Contents
Agentic coding is possible in regulated industries – banking, insurance and healthcare – if three things are ensured: control over which data and code leave the organisation, complete traceability of every change, and clear human accountability with the four-eyes principle. Coding agents do not change your regulatory obligations; they change how you have to meet them. This article explains what matters.
Why regulated industries need to be especially careful
In financial and healthcare organisations, software development is subject to additional requirements: supervisors expect orderly change management, segregation of duties, risk assessments for IT service providers and documented information security. In the financial sector, the EU Digital Operational Resilience Act (DORA) consolidates many of these requirements and has applied since early 2025. In healthcare, the special protection of health data under Art. 9 GDPR, professional secrecy obligations where applicable and – for software with a medical purpose – medical device regulation come on top.
In this context a coding agent is two things: a tool that changes code and therefore falls under change management, and often an external service to which code and possibly data are transferred. You need to cover both perspectives. This article provides an overview and is not legal advice; the specific assessment belongs with your compliance, data protection and information security functions.
The key requirements at a glance
| Requirement | Meaning for agentic coding | Typical implementation |
|---|---|---|
| Data protection and confidentiality | No uncontrolled outflow of personal data or trade secrets | No real data in agent environments, data processing agreement, reviewed data flows |
| Third-party risk management | Assess AI vendors as IT service providers | Risk analysis, contract review, exit strategy |
| Change management | Every agent change must reach production in a controlled way | Pull requests, CI checks, documented approvals |
| Segregation of duties and four eyes | Agent and requester must not approve alone | Mandatory review by a second person, protected branches |
| Traceability | Who changed what, when, with which tool? | Tagging agent contributions, logs, ticket references |
| Information security | Agents must not open new attack paths | Sandboxing, least privilege, secret scanning, security tests |
| AI literacy | Staff must use AI tools competently | Training and policies (cf. Art. 4 EU AI Act) |
Data protection and data sovereignty: what may reach the agent
The first question is: what information does the agent see, and where is it sent? With cloud-based agents, source code – and everything that ends up in the context – leaves the organisation. Regulated organisations have found three principles useful:
- No production data in the context: agents work with synthetic or anonymised test data. Logs, database dumps or error messages containing customer data do not belong in agent sessions.
- Contractual safeguards: review terms on data use, retention and training, the data processing agreement, server locations and sub-processors.
- Tiered tool choice: for particularly sensitive code areas, models running in a controlled environment – in your own cloud region or locally – may be the better choice. Our article on local LLMs for software development covers the pros and cons.
More detail on sending source code to AI services is in AI coding tools and data protection.
Traceability and change management
Auditors care less about whether an agent was involved than about whether every change was controlled, justified and approved. The good news: a clean agentic coding workflow is often more traceable than manual development, because every step is documented through pull requests, pipeline runs and tickets. Make sure of the following:
- Ticket reference: every agent task starts from an approved ticket or change request.
- Tagging: pull requests state that and how an agent was involved.
- Human authorship: a named person is accountable for the content of each change.
- Independent review: a second person reviews and approves; agents have no merge rights on protected branches.
- Automated checks: tests, static analysis, dependency and licence scans run on every change.
- Retention: pull requests, review comments and pipeline results are archived according to your retention rules.
Security: agents as part of the attack surface
Coding agents run commands, install packages and read content from the repository, tickets or the web. That makes them a potential target for manipulated content, such as hidden instructions in dependencies or issue texts. In regulated environments, agents should therefore always run in isolated environments with restricted network access, never use production credentials and require human approval for risky actions. Secret scanning and security tests in the pipeline catch typical mistakes before they reach review.
Adoption in five phases
- Inventory: which systems, data and code areas exist, and how sensitive are they? This yields a classification of where agents may work first.
- Tool assessment: evaluate vendors on data protection, contracts, operating model and security features, aligned with compliance, data protection and information security.
- Write a policy: define approved tools, data classes, approval processes and responsibilities – as part of your AI governance.
- Pilot in low-risk areas: internal tools, test automation or documentation rather than the core banking system or patient records.
- Controlled expansion: evaluate experience and audit findings, tighten controls and extend step by step to further areas.
If you are planning new applications in these industries, you will find sector-specific requirements in developing a fintech app. For teams introducing agentic coding under regulatory constraints, we place AI developers from real-world practice as trainers – including in-house and focused on your policies; offers from 799 € net by e-mail.
Planning an app? Request a free fixed-price quote in 2 minutesConclusion
Regulation does not prohibit agentic coding; it demands discipline. If you control data flows, run agents in isolated environments with minimal rights, make every change traceable through tickets, pull requests and four-eyes review, and train your staff, you can capture the productivity gains in banking, insurance and healthcare too – and often fare better in audits than with manual processes.
Frequently Asked Questions
Are banks and insurers allowed to use coding agents at all?
In principle yes, provided supervisory requirements for third-party risk, information security and change management are met. The specific assessment should be made with compliance and information security; this is not legal advice.
May code dealing with health data be sent to a cloud agent?
Source code itself usually contains no health data, but test or log data may. Real patient data does not belong in agent sessions; for code and context, check contracts, server location and data use.
Is agent-generated code a problem in an audit?
Not if every change goes through a ticket, pull request, automated checks and documented four-eyes review. What matters is traceability, not the tool.
Are local models mandatory in regulated environments?
Not necessarily. They can make sense for particularly sensitive areas but come with their own operating effort; a tiered model combining cloud and controlled environments is often practical.