EU data residency for AI in finance: where models run, what gets logged and how long it stays
An EU hosting region is only the start. Follow one expense receipt through an AI bookkeeping workflow and see where the data rests, who can reach it and when it should be gone.
Manoj Tutika7 min read
Working on this yourself? Talk to Mark, our co-founder. We’re here to solve it with you.
On this page
EU data residency for AI in finance means that every copy of your financial data, including the prompts sent to a model, the model's output, the logs and the backups, is stored and processed inside the EU and cannot be reached from outside it. A hosting region on a sales page covers only one of those copies. The rest is where most of the risk sits.
This post is for CFOs, controllers and accounting firm partners who are about to let software read invoices, bank lines and payroll journals. We follow one ordinary document through an AI bookkeeping workflow, show where the data rests along the way, and end with the questions to put to any vendor, including us.
Is bookkeeping data really personal data?
More of it than most finance teams assume. The GDPR covers any information about an identifiable person, and a ledger is full of people.
- Purchase invoices from sole traders and freelancers carry a name, an address, a bank account and often a personal VAT number.
- Expense claims show where an employee was, what they ate and who they met.
- Payroll journals, even when booked as totals, sit next to the payslips they came from.
- Bank lines name the counterparty and carry an IBAN, and card statements list where your staff spent money.
So when an agent reads your documents, you are the controller of that data and the software provider is your processor. Article 28 of the GDPR requires a written agreement between the two, sets out what the processor may do, and says a processor may only bring in other processors (sub-processors, meaning the companies a vendor itself relies on, such as a cloud host or a model provider) with your prior authorisation.
Where does the data go when an agent books one receipt?
Take a single expense receipt. An employee forwards a €96.80 restaurant receipt (€88.81 plus €7.99 BTW at 9%) for a client lunch. In a typical AI bookkeeping flow, that one file creates at least six copies or derivatives.
- The original file, stored by the vendor so the entry can always point back to its evidence.
- The extracted text and fields, which the model produces and the system stores.
- The prompt and the model's response, which pass through a model provider's infrastructure and may be retained there for abuse monitoring.
- Application logs and traces, which developers use to debug and which often capture full inputs and outputs by default.
- Error reports sent to an error-tracking service when something fails.
- Backups of all of the above.
Data residency is a property of all six, not of the first. A vendor can keep its database in Frankfurt and still send prompts to a model endpoint outside the EU, ship traces to an observability tool hosted elsewhere, or give a support team in another country access to production. Each of those is a separate decision, and each should appear in the sub-processor list.
Where do the models actually run?
Large language models are mostly offered by a handful of providers, and many of them now sell EU-hosted endpoints. That is a real improvement, but it helps to separate three questions.
Where is the inference done? Inference means the model processing your prompt. If the endpoint runs in an EU region, the content is processed there. Ask whether the vendor pins that region explicitly, or whether requests can fail over to another region when capacity is short.
Who can reach the data? Chapter V of the GDPR (Articles 44 to 49) restricts transfers of personal data to countries outside the EU and EEA. The EDPB, the body of EU data protection authorities, has made clear in its guidelines on the interplay between Article 3 and Chapter V that making data available to someone in a third country, for example through remote access, can be a transfer even when the servers never move. A support engineer abroad who can open your documents is a transfer question.
Which law applies to the provider? A provider headquartered outside the EU may be subject to its home country's rules on government access, even for data held in an EU data centre. Adequacy decisions and standard contractual clauses exist to manage this, and the European Commission publishes the current list of adequacy decisions. Whether that residual risk is acceptable is a judgement for you and your data protection adviser, but you can only make it if the vendor tells you who is involved.
What gets logged, and why does it matter?
Logs are the copy nobody plans for. Modern AI tooling makes it easy to record every prompt and response for debugging and quality review, and that record is valuable. It is also a full copy of your financial documents, sitting in a system with different access rules and usually a different retention period from the ledger.
A sound design treats logs as a data store in their own right.
- Log references, not content, by default. A trace that says which document, which step and which model version is enough to debug most problems.
- Capture full content only for specific, time-limited investigations, and keep it in the same region and under the same access controls as the ledger.
- Redact IBANs, names and email addresses before anything leaves the core system, including error reports.
- Give logs their own short retention period, measured in weeks rather than years.
- Record who looked at customer data and why. Staff access should itself be logged and reviewed.
How long should financial data be kept?
Two clocks run at the same time, and they point in different directions.
The fiscal clock says keep it. In the Netherlands, the Belastingdienst requires you to keep your administration for 7 years, and data about real estate for 10 years. If you use the EU One Stop Shop (the Union or Import scheme), records for those supplies must be kept for 10 years. In the UK, a limited company must keep its records for 6 years from the end of the last financial year they relate to, and longer in some cases.
The GDPR clock says keep only what you need. Article 5 requires data to be kept no longer than necessary for its purpose (storage limitation) and limited to what is needed (data minimisation). Article 17 gives people a right to erasure, but that right does not apply where you must keep the data to meet a legal obligation, which covers your statutory bookkeeping records.
The way to reconcile the two is to separate the record from the working material.
- The ledger entry and its source document are the record. They stay for the statutory period and are then deleted.
- Prompts, raw model output and intermediate extractions are working material. Once the entry is booked and the evidence is attached, they serve no further purpose and can go within days or weeks.
- Logs and error reports follow the short clock too.
Retention periods are set by law and by your own situation, so check the exact periods with your accountant or tax adviser before you configure them.
Does my data train someone else's model?
This is the question most finance teams ask first, and the answer should be in writing. Under Article 28, a processor may only process your data on your documented instructions. Using your invoices to improve a general model is a different purpose, and you should expect a contract to exclude it, both between you and the vendor and between the vendor and every model provider it uses. Default settings on a provider's consumer product are not a contract.
What should you ask a vendor before connecting your books?
Put these questions in writing and expect answers in writing.
- Where are documents, extracted data, prompts, logs and backups stored, and in which region is the model run?
- Can any request be routed outside the EU, for example on failover?
- Who are your sub-processors, where are they established, and how will you tell us about changes?
- Which of your staff can access customer data, from where, and is that access logged?
- Do your contracts with model providers exclude training on our data, and do they allow the provider to keep prompts? For how long?
- What do you log by default, and how long is it kept?
- How do you separate the statutory record from working copies, and what happens to our data when we leave?
- How quickly will you tell us about a breach? Under Article 33, a processor must inform the controller without undue delay, and the controller generally has 72 hours to notify the authority.
A vendor that answers these quickly has thought about them. A vendor that needs a week to find out probably has not.
How we approach this at Thred
We hold ourselves to the list above. Your data is stored and processed in European data centres, with no transfers to third countries. It is used to do your work and nothing else, and our contracts with every model provider we use say it never trains a model. Data is encrypted in transit and at rest, with credentials and tokens in a dedicated secrets vault. Access for your team is role-based, with single sign-on and two-factor authentication, and Thred staff access to customer data is restricted, logged and reviewed.
Every entry can be reviewed and reversed, and nothing is deleted from the record. Thred is a Dutch company under EU law, a data processing agreement is part of every contract, and the sub-processor list is available on request. ISO 27001 certification is in progress. Agents read by default and write only with your permission, as described on our security page.
If you run finance for a growing business, the same agents work on top of the ledger you already have through Thred Layer, and firms can apply them across client administrations with Thred for accounting firms.
Want to walk through the data flow for your own books? Book a demo.
Frequently asked questions
Is an EU hosting region enough for GDPR compliance with AI tools?
No. The region usually describes the main database only. Prompts may go to a model endpoint elsewhere, logs and error reports may be shipped to other services, and staff outside the EU may have remote access, which the EDPB treats as a possible transfer.
Do invoices and bank transactions count as personal data?
Often they do. Invoices from sole traders, employee expense claims, payroll records and bank lines naming people or carrying IBANs all relate to identifiable individuals, so the GDPR applies to them.
How long must a Dutch company keep its bookkeeping records?
The Belastingdienst requires 7 years for the administration and 10 years for data about real estate. Supplies declared under the EU One Stop Shop schemes must also be kept for 10 years. Check your own situation with your accountant.
Can someone ask me to delete their data from my ledger?
The GDPR right to erasure does not apply where you must keep data to meet a legal obligation, which covers statutory bookkeeping records. Working copies such as prompts and logs have no such basis and should be deleted on a short schedule.
What should a data processing agreement with an AI vendor cover?
It should set out what the vendor may do with your data, list or govern its sub-processors, cover security and breach notification, and say what happens to the data at the end of the contract. Ask for model-provider terms that exclude training on your data.
Sources
Read next
- Connecting Claude or ChatGPT to your accounting system: what is safe
- Switching accounting software without losing your history: what to migrate, what to archive, and when not to switch
- Have you really outgrown your accounting software? A test to run before you migrate
Where Thred comes in
Hosted in the EU, GDPR compliant, encrypted at rest and in transit. Thred reads by default and writes only with your permission.



