Short answer. Most banking AI use cases reach production in 3 to 9 months. A low-risk internal assistant can ship in about 90 days, while a high-risk system such as credit scoring runs 6 to 12 months because it needs a conformity assessment, bias testing, and human oversight before deployment. The clock is set by data readiness and governance sign-off, not by model training. In 2025 Ableneo shipped 34 production AI projects, and 4 of 5 reached production.
The honest answer to “how long” is a range, and the range is driven by risk class and data condition, not by the algorithm. Three bands cover most banking work.
Model training is rarely the bottleneck. A useful model can often be built in days. The months are spent connecting it to a 20-year-old core, cleaning and governing the data that feeds it, and getting risk, compliance, and audit to sign the release. Ableneo runs 94% of its projects on large language models, and the delivery pattern is consistent: the model is the fast part, the surrounding system is the work.
Concrete outcomes follow the same logic. Ableneo has cut invoice processing time by 37% and reduced quality-control errors by 22% in production settings. Those numbers landed because the work fixed data flows and integration, then measured against a baseline, rather than stopping at a demo.
Two variables move a use case between the bands. The first is data condition: a use case sitting on clean, well-defined data can jump a band, while one that depends on reconstructing history from three source systems drops a band. The second is decision weight: the moment a model influences a customer outcome, a regulatory clock starts and the timeline extends. Naming both variables at the start turns “how long” from a guess into a plan a board can hold you to.
In banking and insurance the timeline is a governance question before it is an engineering one. Under the EU AI Act, credit scoring and creditworthiness assessment are classed as high-risk under Annex III, point 5(b). A high-risk system cannot go live until it has a risk management system, data governance documentation, bias testing, logged human oversight, and a completed conformity assessment under Article 43. The high-risk obligations apply from 2 August 2026, so the paperwork is now part of the critical path, not an afterthought.
DORA adds a parallel track. Any AI running in production is ICT that must be inventoried, tested for operational resilience, and covered by third-party risk controls. A bank that treats these obligations as a final gate discovers them late and slips by months. A bank that builds the evidence while it builds the model keeps the timeline intact. The obligation set is fixed, so the only variable a bank controls is when it starts producing the evidence.
The ECB has been direct that the challenge is governance rather than the technology itself. For a bank, that reframes the timeline: the fastest route to production is the one where risk and compliance are in the room from the first sprint, signing off on the control design as it is built. The slowest route is the one where the model is finished, then handed to a governance queue that has never seen it. The same model, the same data, can differ by three or four months on that single choice.
Most banking AI use cases reach production in 3 to 9 months; high-risk decisioning systems run 6 to 12.
Four things, in order of impact. Data readiness is first: fields are missing, definitions differ across systems, and history is incomplete. The OECD reports that data governance and limited visibility into models are among the leading constraints on AI adoption in finance, and underestimated data work is the most common reason a plan slips. Second is legacy integration: a real-time model on a batch core needs a clean data pipeline before it can be trusted. Third is governance sign-off across risk, compliance, and audit. Fourth is change management, since a model nobody uses has zero return regardless of how fast it shipped.
It adds a fixed block of evidence work that must finish before launch. For a credit scoring model, the provider must produce Annex IV technical documentation, run bias and quality testing on the training data, design and log human oversight, and complete a conformity assessment before the system is placed on the market. Most Annex III use cases, including credit scoring, allow an internal self-assessment rather than a third-party audit, which shortens the path. The practical effect is a 2 to 4 month governance layer on top of build and integration. Banks that assign this work on day one absorb it in parallel. Banks that discover it at the end restart, and a 6-month plan becomes a 10-month one.
Scope narrow and ship one decision, not a platform. A single use case with a clear owner and a measurable baseline reaches production far faster than a program that tries to transform a department at once. Run modernization in parallel: build the use case and clean the data foundation at the same time rather than waiting for a full core replacement. Reuse governance artifacts, because the risk template and monitoring setup from the first project cut weeks off the second. This compounding is visible in Ableneo’s portfolio, where the average client runs 1.7 projects, and the second lands faster than the first on the same governed foundation.
They confuse a working demo with a production system. A demo proves the model can answer. Production proves the system is reliable, observable, integrated, and governed under load. Only about 13.5% of EU enterprises used AI in 2024, and the gap between experimenting and operating in production is where most timelines break. The fix is to define “done” as the audit-ready, monitored, live system from the first planning session, then measure progress against that definition rather than against demo milestones.
A credible timeline comes from a track record, not a brochure. Ableneo shipped 34 production AI projects in 2025 across four countries, with roughly 4 of 5 reaching production, and works inside the same regulated, legacy-core environments the question is really about, including banks and insurers in Central Europe. That production rate is the difference between a plan and an estimate. For the related decision of moving a project that has already stalled, see our guide on how to get a stalled AI pilot into production.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.