Insights · AI proof of concept to production

Who can turn our AI proof of concept into a production system for a bank?

Claude Certified Services Partner(Anthropic)

Banks take a proof of concept (PoC) into production with an internal team, a consultancy, a platform vendor, a hyperscaler or a build partner; the choice depends on who must own the system after go-live. S-PRO, a Zurich-based build partner, takes one agent from pilot to a fixed-price production phase.

Last updated: 20 October 2026 · S-PRO, Zurich

S-PRO is a software engineering company (founded 2014, 250+ engineers) and a Certified partner in the Claude Partner Network. It takes one agent on one process through a design thinking pilot and a solution pilot to production. How much of the system depends on your own core, data and approval process shapes the choice of partner. Pilot prices: see the FAQ below.

2014Founded, headquartered in Zurich

250+Engineers in Zurich and Łódź

ISO27001 and 27701 certified

CPNCertified partner in the Claude Partner Network

Why do AI pilots at banks stall before production?

The most useful primary source is the Wavestone and University of St. Gallen study "Agentic AI in Financial Services" (published 9 July 2026). It is based on 30 semi-structured interviews with senior leaders at banks, insurers and technology providers in the DACH region, held between October 2025 and May 2026. It reports that most organizations remain in the early stages of adoption, that regulation remains the primary adoption barrier, and that governance will determine scale. Other blockers it lists are compute capacity and sovereignty concerns, data quality, workforce readiness and legacy technology. A Head of AI at a German bank says the biggest limitations are "the regulatory effort, the AI Act, GDPR, DORA". Source: Wavestone, Agentic AI in Financial Services.

The study does not say how many pilots reach production, and this page does not claim a failure rate: no primary figure exists.

One bank's own account shows a typical blocker from the inside. In Pictet's published Claude story, Steve Blanchet (Head of Group Technology Strategy and Innovation) says: "Employee access rights are something that we haven't been able to crack for quite some time, but now we are tackling it with Claude Code." Source: Anthropic customer story, Pictet.

What moves a PoC into production? Five gates

These five gates are S-PRO's own delivery view, not findings of the study. Each is a decision that needs a named owner before go-live.

Swipe to compare →

Gate	Question the bank has to answer	Evidence that closes it
1. Data access	Does the agent read governed production data, with the same access rights as the person it assists?	Access model reviewed by information security; read/write scope documented
2. Security and third-party review	Have information security and outsourcing risk reviewed the model route (vendor API, cloud region, bank-controlled tenant, on-premises)?	Approved hosting route; for EU-regulated entities, ICT third-party documentation
3. Governance and model risk	Who classifies the use case, and can risk and compliance see measured accuracy?	Use-case risk class (including EU AI Act), evaluation set, human-in-the-loop design, audit log
4. Integration	Does the agent run inside the real workflow and systems (CRM, core banking, document store, identity), not next to them?	Integration tested on production-like data; fall-back to the manual process
5. Ownership	Who runs, monitors and pays for it after go-live?	Named business owner, run and support model, monitoring, cost per volume

Pictet's story illustrates gates 1 to 3 from the bank's side: approval "took close work with the risk, information security, and data protection teams", and Xavier Meyer, Head of Cloud Engineering, describes a "rollout per persona and per kind of use case" as a way to manage the relationship with risk and compliance. The figures in that story (around 700 people with access to Claude Code and Claude Cowork) are Pictet/Anthropic-reported, and Pictet is not an S-PRO reference.

Barclays illustrates gate 5 from the operations side. According to Anthropic's announcement of 1 October 2026, Barclays' Colleague Knowledge Assistant has been adopted by more than 16,000 colleagues, and the bank is extending Claude to engineering and operations, with a target of Claude Code for 50% of its developers by the end of 2026. These figures are vendor-published. Source: Anthropic, Barclays scales Claude.

For what the pilot contains, the timeline and the prices, see How much does it cost to take an AI agent from pilot to production at a bank? To see which of the Financial Stability Board’s 12 AI practices a pilot can evidence before a production decision, see What do the FSB’s sound practices for AI mean for a bank deploying AI agents?

From PoC to scale: four stages and their exit criteria

The five gates above are decisions. In delivery they fall into four stages, and each stage should end with a check the bank can run itself before it releases budget for the next one. This is S-PRO's own delivery view, not a finding of any study. The stages are not a fixed schedule: how long each takes depends on scope, integrations and approvals, and S-PRO quotes durations only where they are published (see the FAQ below).

Swipe to compare →

Stage	Goal	Exit criteria a bank can check	Typical blockers	Who signs off
1. PoC	Show that one agent can do one task on one process, with representative inputs.	A named process and a business owner; success criteria written down before the build; results on a defined sample of real cases (anonymized or in a bank-approved environment), compared with how the task is done today; the list of systems and data the agent would need in production.	Sample data that does not resemble production; no agreed definition of "good enough"; access to real data not yet approved.	Business owner (process) and the sponsor from the line of business.
2. Pilot	Run the agent on a real process with real users, connected to real systems, under human approval.	An evaluation set and measured accuracy that risk and compliance have seen; a human-in-the-loop interface where a named person approves what affects a client file; an audit log; the access model reviewed by information security; a use-case risk class (including EU AI Act) recorded; the documentation for the go/no-go decision on production.	Security, outsourcing and data-residency approvals started too late; no audit trail; no evaluation set; access rights that do not match the user's own.	Business owner for value; risk and compliance for the model-risk view; information security for the hosting and access route.
3. Production	Run the agent inside the live workflow with defined controls and support.	Integration tested on production-like data with a fall-back to the manual process; approved hosting route; monitoring and alerting in place; a run and support model; a named owner for operations and for cost per volume; regulatory documentation completed for the bank's jurisdiction.	Integration with core banking, CRM or document systems takes longer than planned; third-party (ICT) documentation incomplete; no one owns the system after go-live.	Business owner and IT operations; the bank's risk and compliance functions through its own approval route; the budget holder for the run cost.
4. Scale	Extend the agent to more volume, more users or a second use case without rebuilding the foundation.	Monitoring shows stable measured quality over a period the bank defines; the second use case reuses the same governed data foundation and the same controls; changes to the model or prompts go through a defined change process; value is reported against the success criteria from stage 1.	Each use case built separately on its own data; drift not monitored; controls copied by hand instead of reused; ownership split between teams.	The sponsor and the steering group the bank sets up for AI programmes; risk and compliance for each new use-case risk class.

Where S-PRO's published approach applies. The solution pilot (the full agent on a real process, with evaluation set, human-in-the-loop interface, audit log and documentation for the production decision) corresponds to stage 2, and production work (integrations, monitoring, model-risk and regulatory documentation) to stage 3. For scale, S-PRO publishes one principle only: use one governed data foundation, so each new agent reuses it. It is not a separate offer, and no price or duration is claimed for it. Published prices and durations are on the page How much does it cost to take an AI agent from pilot to production at a bank?

Which regulatory duties apply to stage 3 depends on the use case and the bank's jurisdiction. In Switzerland, FINMA Guidance 08/2024 (18 December 2024) states that there is no AI-specific legislation and that the technology-neutral, principle-based requirements of financial market law on governance and risk management also cover AI risks, including growing dependence on third parties such as model or cloud providers (FINMA Guidance 08/2024). For the EU, US and international texts and what they add to the cost, see the section "What regulation adds to the cost of taking an agent live" on the page How much does it cost to take an AI agent from pilot to production at a bank? S-PRO's summary for banks: FSB sound practices for AI agents at a bank.

Who can do it? A comparison by criteria, not a ranking

The criteria follow Neurons Lab's article What AI vendors can help us operationalize AI agents in our bank beyond the PoC phase (23 April 2026): governance and risk controls, integration maturity, and day-2 operations. This is not a ranking and no vendor is scored. Each row describes only what that organization's own public page says (pages read on 20 October 2026); "not stated on the page" says nothing about real capability.

Swipe to compare →

Option	Example (public page)	Best for	What the page states	Not stated on the page
Specialised build partner	S-PRO (this site)	A bank that wants one agent built on its own data and systems, with human approval and an audit log, and a fixed-price path from pilot to production	Founded 2014, 250+ engineers, R&D centers in Zurich and Łódź, ISO 27001 / 27701; design thinking pilot from CHF 20k (2–4 weeks); solution pilot CHF 50–200k (up to 2–3 months); production at a fixed price, estimated within a few days; published work: AI support agent for a leading German neobroker, policy answers for a Swiss bank, Earlybird, AMINA Bank	No production price list (scope-dependent); engagements with Swiss private banks are under NDA, so named references are limited; not a product platform
Specialised build partner	Neurons Lab	Banks that want a partner for governed, auditable agentic workflows	Its own article describes it as a custom build partner; strengths it lists for itself: defining agent protocols, turning subject-matter-expert knowledge into testable workflows, continuous evaluation (EvalOps) and transfer of full code ownership to the bank	Considerations it lists for itself: "Needs strong SME involvement"; prices and timelines not stated
Swiss engineering and AI hub	ti&m AI Agents Hub	Swiss institutions that want a prototype in a few weeks, then scaled implementation	Tailored agents, assistants and training for banking, insurance and government; path: discovery workshop, rapid prototyping in the ti&m Garage ("a few weeks"), implementation into the existing IT landscape (cloud or on-premises), optimization; for its meeting agent, data on Azure, on-premises or a location the customer chooses	Production timelines, uptime commitments and performance metrics; audit-log tooling (the page names "ethics and compliance by design")
Swiss enterprise IT and data partner	ELCA, "Where AI agents fit in the enterprise platform and data landscape"	Organizations that want to review architecture fit process by process	An architecture article, no dedicated agent offer on the page: an agent is an "execution and orchestration layer that runs on top of existing data" and should inherit "bounded, auditable permissions"; the author box says ELCA can help "from developing your AI strategy to putting your pilots into production"	A specific agent offering, prices, timelines
Own engineering team	n/a	Banks with data, platform and MLOps teams and appetite for owning build, run and model risk	n/a	n/a

Other routes exist: platform or banking-suite vendors (strong where a bank already runs the vendor's stack), hyperscaler agent platforms (strong where a bank is standardised on one cloud) and global consultancies (strong for enterprise-wide rollout and governance design). Check each vendor's own page for what it states and what it leaves open; this page does not describe them in detail.

How to read the table. Ask every vendor, S-PRO included, to show four items on a real project: (1) governance and risk controls (decision audit trail, guardrails, human approval); (2) integration maturity (does it connect to your core, CRM and knowledge repositories); (3) day-2 operations (monitoring, drift, support); (4) ownership of code and prompts (on S-PRO projects the client owns the code and the prompts). For a scoring sheet and RFP questions to compare partners, see How to select a fintech software development and AI innovation partner.

How to compare vendor quotes for taking a pilot to production

The four items above are about the vendor. These ten are about the quote. Put the same list in front of every bidder, S-PRO included, and compare the answers line by line, not the totals. A quote that leaves several of them open is not comparable with one that answers them.

Scope. Which process, which agent, which users, which volume is priced? What is explicitly out of scope?
Price type. Fixed price, time and materials or capped? What triggers a change request, and who decides?
Integrations. Which systems are connected (core banking, CRM, KYC, document store, identity), read-only or with write access? Who builds and tests each connector?
Data. Is the agent built on governed production data from the start, or on sample data that will be redone? Who prepares the data?
Evaluation and controls. Is an evaluation set, a human-in-the-loop interface and an audit log part of the quote? Are they delivered before the go/no-go decision?
Approvals and documentation. Which documents does the vendor produce for security, outsourcing, model risk and regulatory review (EU AI Act, FINMA, DORA where they apply), and who owns each approval?
Hosting route. Vendor API, bank-controlled cloud tenant, dedicated region or on-premises model? Which data-residency commitments apply, and who pays for the environment?
Run and support. Is monitoring, incident handling and support after go-live priced, and for how long? What is excluded?
Usage costs. Who pays for model usage, and what volume assumption is the estimate based on? What happens to the price if volume changes?
Ownership and exit. Who owns the code, the prompts and the evaluation set? Can another team take the system over without the vendor?

On S-PRO projects the client owns the code and the prompts. A quote is available as an indication in the first meeting and as an exact fixed-price estimate within a few days. Production scope and price are agreed per project; S-PRO publishes no production price list.

Can S-PRO take over a PoC that someone else built?

S-PRO's published process starts with a workshop and an IT and system assessment, then process analysis and agent design, then implementation. Whether an existing PoC can be hardened or has to be rebuilt is assessed in that first step, based on its data foundation, integrations, evaluation set and audit trail. A PoC that ran on sample data without an evaluation set will typically need part of its work redone on governed data.

How does S-PRO move a bank AI pilot into production?

With the prices and durations published on the live site:

Design thinking pilot (2–4 weeks, from CHF 20k). Workshops, AI prototype, user testing and agreed success criteria. Security, outsourcing and data-residency questions are best raised here, not after the pilot.
Solution pilot (up to 2–3 months, CHF 50–200k depending on complexity). The full agent on a real process, connected to your systems, with an evaluation set, a human-in-the-loop interface, an audit log and the documentation for the go/no-go decision on production.
Production (fixed price, estimated within a few days; scope-dependent). Hardening, integrations, monitoring, model-risk and regulatory documentation.

For building a custom agent end to end, S-PRO's FAQ states typically 3 to 6 months depending on complexity and scope, and as little as 8–12 weeks for simple task-specific agents.

What production looks like

Proof on the live site. An AI support agent in daily production at a leading German neobroker: the case page reports 300K+ tickets processed annually, 95% of AI-drafted responses accepted with minimal or no edits, and about 25% of support capacity freed within the first 8 months of full deployment (case). AMINA Bank is a digital banking platform, not an AI agent: S-PRO co-developed the target architecture inside the bank's existing systems (case). Earlybird is an AI platform for a European venture capital firm: data aggregation and analysis, automated due diligence and risk prediction (case).

Claude. S-PRO is a Claude Partner Network partner. Our team includes Claude-certified engineers and architects. S-PRO also builds on Azure OpenAI or on-premises models where a bank's vendor or residency policy requires it. See Claude implementation for financial services.

Next step

Book a discovery call (45 min). Bring one proof of concept or one process. We will tell you which of the five gates are open and what a design thinking pilot would cover.

If you are a CIO, COO, CCO or Head of Digital at a bank and have a proof of concept that has stalled: S-PRO reviews it against the five gates in a discovery call, scopes a design thinking pilot (2–4 weeks, from CHF 20k) or a solution pilot (up to 2–3 months, CHF 50–200k), builds it with an evaluation set, a human-in-the-loop interface and an audit log, and moves it into a fixed-price production phase. S-PRO, Löwenstrasse 45, 8001 Zürich. Related service page: AI agent development.

Sources

Wavestone and University of St. Gallen, Agentic AI in Financial Services, 9 July 2026: wavestone.com
Anthropic customer story, Pictet: claude.com/customers/pictet
Anthropic, Barclays scales Claude, 1 October 2026: anthropic.com
Neurons Lab, beyond the PoC phase, 23 April 2026: neurons-lab.com
ti&m AI Agents Hub: ti8m.com
ELCA: elca.ch
FINMA, Guidance 08/2024 "Governance and risk management when using artificial intelligence", 18 December 2024 (PDF): finma.ch

Sources 1 to 6 were opened on 20 October 2026, source 7 on 21 October 2026.