Short answer
Banks and insurers can train engineers on coding agents if the workshop is designed as a controlled activity: documented tool assessment, defined data classes, change management that treats agent output like any other code, and evidence for auditors. The training itself becomes part of the control environment rather than an exception to it.
On this page
The regulated starting point
Engineering teams at banks, insurers and payment firms work inside a dense control environment. The EU Digital Operational Resilience Act (DORA) sets requirements for ICT risk management, incident reporting, resilience testing and the management of ICT third-party risk. National supervisors add their own expectations, and many institutions still structure their internal IT controls along the lines of supervisory catalogues such as Germany's BAIT and VAIT. On top of that sit internal policies on outsourcing, change management, access and data classification.
None of this forbids coding agents. It does mean that an agent is, from a control perspective, an ICT service from a third party that touches source code and possibly data, and that its output enters production through the change process. A workshop that treats this as a footnote will not survive the first conversation with the second line of defence.
At neooutput we therefore design regulated workshops so that they produce both engineering skill and the artefacts risk and compliance functions need.
Map the workshop to your control framework
Before any hands-on session, we run a short alignment with the CISO office, operational risk and, where relevant, the outsourcing or third-party risk function. The goal is a one-page mapping like this:
| Control area | Question for coding agents | Typical workshop answer |
|---|---|---|
| ICT third-party risk | Is the tool provider assessed and registered? | Pilot under existing assessment or a time-boxed exception |
| Data classification | Which data classes may reach the model? | Source code of defined repositories, no customer data |
| Access management | Who may use the tool and with which rights? | Named pilot users, SSO, no shared accounts |
| Change management | How does agent output reach production? | Same pipeline, same four-eyes review, labelled PRs |
| Logging and evidence | What can we show an auditor? | Tool usage logs, PR history, review records |
| Secure development | How are secrets and injection risks handled? | Permission modes, secret scanning, sandboxed runs |
This mapping is often the first document of its kind in the institution. It turns an abstract debate into specific decisions.
Keep personal and confidential data out of prompts
The most common concern is customer data leaking into prompts: test fixtures copied from production, log excerpts with account numbers, or a support ticket pasted to explain a bug. The workshop addresses this directly:
- Synthetic test data generation as a standard exercise
- Rules for logs and stack traces before they are shared with an agent
- Repository hygiene: no real data in fixtures, secrets in vaults, not in files
Where text that may contain personal data has to pass through an AI tool, a local anonymization layer helps. Textwash Pro anonymizes text with local AI, fully offline or as a privacy layer in front of ChatGPT and Claude, so incident descriptions or support texts can be used without exposing personal data.
Workshop design
Day one: controlled hands-on practice
The morning focuses on skills on a pre-approved repository: orientation in the code, context files, small bug fixes with tests. Participants learn to keep agent changes small and traceable, because traceability is what auditors will look for.
The afternoon is about quality: test generation for legacy modules, characterization tests before refactoring and documentation that explains business rules. In financial software, many business rules are only encoded in code, so having an agent extract and document them is valuable and low risk.
Day two: change management and evidence
We practise the full path from task to merged change under the institution's real process:
- Ticket with acceptance criteria and risk classification
- Agent-assisted implementation on a branch
- Pull request labelled as agent-assisted, with a generated risk summary
- Four-eyes review by a human who did not drive the agent
- Pipeline with tests, static analysis, dependency and secret scanning
- Merge with the usual approvals
We then look at what evidence this produces and whether it satisfies internal audit expectations. Gaps become action items, not blockers.
Optional module: resilience and incident scenarios
What happens if the agent provider is unavailable? Teams should be able to work without it. What if an agent introduces a vulnerable dependency? We walk through the detection and response path. These scenarios connect directly to DORA's resilience expectations.
Common objections and how we address them
"The model provider could train on our code." Enterprise terms of major providers address this contractually. The workshop includes a review of the actual terms the institution would sign, not assumptions.
"Agents write code nobody understands." That risk exists with any author. The answer is smaller changes, mandatory tests and reviewers who did not drive the agent.
"We cannot prove what the agent did." Pull request history, review records and tool logs provide a trail. We define which of these are retained and for how long.
From workshop to governed pilot
The workshop should end with a pilot plan: participating teams, repositories, data classes, duration, metrics and an exit review with the second line. Our guide to writing an AI coding tool policy and the deep-dive on security and data privacy of AI coding tools are useful companions for the people writing that plan. Teams with a strong security focus may also want the secure agentic coding workshop as a follow-up.
neooutput delivers this format in English or German, on site or remote, and adapts it to banking, insurance or payment contexts. We do not replace your legal or compliance assessment; we make sure the engineering practice gives that assessment something concrete to evaluate.
FAQ
Does DORA prohibit coding agents?
No. DORA sets requirements for ICT risk management and third-party risk, among other things. Coding agents have to be handled within that framework like other ICT services, which is why a documented assessment matters.
Can we use real production code in the workshop?
Often yes, if the repository contains no customer data or secrets and the tool has been approved for a pilot. Otherwise, start with a sanitized repository and move to real code once approval is in place.
Who should attend besides engineers?
Invite someone from information security and operational risk to the first module. Their presence speeds up later approvals considerably.
How do we label agent-assisted changes?
A simple pull request label or template field is enough. What matters is consistency, so that reviews and audits can filter agent-assisted changes when needed.