Recorded,Not Rated
Every wealth manager in Europe owns a dataset it has never read: years of recorded Client dialogue, kept because MiFID II says so. It can tell you which relationships are at risk and prove your advice matched what the Client said. The EU AI Act decides whether that reading is an asset or a high-risk system. This paper shows where the line runs.
Published 7 October 2026, updated 9 October 2026 (v1.1).
Three numbers
- 81%of next-generation inheritors plan to switch firms within one to two years of inheriting
- 41%of Relationship Manager time goes to operational tasks, not Clients
- 2 Dec 2027the day Annex III high-risk obligations apply
Sources: Capgemini World Wealth Report 2025 and 2026 press releases; Regulation (EU) 2026/1744.
| 2026-03-04 | Call, 14 min | Quarterly review |
| 2026-03-04 | Attendance | Lena Klein absent from quarterly call, third time running |
| 2026-04-11 | Reporting pack sent | |
| 2026-05-02 | Call, 22 min | Asks, for the second time: "what happens to the mandate after me" |
| 2026-06-18 | Meeting | Allocation discussion, minutes filed |
| 2026-07-09 | Requests a liquidity plan for a property purchase. No follow-up on file. | |
| 2026-08-21 | Fee statement |
Future financial commitment mentioned in dialogue, not on the profile.
Everyone records. Nobody understands. The archive you already keep can answer two questions, if you follow one rule: score the relationship, never the Relationship Manager.
Forty minutes of reading for the people who will have to defend the decision. Here is the one-page version.
The archive nobody readsMiFID II makes you record calls and messages that lead to transactions and keep them for five to seven years. In most firms the archive is a fire extinguisher: opened after a complaint, a dispute or a supervisory request. Nobody looks at it while the building is standing.
Two answers in one recordThe same dialogue shows whether a relationship is cooling (an heir who stops attending, a fee question asked twice) and whether the advice matched what the Client said. The first becomes a Health Score with every point traceable to its message. The second becomes a suitability prompt with the evidence attached.
The line, in the Commission's wordsAI that monitors and evaluates workers is high-risk under Annex III point 4(b). The Commission's draft examples put a tool that scores "responsiveness to emails" inside, and a tool that reports deviations only to the employee in charge outside. Build on the right side and the archive is an asset.
One record, two answers
The model in the paper, drawn once. Press a marker to read why each part is built the way it is. Nothing in the right-hand column is keyed to an employee.
Recorded Client dialogue
Already kept under MiFID II Article 16(7). Read with a documented legal basis.
- Email and calendar (M365)
- Calls with transcripts
- Meeting notes and CRM (Salesforce)
One substrate
Statements dated and linked to their source message. Indicators keyed to the Client relationship.
What it must not become
A per-RM score, ranking or productivity figure. Hatched because it is drawn only to be ruled out.
Relationship health
A Health Score, 0 to 100, for the relationship. A meaningful drop becomes a Signal that goes first to the RM who owns it.
Suitability evidence
A gap between what the Client said and the profile on file becomes a prompt to update, with the evidence trail attached.
Start with any marker
Five short notes, one idea each: the legal basis for reading the archive, the unit of analysis, why no per-RM score exists, why the RM sees a Signal first, and what actually enters the suitability file.
Reading the archive needs its own legal basis
MiFID II recordings are processed under a legal obligation, for an evidential purpose. Reading them for suitability evidence sits close to that purpose and can pass a compatibility assessment. Reading them for relationship health is further away and needs legitimate interest, a balancing test and a DPIA. The GDPR question arrives before the AI Act one.
Paper, Section 5.6
The unit of analysis is the Client relationship
Every indicator, every score and every prompt is attached to a Client or household. No output is keyed to an employee identifier, so there is nothing to rank. That is rule A.1 of the standard, and it is the single design choice that most of the rest follows from.
Appendix A, rule A.1
Why the hatched box exists
The Commission's draft example of a law firm that aggregates "responsiveness to emails" into a score is inside Annex III point 4(b). The paper draws that outcome in order to rule it out: no per-RM ranking, league table, productivity score or trend, computed, stored or displayed. A human reviewing such a score does not change the classification.
Paper, Sections 5.2 and 5.4
Why the RM sees the Signal first
The Commission's courier example is outside 4(b) because the system informs only the employee in charge and supports the work rather than evaluating the worker. A Client Signal routed first to the RM who owns the relationship has the same shape. Management views aggregate at firm level; grouping by book is navigation, never comparison.
Paper, Section 5.2; Appendix A, rules A.4 and A.5
What enters the suitability file
The evidence trail, not the model's opinion: the dated statement, its source message, and the discrepancy with the profile, presented to the RM and, where the firm's procedure requires, to the control function, as a prompt to update under ESMA's suitability guidelines (¶54).
Paper, Sections 3 and 4
Dialogue ↓ Statement ↓ Profile gap → Prompt with evidence
Where the line runs
The European Commission's draft guidelines of 19 May 2026 contain two worked examples that could have been written for wealth management. Same kind of data, opposite outcome, and the difference is intended purpose.
A law firm allocates work by behaviour
"The system ingests data on billing hours, turnaround times on prior assignments, responsiveness to emails, and voluntary participation in firm activities ... These behavioural inputs are aggregated into a score, which serves as a basis for the allocation of client matters."
Replace "associates" with "Relationship Managers" and "client matters" with "Clients", and you have the tool most RM dashboards quietly become.
Commission draft guidelines, AI Act Service Desk, employment examples
A courier app notices a detour
"If the AI system detects deviations, it will inform only the employee in charge ... supporting workers in their tasks rather than evaluating workers' performance, the monitoring activity is incidental and falls outside the use case of point 4(b) of Annex III."
A Client Signal that goes first to the RM who owns the relationship has exactly this shape: it supports the work and leaves the decision with the person in charge.
Commission draft guidelines, AI Act Service Desk, employment examples
The guidelines are a draft and not binding; the final text is expected by the end of 2026. A human in the loop does not change the classification, and the paper explains why relying on the Article 6(3) filter is the weaker position.
The RM-neutral standard
Seven design rules, published as an open proposal. Not a legal guarantee: a set of choices documented so that a supervisor, a staff delegation or a sceptical CCO can test them. Hand them to any vendor and ask for evidence.
Every output is attached to a Client or household. No output is keyed to an employee.
None displayed, in any view.
Written into the intended purpose and the marketing. The contract clause is being drafted.
The same "employee in charge" shape as the Commission's courier example.
Grouping by book or team is navigation, never comparison.
No inference of RM stress, mood or engagement. Prohibited in the workplace anyway.
Under Article 6(3) and Annex III, versioned on every material change, open to supervisors.
Five lines worth repeating
Click to copy. Use them in your next steering committee, with or without attribution.
Article 50 transparency obligations apply.
Marking of generated content for systems already on the market.
Final Commission high-risk classification guidelines, announced.
Annex III high-risk obligations, including 4(b), apply.
Retail Investment Strategy expected to apply, 30 months after entry into force.
The Luxembourg layer
The first regulator a Client-communications tool meets is not the CSSF. It is the firm's own staff delegation, with the CNPD one letter away. Three numbers from the Labour Code.
After the employer's prior information, the staff delegation may ask the CNPD for an opinion on the planned processing (Article L.261-1).
The CNPD's time to answer. The request has suspensive effect: nothing goes live while it is pending.
From this size, installations that control employees' behaviour and performance need the delegation's agreement (Article L.414-9).
The paper includes a briefing template for the staff delegation, with the formal no-reuse commitment the law requires, and explains who will supervise AI in the financial sector once Bill 8476 is adopted.
Ten questions to ask any vendor
Before an AI system reads your firm's recorded Client communications, put these ten questions to the vendor. Each comes with what a good answer looks like. This is Appendix B of the paper, published here in full; it applies to every vendor, Luscent included.
- What is the documented intended purpose of the system?
A good answerA written statement, consistent across the instructions for use, the contract and marketing, that names Client relationship management and suitability evidence as the purposes and expressly excludes monitoring or evaluating employees.
- Does any output exist, even internally, per employee?
A good answer"No." Architecture diagrams show outputs keyed to Client identifiers; the vendor can show that no per-RM table exists in the data model; management views aggregate at firm level.
- Does the system perform profiling, of whom, and why?
A good answerClient profiling occurs (relationship health, suitability indicators), with the legal basis and a DPIA template. Staff profiling does not occur, and the vendor explains how this is enforced.
- Where is your Article 6(3) and Annex III self-assessment?
A good answerA versioned document, shared under NDA, that analyses 4(b) against the Commission's draft examples, states whether the filter is relied on, and commits to updating it when the final guidelines are published.
- Which Article 50 obligations apply, and how do you meet them?
A good answerMapping of Client-facing chat (disclosure at first interaction), generated text (machine-readable marking or a reasoned exemption, Code of Practice status), and confirmation that no emotion recognition is used.
- Where are models hosted and data processed?
A good answerNamed hosting provider and region, model families used (for example open-weight models on EU infrastructure, as one option), no training on Client data without a contract, and the transfer position if any processing leaves the EEA.
- What do you provide to support our DORA obligations?
A good answerThe information needed for the buyer's register of information (DORA Article 28), acceptance of mandatory contract terms (Article 30), incident notification cooperation, and audit and access rights. The buyer owns these obligations; the vendor enables them.
- Who are your sub-processors?
A good answerA current list with location and function, advance notice of changes, and a right to object.
- What do you log, and how long do you keep it?
A good answerLogs sufficient to trace every Signal and suitability prompt, retention aligned with the firm's MiFID II schedule, and access controls that stop logs being used to reconstruct RM performance.
- How do we exit?
A good answerExport of all Client-level outputs and evidence links in open formats, deletion certification, a transition assistance period, and no lock-in of the underlying recordings, which stay in the firm's archive.
Who wrote it
Two founders with a commercial interest in the thesis, which the paper says in its own limits section.
CEO and co-founder of Luscent. EFPA European Financial Advisor, former head of the high-net-worth and affluent business at a bank in Luxembourg, former family office adviser at a strategy consultancy, co-author of an ANBIMA 2024 study on foreign investment into Brazil.
Co-founder and Chief Science Officer of Luscent. Co-author of the RM-neutral design standard in Appendix A.
Frequently asked questions
The questions we hear first.
Is this paper legal advice?
Does reading recorded communications make a firm a high-risk AI deployer?
What about GDPR? The recordings were made for MiFID II.
Why is there a Luxembourg chapter?
Isn't this what communications surveillance already does?
Is the Retail Investment Strategy already an obligation?
Get the paper
Eleven sections and three appendices, about 40 minutes to read. Written for CCOs, Heads of Private Banking, COOs and Heads of Digital at private banks, family offices, EAMs and IFAs.
- 1The one-page version
- 2Clients follow people, and the people are stretched
- 3Recorded, never read: the evidence gap
- 4One record, two answers
- 5Where the AI Act draws the line, with the GDPR twin
- 6The RM-neutral standard
- 7The Luxembourg layer, and a note for Swiss readers
- 8What changes by 2029
- 9Ten questions to ask any vendor
- 10What to do this quarter, by role
- ADesign standard, vendor checklist, staff delegation template
Ask for the PDF by email.A person replies with the paper.
Opens your email app with a short request. We reply once, with the paper. No newsletter, no sequence. If you would also like a note when the final Commission guidelines land, say so in your email.
Or write to armin@luscent.io.