Local AI or EU cloud? Data protection, quality and cost

Start with the task and its information boundary. Describe the source material, allowed recipients, expected output and what an employee must check. Then identify who maintains the application and restores it after a failure. These requirements should stay constant while you compare deployment options.

Use the table to compare the options against your requirements.

Separate mandatory conditions from preferences. If a destination is not permitted, a lower price cannot resolve that condition. If both candidates pass the required controls, you can then compare operating effort and the time employees spend checking or correcting answers.

In typical EU projects, we use models through AWS Bedrock and Azure OpenAI in EU regions for client data. Check the resource location and actual processing location separately for the selected service configuration.

AWS documents that a geographic Bedrock inference profile stays within its geography but may route beyond the source region. Record the selected profile and its destination list rather than relying on the region used to submit a request.

A documented example is the Claude Haiku 4.5 EU profile, eu.anthropic.claude-haiku-4-5-20251001-v1:0. AWS distinguishes on-demand access through geo/global profiles on the runtime endpoint from regional access documented for the mantle endpoint. Check the endpoint as well as the model identifier.

Azure likewise separates the resource from inference scope. Global deployments may process worldwide; geography-based and Data Zone deployments have different boundaries. The deployment guide includes EFTA locations in the EU zone. Record the deployment type, not just the resource region.

Microsoft's EU Data Boundary includes the EU and EFTA and has documented exceptions. It is not an EU-member-only promise.

There is also an unresolved documentation conflict: the Azure privacy page uses EU-member wording for Data Zone, while the deployment guide includes EFTA. Confirm the contractual scope and destination list before promising processing only in EU member states. Record that confirmation in the project evidence.

AWS says Bedrock inputs and outputs are not used to train AWS or third-party models. That addresses training use, not every form of storage or access.

Its abuse-detection documentation describes model-dependent retention, including up to 30 days for certain named models. Retained cross-region traffic is stored in destination regions. Do not apply that period to every model or infer the complete Haiku retention arrangement from it.

Microsoft states that inputs and outputs are not used to train foundation models without customer permission or instruction. Stateful features and abuse review can nevertheless involve storage. Check the features actually enabled and any applicable review arrangements.

Draw the full data path: source, prompt, retrieved passages, answer, trace and backup. Record retention and access for each copy, including those created by your application. The document-processing privacy guide helps organise this inventory. Leave unknown retention rules open until the selected configuration has been checked.

Local hosting does not remove requirements concerning purpose, lawful processing, minimisation and retention. Use the EDPB's data-protection basics and lawful-processing guidance to frame the review. Use the following checklist to collect the evidence for your review.

AWS also assigns security configuration and management tasks to the customer. A managed model service does not remove the need to secure the surrounding application. Assign each open checklist item to someone who can produce evidence before approval.

Professional secrecy is a separate review. German tax-adviser outsourcing has contractual requirements under section 62a StBerG; Austrian professional secrecy is addressed in section 80 WTBG. EU hosting alone does not settle those obligations. The tax-firm confidentiality article and law-firm service page provide the relevant project context.

Do not assume a local-model quality deficit. Prepare representative tasks and expected outcomes, then run the local and cloud candidates with the same permitted information. Fix versions and settings and retain the results. Have subject reviewers assess correctness, evidence and the work needed to make an answer usable.

Include difficult cases, missing information and permission boundaries. An answer that appears helpful but uses a forbidden source should fail the comparison. Measure successful completed tasks alongside latency, rather than treating generated text as completed work. The knowledge-assistant acceptance guide provides an existing basis for that exercise.

Keep quality and capacity conclusions separate. A correct response under isolated testing does not establish acceptable behaviour during the expected burst. Likewise, a quick answer does not prove that employees can rely on its content. Agree the acceptance conditions before reviewing the candidate results.

As of October 2026: the following assumptions were chosen for this planning example on 4 October 2026. They are not measurements, staffing recommendations or market tariffs. Use your own observed values before approving expenditure.

The assumed workload gives 20 million input tokens and 5 million output tokens monthly. Anthropic's public list (page 6) effective 27 May 2026 gives Claude Haiku 4.5 on Bedrock Standard, In-Region & Geo Cross Region, at USD 1.10 input and USD 5.50 output per million tokens, in the context band up to 200K. This is a dated price input, not a confirmed October tariff.

Under those assumptions, inference is 20 × USD 1.10 + 5 × USD 5.50 = USD 49.50 per month. Add the assumed cloud labour and provision: USD 49.50 + EUR 70 per month. Keep the currencies separate until a dated exchange rate and tax basis are available.

The local purchase input is the Mac Studio M5 Max starting price of EUR 2,999 including German VAT, from Apple's announcement of 25 August 2026. It is not a current checkout quote.

Using the assumptions above, equipment allocation is EUR 2,999 / 36 = EUR 83.31 monthly, rounded. Energy is 0.100 kW × 160 hours = 16 kWh, or EUR 4.80 at the assumed tariff. Adding EUR 120 labour and EUR 10 provision produces an incomplete local subtotal of EUR 218.11 monthly. This allocation period is not a statutory tax depreciation claim.

Neither subtotal includes initial integration, security setup, retrieval storage and embeddings, redundancy, downtime or application licences. The local machine has not been demonstrated to complete the assumed workload within acceptable response times. The purchase input includes German VAT, while cloud tax treatment remains unresolved. These are not comparable final quotations and establish no price winner.

Check the current rates on the Azure price page for the selected model, deployment, region, currency and tax basis before adding an Azure calculation. Do not treat a placeholder as a zero price.

The RAG cost guide covers retrieval and knowledge-maintenance budgeting. The self-hosting comparison helps assign operating and recovery work. Include these responsibilities even when the model's own usage charge appears small.

Record the approved task, data flow, selected model and deployment, confirmed processing scope, retention rules, contracts, test results and comparable cost assumptions. List unresolved points with an owner. Avoid a break-even claim until tax, exchange rate, successful throughput and operating effort are on a common basis.

Approve a bounded use case, then revisit it when sources, permissions, models or service settings change. Training is part of the project: employees need to understand permitted inputs, answer checking and the fallback when the service fails. Capture those rules in the company AI usage policy. See our guide to local AI for further planning.