Using AI in an interim assignment while protecting data

For each task, I therefore define the required outcome and which data may be used. An agent combines an AI model with a work instruction and approved tools. It can plan and carry out steps. Professional judgement and approvals remain with the people responsible.

Agents for everyday work

To analyse large datasets from the ERP system, an agent uses approved queries and tested programs. The model explains selected results; complete accounting, sales or inventory datasets are not transmitted without prior review. The finance team checks completeness and reconciles totals with the source data. The period, quantities and currencies must be clear. Anomalies initially flag matters for review. The agent must show which data and calculations support its conclusions before action is taken.

For market analysis, a research agent can gather public information about competitors, demand or prices. Key statements must be accompanied by verifiable sources and the date to which the information relates. Differences in periods and definitions must remain visible. I have the information checked against the original sources and any contradictions highlighted. Confidential company information does not belong in public search queries. Internal strategic considerations remain separate from the research; only afterwards does the finance team combine them with the verified findings and assess their significance for the business.

When reviewing contracts, an agent can identify payment terms, deadlines, price changes and liability provisions. Every finding must be supported by a passage in the original contract. Ambiguous wording and missing information are flagged. I assess the financial implications, while legal questions go to the appropriate specialists. Amendments and the conclusion of contracts remain the responsibility of the designated people. Before processing begins, it must be clear which AI model and agent may receive the contract, as individual clauses can contain personal data or confidential commercial terms.

Choosing the right AI model

Different AI models may be suitable for these tasks. Claude Code is a tool that can send requests to an external model. The Claude models themselves cannot be installed on a company’s own computers. Self-hosting requires a different model that is available for this purpose, along with compatible software. None of the following options complies with data protection requirements simply because of its name. When selecting an option, I assess the protection the data requires and whether the company can provide ongoing support.

Company-owned infrastructure. A suitable model runs on the company’s computers or servers. The licence and hardware must suit the task. The company manages updates, user permissions and backups. Local does not automatically mean offline: extensions or file transfers can establish external connections. Before investing, I test the quality of the results on typical tasks.

Rented GPU computing power. The company runs a suitable model on rented servers equipped with graphics processors for model computations. Alongside licensing and operating costs, the infrastructure provider’s access rights must be examined. Security and support require clear responsibilities. The economic assessment must consider actual utilisation together with the effort needed for maintenance and oversight.

European model services. A provider operates the model. The developer and operator may be different companies, so it must be clear who processes the data. A European registered office does not guarantee that processing takes place exclusively in Europe. Contractual terms, other service providers and storage must be checked for the specific offering. The tool used must support the model.

Direct access to US providers. Claude can be accessed directly through Anthropic, and other models through their respective providers. The company does not have to operate the model itself. The terms of the selected business account and the lawfulness of the data transfer are decisive. A US provider is not ruled out solely because of its location.

Cloud service with an EU region. Claude Code supports services such as Amazon Bedrock and Microsoft Foundry. This allows existing cloud agreements and user permissions to be used. It remains necessary to check whether the model is available in the intended region and whether requests are routed elsewhere. Logs and support access form part of the review; the chosen location does not automatically cover every data flow.

Sovereign cloud services. These services promise greater control over operations and access. It must be possible to verify who operates the service and who can read or decrypt data. Independent assurance must cover the specific offering used. It does not automatically establish the security of the company’s own application. The required models and tools must actually be available.

Data protection and operational responsibility

Agents initially receive read-only access to the source data they need. Additional connections are checked before use. IT enforces restrictions through technical controls; an instruction to the model alone is not enough. External documents and websites must not trigger additional actions or data disclosures. Even program output and error messages can contain confidential information.

Replacing names with identifiers amounts only to pseudonymisation if additional information can still link the data to individuals. Data protection rules continue to apply. For data to be anonymous, identification must no longer be possible by means reasonably likely to be used. Trade secrets must be protected regardless of whether they relate to individuals.

A provider may store inputs even if it does not use them to train the model. Retention, access and deletion must therefore be clarified separately. International transfers require their own assessment. Under the CLOUD Act, covered providers can be required, subject to the statutory conditions, to disclose data under their control, even when it is stored outside the US.

The Swiss Federal Act on Data Protection (FADP) and, where applicable, the European General Data Protection Regulation (GDPR) apply. The purpose of processing and information obligations must be clarified. Where a service provider processes personal data on the company’s instructions, appropriate contractual arrangements for processing on its behalf are required. A data protection impact assessment is required where processing is likely to pose a high risk to the individuals concerned.

As CFO, I define the business requirements. IT and the responsible data protection function or legal adviser review their respective areas before management approves the use. I assess the benefits including the effort required for review and corrections. At handover, work instructions, responsibilities and approvals must be documented. At the end of the assignment, access is revoked and working copies are deleted or handed over in accordance with the agreed rules.