Conversational AI in Zurich addresses private banking client service, pharmaceutical adverse event intake, insurance and wealth management where Swiss German alongside standard German, French, Italian and English must be handled, banking secrecy enforced per role and authorisation, and initial adverse event capture distinguished from accountable safety assessment. FINMA scope, FinSA service-type-dependent suitability and FADP data protection principles drive architecture.
ZTABS is remote-first by design with an office in Houston, Texas. Our conversational AI services start with a knowledge source inventory, FINMA scope assessment and FADP basis review before system design. ZTABS builds custom white-label AI products for clients to brand and resell.
Sector landscape
| Sector | System scope | Key constraint |
|---|
| Private banking | Client service, portfolio info, settlement checks | Banking secrecy Art. 47 per role; FinSA suitability by service type |
| Pharma | Adverse event intake, medical info, knowledge management | Initial capture only; accountable safety assessment by MAH PV function |
| Insurance | Policy info, claims status | No coverage determinations |
| Wealth management | Client service, mandate status | FinSA suitability and appropriateness by service and exceptions |
| Multilingual | Swiss German, standard German, French, Italian, English | Agreed tested patterns; standard German fallback for uncertain input |
Buyer profiles and use cases
Private banking with mandate verification
Zurich private banks serve clients across language regions and internationally. The system handles read-only client service: portfolio positions, transaction status, document requests and settlement verification from the bank's secure systems. Banking secrecy under Article 47 Banking Act is enforced per role and authorisation: the system restricts information display to what the authenticated user is authorised to see, not as a blanket guarantee of non-disclosure. FinSA suitability and appropriateness requirements depend on the type of financial service (execution-only, advisory, portfolio management) and applicable exceptions. The system identifies advisory queries and routes to the client's relationship manager with minimum redacted context. Multilingual service covers German, French, Italian and English per client preference.
Pharmaceutical adverse event intake
Zurich pharmaceutical companies require systems for adverse event intake from healthcare professionals. The system captures initial reports: reporter information, patient demographics (age, sex), suspect product (name, dose, route), event description, temporal relationship and outcome where known. This is initial data capture through structured fields, not safety assessment. The system does not assess causality, determine seriousness, classify expectedness or make medical judgments. Reports route to qualified safety personnel for processing into complete valid ICSRs for Swissmedic reporting. The accountable safety assessment, MedDRA coding, causality assessment and regulatory reporting are performed by the marketing authorisation holder's pharmacovigilance function.
Wealth management with FinSA scope
Wealth management firms provide client service where FinSA suitability and appropriateness requirements depend on service type. For execution-only services, suitability assessment may not apply. For advisory and portfolio management services, appropriateness and suitability requirements apply with different scope. The system provides factual information from approved content and identifies advisory queries for routing to qualified advisers.
Insurance customer service
Insurers provide read-only policy information, claims status and coverage explanations from approved content. The system does not make coverage determinations or assess claim validity.
Compliance scope and design constraints
The FADP (nDSG) is technology-neutral. It establishes data protection principles (purpose limitation, proportionality, accuracy, transparency) and notice requirements with defined exceptions. Articles 30 and 31 address private personality rights and justification grounds. Article 6(7) requires explicit consent for specific categories where consent is the relied-upon basis. Articles 16 and 17 provide scoped safeguards and exceptions for outbound transfers. The FADP is not characterised by a blanket information-and-opt-out approach contrasted with GDPR consent.
Article 21 requires the controller to inform the data subject about exclusively automated decisions with legal or significant effect. The data subject has the right to state their position and, on request, obtain review by a natural person. Exceptions: directly connected to contract conclusion or performance where the request is granted, or explicit consent.
FINMA Guidance 08/2024 addresses AI governance and risk management for supervised institutions. Circular 2023/1 covers operational risk and resilience for applicable supervised institutions. Circular 2018/3 addresses material outsourcing for scoped institutions. Banking secrecy (Article 47) protects client information per role and authorisation, not as a blanket assurance. FinSA suitability and appropriateness depend on service type and applicable exceptions.
LLM provider transcript retention and training terms are documented. Processing location (Switzerland versus international) is specified for FADP transfer provisions.
System architecture
Swiss multilingual NLP
NLU for standard German with Swiss German expression handling for agreed tested patterns. French, Italian and English per client or employee base. Standard German fallback for uncertain Swiss German input. Domain-specific terminology validated per language.
Private banking client service
Avaloq, Temenos or equivalent core banking integration for read-only portfolio data, transaction status and settlement verification. Banking secrecy enforced per authenticated user's role and authorisation. Relationship manager routing for advisory queries with minimum redacted context (query category, client preference, no full transcript). Multilingual service per client preference.
Pharma adverse event intake
Adverse event intake is a controlled write in an approved secure channel. Structured intake fields: reporter, patient demographics, suspect product, event description, temporal relationship, outcome. Minimum valid ICSR per ICH E2D(R1) section 2.2: identifiable reporter, identifiable patient, suspect or interacting medicinal product and suspected adverse reaction; separate from local reportability and other ICH observation categories. Missing fields do not block escalation. The system verifies handoff receipt and reconciles duplicates against existing reports. Initial awareness, complete report and regulator-submitted case are distinguished. No causality assessment, seriousness determination or MedDRA coding. The MAH's pharmacovigilance function owns assessment and reportability. GxP documentation where the system affects regulated processes.
Settlement and mandate verification
Dated position display from core banking system with timestamp. Settlement status verification against the authoritative system. Readback confirmation for status queries (the system displays what the system of record shows, not a cached or stale value).
Review your pharmacovigilance intake and FINMA scope
Tell us which knowledge domains, FINMA scope and pharmacovigilance requirements apply. We document banking secrecy role boundaries, FADP basis and adverse event intake scope before architecture.
Delivery and testing
ZTABS delivers remote-first. Testing covers: Swiss German expression comprehension for tested patterns; standard German fallback accuracy; banking secrecy role-based access enforcement; FinSA advisory boundary identification; adverse event intake field completeness; settlement verification against authoritative system; Article 21 notice for automated decisions where applicable; scope boundary and adversarial testing; and fallback behaviour.
Illustrative Example
Pharma adverse event intake with initial capture and safety personnel routing
If a Zurich pharmaceutical company were to scope a conversational AI for adverse event intake, the problem would be that healthcare professionals currently report suspected adverse reactions by phone, email or web form, with inconsistent data quality across channels. Safety personnel spend significant time following up for missing information before they can process reports into valid ICSRs for Swissmedic.
Prerequisites would include: adverse event intake structured fields aligned with ICH E2B(R3) standards (reporter information, patient demographics, suspect product details including dose and route, event description, temporal relationship, outcome where known); routing workflow to qualified safety personnel for causality assessment, MedDRA coding and ICSR completion; multilingual support (German, French, Italian, English) for healthcare professionals across Swiss language regions; FADP processing documentation for health-related personal data; GxP documentation requirements where the system affects the pharmacovigilance process; and evaluation metrics for intake field completeness, routing accuracy and follow-up reduction.
The system would capture initial adverse event data through structured questions in the healthcare professional's preferred language. For each field, the system would prompt specifically: suspect product name and strength, dose and route of administration, start and stop dates, event description in the reporter's own words, temporal relationship (when did the event start relative to the medication), patient age and sex, and outcome (recovered, recovering, not recovered, fatal, unknown). Where the reporter cannot provide a field (e.g., exact dose), the system would record "not provided" rather than guessing. The completed initial intake would route to qualified safety personnel with the structured data. The system would not assess causality (not related, possibly related, probably related, definitely related), determine seriousness (meets ICH seriousness criteria or not), classify expectedness against the product label or code the event in MedDRA. These are the marketing authorisation holder's pharmacovigilance function responsibilities.
Held-out evaluation would include: reports with minimal information (reporter knows only the product name and one symptom), reports in mixed languages (e.g., German medical terminology with French colloquial description), reports that describe product quality issues rather than adverse events (which route differently), and reports where the reporter wants medical advice (which the system cannot provide). If intake completeness falls below threshold, fallback is phone intake with safety personnel for affected product categories.
If the company updates its product portfolio (new product launch, formulation change), the suspect product database would be updated within the documented cycle. If the safety personnel routing system is temporarily unavailable and an approved encrypted queue is configured, the system places the intake into the queue with timestamp and provides a durable local receipt with reference number. If no approved queue is configured, the system routes to the secure urgent staff contact. The system verifies handoff receipt and reconciles against existing reports before any retry. If a reporter submits a follow-up report for a previously reported event, the system would link it to the original report reference if provided, or create a new record for safety personnel to reconcile.
Residual risks: intake field completeness for reporters with limited information, language-specific medical terminology mapping across four languages, product quality issue versus adverse event classification, GxP audit trail requirements for the intake process, FADP basis for health-related personal data processing, and LLM provider transcript handling for conversations containing patient and product information.
This is a hypothetical scenario for illustration. Actual scope, timeline and cost depend on your requirements.
Project costs and planning benchmarks
Costs reference the AI chatbot cost guide. The four published tiers are:
| Tier | Range | Timeline | Scope |
|---|
| Basic FAQ Chatbot | $10K to $25K | 4 to 8 weeks | Pre-defined responses, simple NLU, FAQ matching, basic UI widget |
| RAG Knowledge Base Bot | $25K to $60K | 8 to 16 weeks | Document ingestion, vector search, context-aware answers, source citations, conversation memory |
| Multi-Channel AI Agent | $60K to $100K | 16 to 24 weeks | WhatsApp, Slack, email integration, tool calling, CRM integration, handoff to humans |
| Enterprise AI Platform | $100K to $150K+ | 24 to 36 weeks | Multi-agent orchestration, custom fine-tuning, analytics, compliance, self-hosted LLM option |
2026 US-buyer benchmark for a RAG knowledge base bot
These planning references reflect a RAG-powered knowledge base bot over approximately 5,000 documents with web widget, OpenAI or Anthropic LLM, citations and CRM handoff.
| Vendor type | Typical cost | Timeline | What to check |
|---|
| Freelancer or GPT-wrapper specialist | $6K to $25K | 3 to 8 weeks | Prompt quality, evaluation harness and hallucination control |
| Offshore agency | $15K to $45K | 8 to 16 weeks | Retrieval quality tuning and cost controls |
| Nearshore agency | $25K to $75K | 6 to 14 weeks | Timezone alignment and depth in RAG evaluation |
| Specialist AI consultancy (ZTABS tier) | $35K to $120K | 6 to 14 weeks | Eval-driven development and observability approach |
| Off-the-shelf platform | $0 to $15K setup | 1 to 4 weeks | Ceiling on custom tools, private data and domain-specific tone |
Ongoing operational costs are separate from project development. LLM API costs run $100 to $5,000 per month. Vector database hosting runs $25 to $500 per month. Self-hosted GPU servers, when selected, add $500 to $3,000 per month. These are planning references, not engagement-specific quotes.
Price the pharmacovigilance intake and banking integration
Describe your FINMA scope and pharmacovigilance requirements. We provide a budget estimate referencing the published AI chatbot cost guide with ongoing operational costs and LLM provider terms estimated separately.
When ZTABS is not the right fit
- The engagement requires investment advice, portfolio recommendations or suitability determinations. ZTABS builds the information system; advisory activities are the institution's duty under FinSA by service type.
- The project requires adverse event causality assessment, seriousness determination or ICSR completion. ZTABS builds the intake system; safety assessment is the MAH's pharmacovigilance function.
- The project requires Swiss regulatory licensing or compliance determinations. ZTABS builds the system; regulatory determinations require qualified Swiss professionals.
- ZTABS is remote-first by design. Engagements requiring permanent on-site embedded staff are outside this scope.
Risk register
| Risk | Mitigation approach |
|---|
| Swiss German expression gaps | Tested patterns documented; standard German fallback; ongoing corpus expansion |
| Banking secrecy enforcement | Art. 47 enforced per role and authorisation; role-based access testing; no blanket guarantees |
| Adverse event intake completeness | Structured field prompts; "not provided" for missing data; follow-up linked by reference |
| FinSA suitability boundary | Service-type-dependent; advisory routing to relationship manager; execution-only scope documented |
| FADP automated decision (Art. 21) | Notice, position-stating and review opportunities implemented where applicable |
Proposal evaluation criteria
| Criterion | What the proposal specifies |
|---|
| 1. Deliverables and IP | Custom-developed system, core banking integration, adverse event intake, multilingual NLU and monitoring dashboards. Ownership and licence terms specified. |
| 2. Evaluation and holdout method | Swiss German comprehension, banking secrecy role enforcement, adverse event field completeness, FinSA boundary identification and held-out evaluation per language. |
| 3. Deployment and data scope | Infrastructure documented per FADP. FINMA scope where applicable. GxP documentation for pharma. Processing location specified. LLM provider transcript terms documented. |
| 4. Cost structure | Total project cost with milestone payments. Ongoing LLM API, infrastructure and knowledge maintenance estimated separately. |
| 5. Acceptance and fallback | Defined accuracy thresholds per language. Banking secrecy enforcement. Adverse event intake completeness. Documented fallback. Sign-off by named client staff. |
| 6. Roles, languages and handover | Provider delivers development, core banking integration, adverse event intake, multilingual testing and documentation. Proposal lists staffing languages and tested locales (German, French, Italian, English) separately. Customer provides core banking access, pharmacovigilance workflow, domain expertise and owns all financial, pharmaceutical, regulatory and safety decisions. |
Plan a pharmacovigilance intake pilot with FINMA scope
We map banking secrecy role boundaries, document adverse event intake scope and assess FADP basis in a focused initial call.