--- title: "AI Pilot to Production in a Bank: 9 Go/No-Go Questions - S-PRO" url: https://s-pro.io/blog/ai-pilot-to-production-bank lang: en description: "Before a bank scales an AI pilot, nine go/no-go questions need an evidence artefact each. Use this checklist with FINMA, FADP and EU AI Act references." updated: 2026-10-29T05:00:47+00:00 source: S-PRO (https://s-pro.io) --- [Back to insights](https://s-pro.io/blog) Article # AI Pilot to Production in a Bank: 9 Go/No‑Go Questions Author [Igor Izraylevych](https://s-pro.io/author/sproadmin) CEO and Co-founder of S-PRO Published October 29, 2026 Time to read 12 min Table of content - Key takeaways - Table of Contents - What Swiss evidence says about the pilot-to-production gap - Why bank AI pilots stall - The nine go/no-go questions and the evidence for each - The EU AI Act dates your timeline should respect - Which workflows reach production first - Our perspective - Taking a pilot live with S-PRO - FAQ - Sources Share Before a bank scales an AI pilot, it should be able to answer nine go/no-go questions, each backed by an evidence artefact someone can open and review. The search “which consultancy” misses the point: this is the bank’s own checklist, and it exists whoever builds the system. If you want to know who can take a bank AI agent from pilot to production, we answer that once, on a separate page. *Last updated: 27 October 2026* S-PRO Move Bank AI From Pilot to Production S-PRO builds AI agents for financial institutions and helps move AI pilots into production systems integrated with existing infrastructure. [Explore S-PRO’s AI services](https://s-pro.io) ## Key takeaways - A pilot proves a model can work. Production proves the bank can own it: inventory, accountable people, data, tests, fallback, explainability, outsourcing, data location and human oversight. - In Switzerland the gap is visible. In the SFTI/OST study reported by SwissBanking, about 40% of financial-sector AI initiatives were still in ideation or feasibility, 17% had reached deployment and 11% scaling. - Production is not the weak point once cases get there: 65% of production cases met or exceeded expectations in the same report. - FINMA Guidance 08/2024 sets out what the supervisor expects on AI governance. The EU AI Act adds dated obligations for banks with EU exposure. - We found no Swiss survey of pilot-to-production conversion rates, so treat any single “percent of pilots fail” figure with caution. ## Table of Contents - Why bank AI pilots stall - Which workflows reach production first - Our perspective - Taking a pilot live with S-PRO - FAQ - Sources ## What Swiss evidence says about the pilot-to-production gap The most relevant local data comes from a study led by SFTI and OST. [SwissBanking reported in December 2025](https://www.swissbanking.ch/en/media-politics/blog-posts/from-pilots-to-production-scaling-ai-in-swiss-finance) that about 40% of AI initiatives in the Swiss financial sector remained in ideation or feasibility, 17% had reached deployment and 11% scaling. The same report said 65% of production cases met or exceeded expectations. It linked better outcomes to defined business cases, strong data foundations and compliance-by-design, and noted that institutions are becoming more selective, with fewer but better-prepared initiatives. A second Swiss data point concerns private banks. In KPMG’s survey, [27 of 34 responding banks had deployed AI](https://assets.kpmg.com/content/dam/kpmgsites/ch/pdf/kpmg-private-banking-ai-advancement-index-26.pdf.coredownload.inline.pdf) in at least one operational setting, and 20 had done so across multiple areas. The same survey shows how shallow much of that deployment is: 26 of 34 respondents were still at ad hoc or developing maturity, and only two described themselves as optimizing with several production use cases and continuous improvement. Deployment was concentrated on employee productivity and document automation, followed by compliance, KYC, AML and monitoring. Client-facing and investment applications were less widespread. What we did not find is a Swiss survey that measures how many pilots convert to production. Figures such as “88% of pilots never reach production” circulate widely, but they come from vendor content without a Swiss base. The honest reading of the Swiss evidence is narrower: many initiatives are early, and those that reach production tend to do well. ## Why bank AI pilots stall Pilots rarely fail because the model is weak. Academic work points to organizational and structural causes. An [Oxford white paper on AI preparedness](https://www.sbs.ox.ac.uk/sites/default/files/2025-05/OFFTI-ai-preparedness-white-paper.pdf) describes pilots getting stuck without leadership backing, without a suitable tech stack, with unresolved data workarounds, and without business-user adoption. It also notes that time savings may not turn into value unless roles or work are redesigned. A [peer-reviewed study on banking data management](https://www.tandfonline.com/doi/full/10.1080/0960085X.2026.2723258) names legacy IT, siloed or low-quality data and demanding regulation as structural reasons banks fail to scale AI beyond isolated pilots. The [Financial Stability Board](https://www.fsb.org/uploads/P151026-1.pdf) lists third-party dependency and concentration, cyber risk, model risk, data quality and governance among the important AI-related vulnerabilities for financial institutions. It also treats AI risk as systemic, because common external providers and correlated models can concentrate failures across institutions. Each of these causes maps to a question a bank can answer before launch. That is the purpose of the checklist below. ## The nine go/no-go questions and the evidence for each The table gives each question, the artefact that answers it, the regulatory reference that makes the question reasonable to ask, and the bank role that should own it. The references are starting points for your compliance and legal teams, not a legal opinion. Where we write “expects”, we mean the FINMA text sets out expectations for supervised institutions. The guidance records supervisory findings; it does not explain why any particular pilot failed. | Question | Evidence artefact | Source | Bank owner | | --- | --- | --- | --- | | 1. Is the AI application in a central inventory, with its purpose and risk classification? | Inventory entry with use case, model, data sources and materiality | [FINMA Guidance 08/2024](https://www.finma.ch/en/news/2024/12/20241218-mm-aufsichtsmitteilung-08-24/), which expects an AI inventory | COO or Head of Operational Risk | | 2. Is there a named owner and an independent reviewer? | Responsibility matrix and review sign-off | FINMA Guidance 08/2024, which expects clear accountability | Business line head and Head of Risk | | 3. Is the data fit for the use case? | Data lineage record, quality checks, representativeness assessment | FINMA Guidance 08/2024 | Head of Data | | 4. Is the system tested, and is drift monitored after go-live? | Test report on a held-out set, monitoring thresholds, alert history | FINMA Guidance 08/2024 | CIO or Head of Model Risk | | 5. Is there documentation, and a fallback if the system fails? | Technical documentation, runbook, manual fallback procedure tested once | FINMA Guidance 08/2024 | CIO and Head of Operations | | 6. Can the bank explain outputs to the people who rely on them? | Explainability note per use case, reviewer guidance | FINMA Guidance 08/2024 | Business owner and Compliance | | 7. Are outsourcing and exit covered? | Provider contract, subcontractor list, exit plan | [FINMA Circular 2018/3](https://www.finma.ch/en/documentation/circulars/) on outsourcing | Head of Procurement and Legal | | 8. Where is data stored and processed? | Data-location register, access model | FINMA Circular 2018/3 and Circular 2023/1, Rz 79 and 82 (these do not apply to all banks) | CISO and Data Protection Officer | | 9. Is there human oversight where an output affects a client or a decision? | Oversight procedure, override log, escalation path | FINMA Guidance 08/2024 and the FADP rules on automated decisions | Head of Compliance and business owner | Check the scope of each circular against your institution before you rely on it. FINMA circulars differ in how they apply to banks of different size and category. ### Inventory, ownership and review (questions 1 and 2) Most pilots live outside the formal model landscape: a team licenses a tool, builds a prototype and tests it on sample data. Production changes that. The bank needs the application in one inventory with a stated purpose, and it needs one person who answers for it. An independent reviewer, not the team that built the agent, confirms the evidence. The artefact is simple: an inventory record and a signed responsibility matrix. If neither exists, the answer to the first two questions is no. ### Data, tests and drift (questions 3 and 4) Pilot data is usually curated. Production data is not. The evidence here is a lineage record showing where each input comes from, plus a test report on data the system has not seen. After go-live, someone must watch for drift: inputs changing, accuracy decaying, alert volumes shifting. Sygnum’s experience with AI testing, [reported by QA Financial in September 2026](https://qa-financial.com/sygnum-ai-testing-is-a-measurement-problem/), frames this as a measurement problem. The reported agents help analysts separate true from false positives in negative-news screening, and propose KYC profile updates from reviews that can involve more than 40 documents, with a person reviewing and confirming. The pattern is useful: the output is measured against what a reviewer decides. ### Documentation, fallback and explainability (questions 5 and 6) A fallback that has never been exercised is a statement, not evidence. Run the manual process once with the agent switched off, and file the result. Explainability follows the same logic. For a KYC or AML agent, a reviewer needs to see which documents and rules led to a proposal. For an RM copilot, the adviser needs to see which sources a summary draws on. ### Outsourcing, exit and data location (questions 7 and 8) Most bank AI agents rely on an external model, a cloud platform or both. The FSB’s concern about concentration is the reason exit planning belongs in the go/no-go decision: if the provider disappears or changes terms, what happens to the process? Data location is a separate question with a separate register. For Swiss private banks, client identifying data is often the sensitive element, so record where prompts, logs and outputs are stored, not only where the model runs. The Swiss FDPIC states that the [FADP applies directly to AI](https://www.edoeb.admin.ch/en/ai-in-everyday-life) whenever personal data is processed. That means transparency, proportionality, a specific limited purpose, good faith and accurate data. It also says AI providers and users [must make the purpose, functionality and data sources transparent](https://www.edoeb.admin.ch/en/09112023-current-data-protection-legislation-is-directly-applicable-to-ai), and that high-risk processing requires a data-protection impact assessment. ### Human oversight (question 9) Under the FADP, a person affected by a significant automated individual decision [must be informed and may request review by a real person](https://www.edoeb.admin.ch/en/knowing-and-asserting-my-rights). Oversight should therefore be designed into the workflow: who reviews, what they see, how they override, and where the override is logged. A Swiss banking architecture described in a [Camunda case study on Finnova](https://camunda.com/case-studies/finnova/) combines process orchestration with agents that start defined tasks such as ID verification and background checks, and hand over to humans when needed. That is the shape most compliance teams accept. ## The EU AI Act dates your timeline should respect If the bank serves EU clients or the system’s output is used in the EU, the EU AI Act applies to the system itself. Regulation (EU) 2026/1744, the AI Omnibus, entered into force on 27 July 2026. Under the [European Commission’s AI Act page](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai), 2 August 2026 is the general application date, with exceptions. Stand-alone high-risk (Annex III) obligations apply from 2 December 2027. High-risk AI embedded in products (Annex I) applies from 2 August 2028. We do not classify use cases in an article. Check Annex III against the use case, with legal counsel, before you fix a production date. Include the result in the evidence for questions 1 and 6. ## Which workflows reach production first The KPMG data above shows where Swiss private banks have started: employee productivity, document automation, then KYC, AML and monitoring. This order is sensible. These workflows have a human reviewer already, a measurable baseline and a bounded data set. Client-facing advice and investment applications carry more risk and need stronger oversight evidence. If you are choosing the first use case to take live, favor one where the nine answers are easiest to produce. A workflow with clean inputs, a clear owner and a manual fallback will pass faster than an ambitious one. ## Our perspective We build AI agents for banks, and we see the checklist above as a design brief rather than an audit step. If the evidence artefacts are specified in the pilot’s first week, the production decision becomes a review of documents that already exist. If they are left to the end, they tend to become a project of their own. Three choices matter most. First, scope the pilot so it can be measured against a human baseline. Second, keep the agent inside the bank’s existing controls: it proposes, the reviewer decides, and every step is logged. Third, plan integration with core, AML and wealth systems from the start rather than replace them. Our teams work in Zurich and Łódź, and our solutions are designed with EU AI Act, FINMA and DORA requirements in mind. > *— Igor Izraylevych* ## Taking a pilot live with S-PRO S-PRO delivers AI pilot to production for banks, wealth managers and fintechs, including AI agents for banks, KYC onboarding agents and RM & wealth copilots. For what a production budget for a bank AI agent consists of, and what regulation adds to it, see our page on AI agent pilot to production cost. If you already have a pilot and want to test it against the nine questions, talk to us through [s-pro.io](https://s-pro.io/). ## FAQ ### How many AI pilots at banks reach production? We found no Swiss survey that measures pilot-to-production conversion. The nearest Swiss evidence is the SFTI/OST study reported by SwissBanking: about 40% of initiatives in ideation or feasibility, 17% in deployment and 11% in scaling. Figures claiming a precise failure rate usually come from vendor content without a stated Swiss base. ### Does FINMA require an AI inventory? FINMA Guidance 08/2024 expects supervised institutions to manage AI risks, including keeping track of their AI applications. Guidance is a statement of supervisory expectations, so your compliance team should read the text against your institution’s category and activities. ### Does the EU AI Act apply to a Swiss bank? It can, when the system is placed on the EU market or its output is used in the EU. Stand-alone high-risk (Annex III) obligations apply from 2 December 2027, and high-risk AI embedded in products (Annex I) from 2 August 2028. Check Annex III against the use case with counsel. ### Who should own the go/no-go decision? The business owner of the use case, with an independent reviewer from risk or compliance. IT and data teams supply evidence, but they should not be the only sign-off. The decision belongs on the same committee that approves other material operational changes. ## Sources - [From Pilots to Production: Scaling AI in Swiss Finance, SwissBanking](https://www.swissbanking.ch/en/media-politics/blog-posts/from-pilots-to-production-scaling-ai-in-swiss-finance) - [Swiss Private Banks in the Age of AI, KPMG](https://assets.kpmg.com/content/dam/kpmgsites/ch/pdf/kpmg-private-banking-ai-advancement-index-26.pdf.coredownload.inline.pdf) - [AI Preparedness in Established Financial Services Firms, Oxford](https://www.sbs.ox.ac.uk/sites/default/files/2025-05/OFFTI-ai-preparedness-white-paper.pdf) - [Scaling data management capabilities for AI, Taylor & Francis](https://www.tandfonline.com/doi/full/10.1080/0960085X.2026.2723258) - [Sound Practices for Responsible Adoption of AI, Financial Stability Board](https://www.fsb.org/uploads/P151026-1.pdf) - [FINMA Guidance 08/2024](https://www.finma.ch/en/news/2024/12/20241218-mm-aufsichtsmitteilung-08-24/) - [FINMA circulars](https://www.finma.ch/en/documentation/circulars/) - [Sygnum: Testing AI is a measurement problem, QA Financial](https://qa-financial.com/sygnum-ai-testing-is-a-measurement-problem/) - [AI in everyday life, FDPIC](https://www.edoeb.admin.ch/en/ai-in-everyday-life) - [Data protection legislation is directly applicable to AI, FDPIC](https://www.edoeb.admin.ch/en/09112023-current-data-protection-legislation-is-directly-applicable-to-ai) - [Knowing and asserting my rights, FDPIC](https://www.edoeb.admin.ch/en/knowing-and-asserting-my-rights) - [Finnova case study, Camunda](https://camunda.com/case-studies/finnova/) - [AI Act regulatory framework, European Commission](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) Subscribe Get technical insights, and industry news Thank you! Get ready! You will receive handpicked content right to your inbox. --- Link copied to clipboard [Get in touch](https://s-pro.io/get-in-touch)