothing in the process treated a call during a sell-off as a reason to look at his file. At his March 2023 review, the questionnaire was simply filled in again and produced the same answer, because Paolo was asked the same hypothetical question on another calm day.
Reconciling an overstated profile also usually means proposing a more conservative allocation, which can mean a lower-fee portfolio and telling a successful entrepreneur that his own record contradicts his self-image. Leaving the questionnaire alone is easier for everyone, until the next sell-off, when the client sells at the bottom and asks why the firm had him in a portfolio he could not hold.
The regulatory direction runs the other way. ESMA's guidelines ask firms to counterbalance self-assessment "by objective criteria," to address the risk that clients "overestimate their knowledge and experience," and to be alert to contradictions between pieces of information. They also expect firms to contact clients to resolve material inconsistencies, but they do not prescribe how contradictions should be found or ask firms to analyse recordings or order histories; reading the three records together is the approach this article proposes. Holding a record is also different from being free to use it for any purpose: reusing recordings for suitability must be assessed under the applicable data-protection rules, including the need for a documented purpose and legal basis, appropriate transparency towards the client and data minimisation. Within those limits, a firm that holds years of a client's recorded reactions to falling markets, and relies instead on a self-assessed questionnaire that those records contradict, will find it hard to argue that its information was reliable and consistent.
What reconciling the records would require
Putting the three records side by side is harder than a reporting exercise, because they differ in form and in how far they can be trusted. A system that did it properly would have to meet six requirements.
- Anchor every reaction to the portfolio at that moment. An instruction or a call means little until it is matched to how far the client's own portfolio had fallen that day, from which peak, and in which currency. That figure has to come from the same performance engine the firm uses for reporting, so that the 8% discussed in the review is the 8% on his statement.
- Read the unstructured records, and cite them. The useful evidence in a recording or a note is a sentence: "I can't watch this," "sell half," "I'd rather sleep at night." Finding them across years of notes and transcripts is reading work, and every finding has to link back to its source for checking.
- Compare against the stated profile, and say how the measures differ. The questionnaire states a loss over a year; the observed reactions are drawdowns from a peak. The comparison should show both, as percentages and in the client's currency, with the difference in measure stated: €2.1 million over a year in the file, against a fall of roughly €670,000 from the previous peak when Paolo acted.
- Treat divergence as evidence for a person to weigh. The output is a flag and a stated hypothesis, with its supporting records, routed to the advisor and, where the firm's policy requires, to compliance. It never changes a profile, an allocation or a risk category by itself.
- Record the decision without erasing the past. Whether the firm keeps the profile, revises it or schedules a conversation, the decision, its reasons and its evidence are recorded beside the earlier questionnaire, which stays in the file.
- Review on events, not only on the calendar. A drawdown beyond a threshold, a cluster of calls or an override request should prompt a look at the profile while the evidence is fresh.
Two constraints sit over all six. The first is symmetry: a firm must use the same process when the records suggest a client can bear more risk as when they suggest less, or it will end up using behavioural evidence selectively. The second is access. Recordings and notes are sensitive, so each person should see only what their role permits, and the client should know that his recorded instructions and conversations inform the assessment of his profile, on a legal basis the firm can document, with every use logged so that it can be audited.
How Pivolt approaches it
Pivolt keeps the three records in one data model. The questionnaire and its history sit in the onboarding and suitability module, instructions and transactions in the portfolio and order records, and notes in the CRM. Call recordings stay in the firm's recording system and are brought in through an integration that links each call and its transcript to the client and the date. The analysis uses only the records the firm holds and is authorised to use for this purpose. Because everything refers to the same client and portfolio, each instruction and conversation can be placed next to the portfolio as it stood that day.
Take Paolo's annual review, which his advisor has scheduled for late November 2026. Preparing for it, she asks the AI Advisor whether his risk profile is consistent with how he has behaved. The answer starts from the stated record: category 5, a 25% annual loss accepted, last confirmed in March 2023. It then lists the three episodes, each with its date and the drawdown of his portfolio from its previous peak, in percent and in euros. Those numbers come from Pivolt's performance engine, not from the language model, which only reads the records and explains what it found. Each episode links to its sources: the order tickets from February 2020, the meeting note from May 2022 with the sentence about sleeping at night, and the transcripts of the three calls in April 2025. She can open each one and check that the summary is fair.
The AI Advisor presents the result as a hypothesis, a potentially overstated profile for the advisor to review, and notes what the records do not show: no withdrawal, liquidity need or change in circumstances was recorded around any of the three episodes. It suggests how to use the meeting: show Paolo the three episodes, ask the loss question again in euros rather than percentages, and discuss whether 60% in equities still fits. It drafts that agenda for her. It does not change his category, propose trades or contact him. In the meeting, Paolo looks at the April 2025 figures and says that a loss of around €700,000 is where he starts to lose sleep. The advisor uses that answer as part of a broader review of his risk tolerance, investment objectives and capacity for loss. The revised assessment places him in category 4. She proposes a move to 50% equities, and the change goes through the firm's normal approval workflow. The 2023 questionnaire stays in the file beside the new one and the evidence behind it, so a supervisor can later see why the profile changed.
Compliance uses the same capability across the book. In this illustrative firm, a query across its 600 client families identifies 43 cases for advisor review because their enacted or spoken records appear to diverge materially from the stated profile: 29 potentially overstated, 8 risk-seeking and 6 asset-specific. These are preliminary categories for triage, not findings of unsuitability. The list goes to the responsible advisors as review tasks, with evidence attached. Because risk-seeking cases appear in the same list as overstated ones, the process cannot quietly become a way to justify riskier products.
Three guarantees hold throughout. Every figure, from a drawdown to its euro amount, comes from the deterministic engine. No profile, allocation or client communication changes without a person confirming it. And each user sees only what their permissions allow: the advisor sees the transcripts of her own clients' calls, compliance sees the flags across the book, and the AI Advisor's answers respect the same boundaries.
A profile that answers to the record
The clients of that Italian bank did not mislead anyone in 2007. They described their appetite for risk accurately for the conditions they were in, and did so again two years later, in very different ones. What failed was the assumption that the first answer would hold.
Many wealth managers make that assumption, because in their suitability process the questionnaire is the principal record built to be read. Yet the same firms are required to keep years of recorded order-related calls and communications, and they already store every order with a timestamp. For a client like Paolo, those records contain behavioural evidence that the questionnaire alone cannot capture, evidence the suitability assessment should weigh alongside his stated tolerance and his capacity for loss, and it was available well before his last review. The practical change is to let the suitability assessment read them: put the stated, enacted and spoken records side by side, show the advisor where they disagree, and record what the firm decided. A risk profile built that way will still be wrong sometimes, but when it is, the firm will see it before the next sell-off rather than after.
Sources
- "Time Varying Risk Aversion", Journal of Financial Economics (2018); publication record. Used: 1,686 clients surveyed in 2007; 666 re-interviewed in June 2009; share choosing "low return and no risk" 15.9% in 2007 and 42.6% in 2009 (Table 7); 46% more risk averse, 10% less; 83% of the most aggressive moved to a more conservative option; no effect of changes in wealth or habit; certainty equivalent 27% lower after a scary video (working paper abstract). Italy, 2007–2009.
- Financial Services Authority, Assessing suitability: establishing the risk a customer is willing and able to take and making a suitable investment selection, guidance consultation GC11/01 (January 2011), finalised as FG11/5 (March 2011). Used: half of files assessed as unsuitable between March 2008 and September 2010 failed on risk; nine of eleven tools reviewed had weaknesses that could lead to flawed outputs. United Kingdom.
- European Securities and Markets Authority, Guidelines on certain aspects of the MiFID II suitability requirements, ESMA35-43-3172, official version published 3 April 2023 (applicable from 3 October 2023), following the final report of 23 September 2022. Used: paragraphs 44 (reliable and consistent information, without unduly relying on self-assessment), 46 (self-assessment counterbalanced by objective criteria; loss-based question), 48 (self-assessment for risk tolerance to be avoided), 51 (relevant contradictions; contacting the client to resolve material inconsistencies), 52 (overestimation of knowledge and experience) and 54 (information not manifestly out of date).
- European Securities and Markets Authority, Q&A on the retention of recordings (ESMA_QA_1770, 10 October 2016). Used: records under Article 16(7) of MiFID II kept for five years, extendable to seven only at the competent authority's request.
- European Securities and Markets Authority, Q&A on the scope of the recording obligation (ESMA_QA_2416, answer published 18 June 2025). Used: communications intended to result in a transaction must be recorded even if no order is placed during the interaction, including discussions of investment options intended to result in a transaction, even if the order is later placed through another channel.
Paolo, his portfolio figures, the wealth management firm, its advisor and the book-level counts (600 families, 43 flagged) are illustrative and were constructed for this article. The market episodes (February 2020, 2022, April 2025) are real; the portfolio drawdowns attributed to them are illustrative.



