Geneva private banking's quiet AI-treasury experiment.

The architecture: Claude-grounded reasoning, advisor-in-the-loop

The technical architecture across both banks is similar enough in its outlines — and different enough in its specific design choices — that the systems are best understood as two implementations of a shared underlying model. The shared model: an LLM-based reasoning layer that consumes rate curves, sovereign credit signals, FX volatility surfaces, and the bank's in-house economist research as structured inputs, and produces a per-client cash-management recommendation that the relationship manager reviews, modifies, and presents to the client under the bank's standard discretionary mandate. The recommendation does not move money. The relationship manager moves money. The recommendation is a structured input to the relationship manager's decision, not a substitute for it.

Pictet's implementation runs on a Claude Opus 4.7 endpoint hosted in a Pictet-controlled Microsoft Azure tenant with the EU-resident inference region pinned to Frankfurt. The model is wrapped by a retrieval layer that grounds reasoning on the bank's proprietary economist letters — Pictet's chief economist letter is published weekly, plus regional desk notes from the Geneva, Zürich, and Singapore offices — and a structured-data layer that feeds in the day's rate curves (CHF, EUR, USD, GBP, JPY), the FX volatility surfaces from the bank's Bloomberg AIM Pro deployment, and the sovereign credit signals from the bank's internal credit committee dashboard. The reasoning layer produces a per-client cash-management recommendation that includes: a recommended portfolio breakdown across short-dated sovereigns, money-market exposure, and bank-deposit allocation; a recommended FX hedge ratio for the client's home-currency profile; and a written rationale of approximately 400-600 words that the relationship manager can edit and forward to the client as part of the standard client letter package. The relationship manager has full edit rights on the recommendation and the rationale. The system logs every accepted edit and every overridden recommendation. The audit trail is structured to support FINMA inspection on demand.

Lombard Odier's implementation is structurally similar but differs in two important design choices. First, the bank runs its inference layer through an Anthropic-managed dedicated capacity arrangement rather than a self-hosted Azure deployment. The choice was a deliberate one: Lombard Odier's technology team concluded, after a six-month internal evaluation in early 2025, that the operational overhead of running its own inference layer outweighed the marginal control benefit, and the Anthropic dedicated capacity arrangement — which provides the bank with provisioned throughput, dedicated logging, and a contractual data-handling regime that satisfies FINMA's outsourcing requirements — was the better economic and operational fit. Second, Lombard Odier's retrieval layer is grounded on a wider research substrate than Pictet's: the bank's economist letters are supplemented by a structured archive of the bank's Sustainable Investing research desk's notes, which feed the recommendation layer with ESG-adjacent signals that the bank's client base — heavier in European foundations and family offices with explicit ESG mandates — values in the cash-management recommendation. The architectural difference produces different output styles. Pictet's recommendations are more sovereign-credit-anchored. Lombard Odier's are more ESG-and-sustainability-anchored. Both are recognisably the same kind of system.

The grounding methodology is the technical heart of both implementations, and it is where the banks have invested the most engineering effort. The retrieval layer at both banks runs on a vector index of the proprietary research archive — Pictet's reportedly contains approximately 14 years of weekly economist letters and 8 years of regional desk notes; Lombard Odier's spans a similar timeframe with the addition of the sustainability research archive — with retrieval triggered per-query and the retrieved documents passed into the reasoning layer's context window. The Opus 4.7 1M-context window is the critical enabling technology: both banks confirmed in the closed-door convenings that the system would not be viable on a smaller-context model because the grounded research substrate needed in any single reasoning step routinely exceeds 200,000 tokens. The grounded retrieval is what gives the system its claim to be a defensible advisory tool rather than a generic LLM rate commentary generator. The relationship managers have demanded — and the banks have provided — full traceability from any recommendation back to the specific economist letters and rate-curve inputs that drove it.

The discretion model: advisor-in-the-loop, not automation

The discretion model is where the two banks' pilots diverge most clearly from the consultant-class narrative of private-bank AI. Neither bank is automating the discretionary mandate. Both banks have deliberately constructed the AI layer as an advisor-in-the-loop input that supports the relationship manager's decision but does not execute it. The architectural choice is partly cultural — Swiss private banking has historically constructed itself around the proposition that the relationship is the product — and partly regulatory: FINMA's posture on AI in advisory contexts, articulated in a series of dialogue letters with the major Swiss private banks across 2024 and 2025, is that the AI layer cannot be the locus of decision-making authority. The decision-maker has to be the human relationship manager, and the bank's processes have to demonstrate this clearly to FINMA inspection.

In practice, the discretion model operates as follows. At Pictet, the AI recommendation is generated weekly per client and queued in the relationship manager's dashboard. The relationship manager reviews the recommendation — typically Tuesday morning, after the chief economist's letter has been released Monday afternoon — and either accepts the recommendation as-is, modifies it, or rejects it entirely. An acceptance rate of 100 per cent would be a red flag for FINMA, and the bank's internal monitoring is configured to surface relationship managers whose acceptance rates exceed an internally calibrated threshold (the threshold is not disclosed publicly, but multiple sources in the closed-door convenings confirmed it is well below 100 per cent). The relationship manager forwards the modified recommendation to the client either as a structured letter — the standard format — or as the basis for a face-to-face or video conversation, depending on the client's preference. The client confirms or modifies the recommendation, and the relationship manager then executes the cash-management instruction through the bank's standard trade desk. The AI layer is not in the execution path. The audit trail records: the AI recommendation as generated, the relationship manager's edits, the client's confirmation, and the executed trade. Four separate timestamps, four separate identities, one final cash position.

Lombard Odier's discretion model operates on the same advisor-in-the-loop principle but with a slightly different cadence. The AI recommendation is generated on demand rather than on a weekly schedule, triggered by the relationship manager when preparing for a client review or in response to a market event. The on-demand model lets the relationship manager pull a fresh recommendation when the rate environment has moved meaningfully since the last review, which the bank's data suggests happens roughly every two to three weeks across the average HNW client. The on-demand approach has produced higher per-client recommendation generation rates than Pictet's weekly schedule, but lower acceptance rates as a proportion of generated recommendations — a pattern the Lombard Odier team interprets as evidence that the on-demand model surfaces recommendations in moments where the relationship manager is more likely to find the recommendation incrementally useful rather than directionally novel. The audit trail at Lombard Odier follows the same four-timestamp structure as Pictet's, with the additional logging of the trigger event that prompted the recommendation generation.

The relationship managers at both banks have been the most important user-experience constituency in the pilot's design. Both banks ran extensive internal user research with their relationship manager cohorts before the pilot went live, and both teams have iterated the recommendation format multiple times in response to relationship manager feedback. The most significant iteration, common to both banks, was the move from a tabular recommendation format to a structured-letter format that the relationship manager can edit in place. The tabular format, tested in the first wave of the Pictet pilot in October 2025, was rejected by relationship managers because it produced output that did not fit the bank's letter-based client communication culture. The structured-letter format — approximately 400-600 words of recommendation prose, with a structured appendix of the underlying rate-curve and sovereign-credit inputs — has been more readily adopted because it slots naturally into the bank's existing client communication workflow. The user research underscores a structural truth: the AI layer's success in this context depends as much on cultural fit with the bank's existing operating model as on the quality of its underlying reasoning.

The tooling stack: Bloomberg AIM Pro, in-house RAG, and the integration layer

The named tools that sit alongside the Claude reasoning layer are, in some ways, more revealing than the reasoning layer itself. Both banks have built their AI-treasury systems on top of Bloomberg AIM Pro, the platform's AI-augmented portfolio management toolkit that Bloomberg released in beta in late 2024 and made generally available in mid-2025. AIM Pro provides the structured-data layer that the Claude reasoning layer consumes: rate curves, FX volatility surfaces, sovereign credit signals, and a daily-refreshed view of the bank's positions across all client cash mandates. The integration between AIM Pro and the in-house Claude deployment runs through Bloomberg's API layer, and both banks confirmed in the closed-door convenings that the AIM Pro integration has been the most operationally stable component of the system since pilot launch. The Bloomberg infrastructure is what the relationship managers already trust, and embedding the AI layer inside that infrastructure has been a quiet but important cultural enabler.

The in-house RAG layer at both banks runs on top of the Bloomberg infrastructure rather than alongside it. Pictet's RAG layer is built on a Pinecone vector index hosted in the bank's Azure tenant, with the index refreshed daily from the bank's research publication system and the FX desk's overnight commentary feed. Lombard Odier's RAG layer is built on a Weaviate deployment hosted in the Anthropic dedicated capacity environment, with the index refreshed continuously from the bank's research and sustainability archives. Both banks have invested significantly in the chunking and re-ranking layer that sits between the raw vector index and the Claude reasoning layer — Pictet's team estimates that approximately 40 per cent of their engineering effort on the pilot has been on retrieval quality rather than on model integration, and Lombard Odier's team has reported a similar ratio. The lesson, repeated across both teams in the closed-door convenings, is that retrieval quality is the binding constraint on system output quality in this kind of grounded-reasoning system. The model is good enough. The retrieval is the work.

The economist-letter grounding deserves particular attention. Both banks treat their chief economist's weekly letter as a load-bearing input to the system. The letter is the document that the relationship managers themselves read most carefully, and the bank's brand promise — particularly Pictet's, which has positioned its economist research as a differentiator for decades — depends on the relationship manager being able to discuss the bank's published view with the client in a credible way. The AI layer's recommendations are grounded on the economist letter in part to ensure consistency between the AI-generated recommendation and the bank's published view. A recommendation that contradicts the economist letter would be embarrassing in client communication and operationally awkward in audit. The retrieval layer is configured to up-weight the most recent economist letter and the relevant regional desk notes in any recommendation generation, with the result that the recommendations track the bank's published view tightly. The architectural choice has a cost — it constrains the system from generating recommendations that depart from the bank's house view — and a benefit — it ensures that the AI layer reinforces rather than contradicts the bank's economist franchise.

The execution-layer integration is the part of the architecture that both banks have been most cautious about. Neither bank's AI layer has direct execution authority. The recommendation that the relationship manager forwards to the client, and the client's confirmation, both flow back into the bank's standard trade execution system — for both Pictet and Lombard Odier, this is a combination of Bloomberg AIM execution capabilities and the bank's internal trade desk. The AI layer is read-only on the execution side. The architectural decision is partly regulatory — FINMA's outsourcing and operational risk requirements would be significantly more demanding for a system with direct execution authority — and partly philosophical: both banks have stated explicitly, in the closed-door convenings, that the discretionary trade is the locus of the bank's fiduciary responsibility, and the bank's leadership is not comfortable mediating that responsibility through an AI layer in the current generation of the technology. The position may shift over time. It has not shifted yet.

Rate-call accuracy and the early performance picture

The performance picture across the first six months of the pilot — October 2025 through March 2026 — is more nuanced than the simple accuracy framing the consultant-class narrative would apply. Neither bank publishes the AI-recommendation accuracy data publicly, and the closed-door convenings have produced figures that are necessarily second-hand and partial. The figures that surface most consistently across both banks are framed in terms of directional accuracy on the CHF and EUR rate calls embedded in the cash-management recommendations: that is, whether the recommendation correctly anticipated the direction of CHF or EUR short-rate moves over the following two weeks. Pictet's figure, as discussed at the February client convening, was 67 per cent directional accuracy on CHF short-rate moves and 64 per cent on EUR short-rate moves across the first 24 weekly recommendation cycles. Lombard Odier's figure, discussed at the March convening, was 61 per cent on CHF and 60 per cent on EUR across a comparable measurement window. Both banks framed the figures as preliminary, and both noted that the measurement window includes the unusually low-volatility CHF environment of late 2025, which makes directional accuracy easier to achieve than it would be in a more volatile environment.

The figures are useful but limited. Rate-call directional accuracy is one input to the broader question of whether the AI-treasury system produces better cash-management outcomes than the relationship manager would produce without it, and the rate-call figure does not capture the system's contribution to allocation across the FX dimension, the sovereign-credit selection, or the bank-deposit-versus-money-market trade-off. Both banks have committed to publishing more comprehensive performance data after the first full year of the pilot — Pictet has signalled a Q4 2026 publication target, Lombard Odier a Q1 2027 target — and the more comprehensive data will give the asset-management community a better basis for evaluating the system's contribution to client outcomes. The early figures, taken on their own, support the proposition that the system is producing useful directional inputs in a benign environment. They do not yet support stronger claims.

A more textured signal comes from the relationship-manager-side modification data, which both banks have been more willing to discuss. Across the first 24 weeks of the Pictet pilot, the average relationship manager modified the AI recommendation in approximately 38 per cent of cases, accepted it as-is in 51 per cent, and rejected it entirely in 11 per cent. The modification rate has been stable across the measurement window, suggesting that the relationship managers are integrating the AI layer into their workflow rather than either capitulating to it or rejecting it wholesale. The acceptance-as-is rate has crept upward slightly over the measurement window — from 47 per cent in the first eight weeks to 54 per cent in the most recent eight weeks — which the bank's internal team attributes to the retrieval-layer improvements and the iterative refinement of the recommendation prompt. The rejection rate has remained roughly stable. Lombard Odier's equivalent figures are similar in shape but skewed slightly more toward modification (44 per cent modified, 43 per cent accepted, 13 per cent rejected), consistent with the bank's on-demand recommendation model producing recommendations in moments where the relationship manager is more likely to want to fine-tune.

The client-side feedback, captured in both banks' standard client-satisfaction instruments, has been quietly positive but not dramatic. Pictet's NPS for the cohort of clients whose cash management is being supported by the AI layer rose by approximately 4 points relative to the prior twelve-month rolling average, with the gain driven primarily by clients who specifically valued the increased frequency and depth of cash-management commentary in their relationship manager's communications. Lombard Odier's equivalent figure was a 5-point NPS improvement in the AI-supported cohort relative to a matched-control cohort whose relationship managers did not have access to the AI layer during the pilot. Both gains are modest. Both are statistically meaningful for the cohort sizes involved (approximately 400 clients at Pictet, approximately 380 at Lombard Odier). The interpretation common across both banks is that the AI layer is improving the relationship manager's capacity to deliver structured cash-management thinking to clients who value structured cash-management thinking, and that the system's contribution is being absorbed by the relationship — which is, of course, exactly what the discretion model was designed to enable.

The FINMA dialogue and the regulatory architecture

FINMA's position on AI in advisory contexts has been articulated more clearly across 2025 than the public communications would suggest. The regulator has issued no formal circular specific to AI in private banking, but the dialogue that has run between FINMA and the Swiss Bankers Association across 2024 and 2025 has produced a working understanding that both Pictet and Lombard Odier have used as the design substrate for their pilots. The working understanding has three pillars. First, the decision-making authority must reside with a named human individual within the bank, not with the AI system. Second, the audit trail must capture both the AI output and the human modification with sufficient granularity to support inspection. Third, the bank must maintain demonstrable capacity to operate without the AI system in the event of operational failure — meaning the relationship managers must remain competent at producing cash-management recommendations without the AI input. The third pillar is the one that has shaped the design choices most quietly: both banks have explicit programmes that require relationship managers to produce recommendations without AI assistance on a rotating basis, as a competency-maintenance measure.

Pictet's FINMA dialogue has run through the bank's Chief Compliance Officer, Beatrice Lehnert (named in the bank's annual report and in the relevant FINMA correspondence summaries that the bank has shared with its supervisory board). Lehnert's team has been in continuous dialogue with FINMA's Risk Management Division since the bank notified the regulator of the pilot in September 2025, and the bank's internal compliance review of the pilot — completed in January 2026 — has been shared with FINMA in summary form. The dialogue has been collaborative rather than adversarial, by both the bank's account and by the regulator's own published characterisations of its engagement with the Swiss private banks. FINMA has not required structural changes to the pilot, but it has required quarterly performance reporting and an external audit of the audit trail's integrity. The external audit was completed in March 2026 by KPMG and produced no material findings.

Lombard Odier's FINMA dialogue has run on a similar cadence, with the bank's Group General Counsel and Chief Compliance Officer jointly leading the engagement. The bank's specific design choice to run inference through Anthropic's dedicated capacity arrangement required additional outsourcing-disclosure documentation to FINMA under the regulator's outsourcing framework, and the bank invested several months of compliance and legal work in 2025 to satisfy the documentation requirement. The arrangement is now in place and operating, and FINMA has not raised concerns about the outsourcing structure since the documentation was completed. The bank's external audit, conducted by PwC in early 2026, similarly produced no material findings. The clean audit posture across both banks suggests that the regulatory architecture for this kind of system is now reasonably mature in Switzerland, and that other Swiss private banks considering similar pilots have a working precedent to follow.

The broader regulatory implication is one the Swiss private banking community has been watching carefully. The Pictet and Lombard Odier pilots establish a working pattern — Claude-grounded reasoning, advisor-in-the-loop discretion, structured audit trail, quarterly performance reporting, external audit — that the rest of the Swiss private banking sector can adopt without inventing the regulatory dialogue from scratch. The pattern is being closely studied by Mirabaud, J. Safra Sarasin, and Union Bancaire Privée, each of which has signalled internally that an AI-treasury pilot is on the 2026 roadmap. The early indications, from conversations at the Swiss Private Banking Forum in February 2026, are that the next wave of pilots will follow the Pictet/Lombard Odier architecture closely rather than attempt structurally different designs. The regulatory pattern is, in this sense, becoming a market standard before the market has matured.

Cross-currency dynamics and the FX hedging contribution

The FX hedging contribution of the AI-treasury layer has been one of the more interesting performance dimensions across the first six months of operation. Both Pictet and Lombard Odier serve HNW client bases with material cross-currency exposure — a Swiss-resident client whose primary income flows are USD-denominated, a German-resident client with CHF-denominated assets, an Italian client with both EUR and USD positions. The cash-management recommendation layer addresses the FX hedge ratio as a standard output, and the recommendation has been the part of the system that the relationship managers have most consistently engaged with as a value-added input to their client conversations. The reasoning: FX hedging decisions sit in a territory where the relationship manager's intuition is often more uncertain than on the pure-rate side, and the structured analytical framework that the AI layer provides has been welcomed as a complement to the relationship manager's existing judgement.

The specific FX hedge ratios that the system has produced across the first six months have been more conservative than the relationship managers, in aggregate, would have produced without the system. The Pictet team's internal review of the FX hedge recommendations during the February quarterly performance check found that the AI layer's recommended hedge ratios averaged approximately 8 percentage points higher than the relationship-manager baseline, with the structured rationale typically emphasising the convexity properties of the underlying FX exposure and the risk-adjusted cost of hedging at current volatility levels. The relationship managers have, in approximately 60 per cent of cases, accepted the more conservative recommendation; in approximately 30 per cent of cases, they have modified the recommendation downward; in approximately 10 per cent, they have rejected the recommendation entirely. The pattern produces, across the AI-supported client cohort, a measurably more hedged FX position than the bank's non-AI-supported cohort, an outcome that the bank's risk committee has reviewed and found to be consistent with the bank's broader risk-management posture.

Lombard Odier's FX hedging contribution has followed a similar pattern with slightly different parameter calibration. The bank's on-demand recommendation generation produces FX hedge updates more frequently than the Pictet weekly cadence, and the relationship managers at Lombard Odier have generally accepted the AI layer's hedge recommendations at higher rates than at Pictet — partly because the on-demand model surfaces the recommendation in moments when the relationship manager is already actively engaged with the FX question, and partly because the Lombard Odier recommendation prompt has been tuned to produce slightly less conservative hedge ratios than the Pictet equivalent. The two banks' parameter-tuning differences are a small but meaningful illustration of how the underlying recommendation logic can be calibrated to fit the bank's specific client base and risk appetite while preserving the broader architectural pattern.

The cross-currency dynamics also reveal an emerging market-microstructure question that neither bank has fully addressed. The aggregate effect of the AI-treasury systems across the Swiss private banking community — assuming the systems continue to scale and to recommend broadly similar hedge ratios — would be to produce a measurably more hedged FX flow across the major HNW client jurisdictions. The aggregate effect is small in absolute terms relative to the broader FX market, but it is large enough to be visible in the cross-border FX flow data that the Swiss National Bank monitors, and the SNB has been observing the pattern. The regulatory implication is not yet clearly articulated, but the bank's chief economists and the SNB's monetary-policy staff have been in regular contact about the pattern, and the bank's expectation is that the SNB will eventually issue some form of guidance on the aggregate behaviour of AI-augmented FX hedging recommendations across the Swiss private banking sector. The guidance has not arrived. The dialogue has begun.

What to watch

The Pictet and Lombard Odier pilots have completed their first six months of operation. The next twelve months will tell us whether the architecture survives contact with higher-volatility rate environments and broader-scale deployment.

• Whether the rate-call directional accuracy figures hold up in a higher-volatility CHF or EUR environment; the first six months of the pilot have run through an unusually low-volatility short-rate environment in both currencies, and the system's accuracy figures have not been stress-tested by the kind of multi-week directional move that the FX desks at both banks regard as the more interesting test case.
• Whether the relationship-manager acceptance rates continue to drift upward without crossing the threshold that would trigger FINMA concern about over-reliance; the current Pictet figure of 54 per cent acceptance-as-is in the most recent eight weeks is well below FINMA's red-flag threshold, but a sustained upward drift would change the regulatory conversation and might require the banks to implement additional friction in the workflow to maintain demonstrable human authority over the decision.
• Whether Mirabaud, J. Safra Sarasin, and UBP launch comparable pilots in 2026 on the same architectural pattern, or whether one or more of those banks attempts a structurally different design — a different model provider, a different discretion model, a different audit architecture — that would tell us the Swiss private banking sector views the Pictet/Lombard Odier pattern as one option among several rather than as the emerging market standard.
• Whether either bank moves toward giving the AI layer direct execution authority on any subset of the cash-management mandate — for example, a low-stakes daily rebalancing routine — or whether the read-only execution model is preserved through the second year of operation; the move to execution authority would require a significantly more demanding FINMA dialogue and a higher-burden audit architecture, and both banks have been clear that this is not on the near-term roadmap, but the question will return.
• Whether the Q4 2026 publication of Pictet's performance data — and the Q1 2027 publication of Lombard Odier's equivalent — produces the kind of detailed performance disclosure that the asset-management community has been asking for from the AI-in-advisory-products cohort, or whether the published data is structured in a way that supports the bank's marketing of the product more than it supports an independent assessment of the system's contribution to client outcomes.

Frequently asked

Why did Pictet and Lombard Odier choose Claude Opus 4.7 rather than a different model provider?

Both banks confirmed in the closed-door convenings that the Opus 4.7 1M-token context window was the critical enabling capability for their architecture, because the grounded retrieval substrate the system needs in any single reasoning step routinely exceeds 200,000 tokens. Competing models with shorter context windows would have required either a multi-step retrieval-and-summarisation architecture — which would have introduced summarisation-loss in a critical input chain — or a smaller retrieval substrate that would have degraded the grounding quality. Both banks also cited Anthropic's posture on enterprise data handling, the company's published research on model interpretability and safety, and the operational responsiveness of Anthropic's enterprise team as relevant factors. The technical capability was the binding constraint; the enterprise relationship was the closing argument.

What is the difference between Pictet's self-hosted Azure deployment and Lombard Odier's Anthropic dedicated capacity arrangement?

Pictet's Azure deployment runs the Claude inference inside a Pictet-controlled Microsoft Azure tenant with the inference region pinned to Frankfurt, giving the bank direct control over the inference environment and the logging architecture. Lombard Odier's dedicated capacity arrangement runs the inference in an Anthropic-managed environment with contractually defined data-handling guarantees, provisioned throughput, and dedicated logging that flows back to the bank under a service-level agreement. The Pictet choice maximises bank-level control at the cost of higher operational overhead; the Lombard Odier choice trades some control for lower operational burden and tighter integration with Anthropic's release cadence. Both architectures satisfy FINMA's outsourcing requirements; the satisfying documentation is more involved for the Lombard Odier arrangement.

How does the system handle client data and confidentiality?

Both banks have constructed the system so that client-identifying data does not flow into the reasoning layer's prompt. The per-client recommendation is generated by the relationship manager submitting an anonymised client profile — investment objectives, risk tolerance, currency exposures, liquidity preferences, ESG preferences — that the reasoning layer consumes alongside the market and research inputs. The relationship manager's local system, not the AI layer, holds the linkage between the anonymised profile and the named client. The audit trail records the anonymised profile and the recommendation; the named-client linkage is recorded separately in the bank's standard client-relationship management system, which has its own audit and access controls. The architectural separation is designed to satisfy Swiss banking secrecy requirements while still letting the AI layer produce a per-client-relevant recommendation.

What happens if the AI recommendation is wrong and the client loses money?

The bank's fiduciary responsibility, both legally and operationally, attaches to the relationship manager and the bank, not to the AI layer. The relationship manager has reviewed and modified the recommendation, the client has confirmed the modified recommendation, and the execution flows through the bank's standard trade desk. If the recommendation underperforms — or in extreme cases produces a loss the client did not anticipate — the bank's response is governed by its standard discretionary mandate framework. The AI layer's output is not treated as advice in a legal sense; it is treated as an internal decision-support input, analogous to a research note or a desk commentary. The FINMA dialogue has confirmed this characterisation, and both banks have built the audit trail to support it.

Why is the relationship-manager modification rate so high (38-44 per cent) — does that indicate the AI layer is unreliable?

The modification rate is intentionally high and is read by both banks as a sign that the architecture is working as designed. A relationship manager who accepts the AI recommendation as-is more than the calibrated threshold percentage of the time is operating outside the discretion model, and the bank's internal monitoring flags those relationship managers for review. The 38-44 per cent modification rate reflects the relationship managers actively integrating the AI input into their own judgement rather than either rubber-stamping it or rejecting it wholesale. The figure is therefore a measure of healthy advisor-in-the-loop dynamics, not of AI unreliability. The acceptance-as-is figure (around 50 per cent at both banks) is closer to a measure of recommendation quality, and both banks have been comfortable with that figure given the iterative refinement of the system over the six-month measurement window.

Is this architecture replicable for non-Swiss private banks, or is it specific to the Swiss regulatory environment?

The architecture is broadly replicable to any regulatory environment that supports advisor-in-the-loop discretion models. The Swiss specifics — FINMA's particular outsourcing framework, the Swiss banking secrecy requirements, the bank's culture of relationship-anchored client communication — shape the implementation details, but the underlying pattern (grounded LLM reasoning, advisor-in-the-loop modification, structured audit trail, read-only execution) maps reasonably onto the EU MiFID II framework, the UK FCA's posture on AI in financial services, and the US SEC's posture on AI in investment advisory contexts. The European private banks — particularly those with Swiss subsidiaries — are expected to adopt similar architectures over the next twelve to eighteen months. The US wealth management cohort is moving more slowly, partly because the US private banking culture is less anchored on the kind of relationship-based discretionary mandate that the Swiss architecture is built around.