Short answer. Staff it small and sequenced. A bank ships its first production AI use case with a core squad of 5 to 8 people: one AI product owner, two engineers, one data engineer, one MLOps or platform engineer, and a part-time model risk lead. You do not need a 50-person AI department to start. Across 34 production AI projects Ableneo shipped in 2025, roughly 4 of 5 reached production with lean cross-functional squads, not large central teams.
The staffing question is really a sequencing question. Banks that try to hire a full AI department before shipping anything spend 12 months recruiting and still have nothing in production. The teams that ship start with one use case and one small squad, then add people as the workload proves itself.
A working first squad in a CEE bank has five seats. Each seat owns a decision, not a job title:
The skill mix has shifted. In 2025, 94% of Ableneo projects used large language models, which moves the demand toward retrieval engineering, prompt design, and evaluation, not training models from scratch. A bank that staffs only classical data scientists ends up short of the people who actually get an LLM system into production. The people who ship pair banking domain knowledge with hands-on delivery, and that pairing is rarer than either skill alone.
In a bank, staffing an AI team is a compliance obligation, not only a hiring plan. Article 4 of the EU AI Act has applied since 2 February 2025 and requires both providers and deployers of AI systems to ensure the people who operate them have a sufficient level of AI literacy. A bank cannot put an unqualified team on a credit-scoring or fraud model and treat skills as a nice-to-have. The team composition is part of the audit trail.
The pressure is real. The OECD reports that around 40% of finance and manufacturing employers who have not adopted AI name a shortage of skills as the main reason. DORA adds a second requirement: named human accountability for every ICT and AI system in production. That means a bank’s AI staffing plan has to answer a governance question, who signs off and who monitors, before it answers a technology question. Staffing decisions made only on cost or speed fail the audit later.
Ship the first AI use case with a squad of 5 to 8, not a department of 50.
Five role families cover a production AI capability in a bank. Product owners translate a business problem into a scoped use case and defend the metric. Engineers build the integration and the application layer. Data engineers own the pipelines, quality, and lineage the model depends on. MLOps or platform engineers own deployment, monitoring, and the rollback path. Model risk and compliance specialists validate the system and hold the sign-off.
The mistake is to over-index on data scientists and under-index on the platform and risk seats. A model that no one can deploy, monitor, or defend in an audit never leaves the pilot. The engineering and platform roles are where most of the production work sits, and the risk seat is what turns a working demo into a system the bank is allowed to run.
For most banks the answer is sequenced, not either-or. Direct hiring for scarce AI roles in CEE runs 6 to 9 months per senior seat, and the first hire cannot ship a use case alone. A partner installs the operating pattern and delivers the first two or three production use cases in 60 to 120 days, while the bank hires a lean internal owner in parallel. As the internal team stabilizes, the partner steps back to specialist support for spikes and hard problems.
Use three tests to decide each seat. Build in-house when the capability is core, permanent, and you can attract the talent. Hire when the role is full-time, stable, and the market has supply. Partner when you need production speed now, when the skill is scarce, or when you want to learn the operating pattern before you commit headcount. Ableneo’s own clients run 1.7 projects each on average, which is the shape of a partnership that transfers capability rather than one that creates dependency.
The first squad proves one use case. The second and third squads prove the pattern repeats. Only then does a bank centralize the shared parts into a Center of Excellence: the MLOps platform, the model risk framework, the reusable data pipelines, and the AI literacy program. Centralizing too early creates a bottleneck team that reviews everything and ships nothing. Centralizing after three shipped use cases means the CoE codifies patterns that already work.
A useful ratio: keep delivery squads close to the business units that own the use cases, and keep governance, platform, and standards central. That split gives you speed at the edge and consistency at the core, which is what a regulator wants to see and what a stalled program usually lacks.
Bake AI literacy into the staffing plan from day one. Under Article 4, every person who operates or oversees an AI system needs documented, role-appropriate training. That is not a single company-wide slide deck. A model risk lead, a customer-facing operator, and a platform engineer need different depth. Record who was trained, on what, and when, and keep it current as the system changes.
Pair literacy with named accountability. Every production AI system needs one accountable owner and one monitoring owner, documented, so a DORA audit finds a person, not a team inbox. Banks that staff the risk and monitoring seats from the first squad pass this test without a scramble. Banks that add them after go-live rebuild the governance under audit pressure.
Hiring for scale before proving value. A bank stands up a 30-person AI unit, spends a year on recruitment and reorganization, and still has zero use cases in production because the team has no shipped problem to rally around. The teams that reach production do the reverse: one squad, one use case, one metric, then grow from proof. Roughly 4 of 5 Ableneo projects reach production because the staffing follows the work, not the other way around.
Ableneo staffs and ships AI in regulated banks across Slovakia, Czechia, and Austria with 80+ specialists, and about half of the work is embedding those specialists directly into client teams rather than delivering from the outside. That model is why 4 of 5 projects reach production and why clients come back for 1.7 projects on average: the operating pattern stays in the bank after the engagement ends. See how this works in practice on the Ableneo AI transformation page.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.