Someone on your team has found an AI tool that drafts replies in seconds. They want to try it on real work: a WhatsApp message asking about opening hours, a complaint with a customer's name and phone number in it, or a supplier order email the tool should turn into a spreadsheet row. The tool is ready for work. What remains is to find out whether your business is ready to use it.

Before the first real customer message goes in, one person should read the AI tool data processing terms. A DPA outlines the parties' rights and responsibilities concerning personal data.¹

This article is a reading guide for that hour. It covers four areas: training, storage and deletion, who else handles the data, and what happens after a breach or when a model is withdrawn. It also names the few answers that mean you should not use the tool at all.

What a data processing agreement is, in plain words

A data processing agreement, usually shortened to DPA, is a legally binding contract that sets out the rights and duties of two parties when personal data is handled.¹

Two words are worth knowing. For this example, call your business the controller and decide what you want the tool to do with customer data. For this example, call the vendor the processor and specify the tasks you want it to perform.

The DPA is not a minor side document. It is a central building block of modern data protection and security architectures.² If the DPA is not alongside the main terms of service, ask the vendor for it.

Ask for both: the general terms page and the DPA. Save a copy of each with the date you downloaded it.

Does your data train the vendor's model?

Look for a clear answer in the agreement: is the data you send used to train models, and how can you change that choice? If the clause allows it, a customer's complaint about a late order could become material the vendor learns from.

A good clause says plainly what happens. OpenAI, for example, states that since 1 March 2023 data sent through its API is not used to train its models unless the customer explicitly opts in.³ The API is the connection other software uses to send requests to the model. That is the kind of sentence you are looking for: a clear default and a clear way to change it.

Check consumer plans and business plans separately. A free app, a business plan and the API can have different data processing terms. Read the clause for the plan your team will actually use, not the one on the marketing page.

The same care applies later, when a tool becomes part of daily work. We describe how a person checks each step in building reliable AI workflows with human review.

Warning

Three answers block a tool outright:³

  • The vendor trains its models on your data and offers no opt-out.
  • The vendor offers no DPA.
  • Check whether the vendor offers a Business Associate Agreement, or BAA, for health data.

Do not read the third as a rule for every business in Latvia: a clinic that handles patient messages needs its own legal check on which agreements apply.

Vendors change their terms. Check the current version on the day you decide, not a copy someone saved last year.

How long the vendor keeps data, and how you get it back

A DPA should say what happens when the contract ends, including whether the vendor returns your personal data or deletes it.¹ Look for a clear time limit and for what "delete" actually covers.

Anthropic's DPA is a useful example. It says Anthropic will return data on request and delete it within 30 days after the agreement ends or expires, with exceptions that the DPA lists.³ Ask the vendor whether any of them could apply to your customers' messages.

Also ask how long messages are kept while you are still a customer. A short deletion period after the contract ends helps little if every message is stored for years before that.

Also check data portability. It means you can take your data out in a file another tool can open, so you can switch vendor or use the data for something else.¹

Here is why that matters. Imagine your team has spent a year drafting replies to customer questions with one tool. Those drafts, and the corrections staff made to them, show how your business actually answers people. If you switch tools, you want to keep them. Before you start, ask the vendor three things in writing: what can be exported, in which file format, and whether the history is included.

Who else handles your data

Sub-processors are third parties a processor engages to process personal data on behalf of a controller.¹ A DPA can allow the vendor to engage sub-processors if it enters a written agreement with data protection obligations no less protective than the main DPA.¹

In practice, a customer's WhatsApp message may pass through several firms before a draft reply comes back. You do not need to check each of those firms yourself. You do need to know who they are and where they operate.

The same goes for the tool's connections to your other systems. If the tool should read supplier order emails from your inbox, ask the vendor how its tool would get the data from that email and where the data goes next.

International transfers mean your data is handled outside the EU. Technical annexes, rules on international data transfers and the different style of legal documents make a DPA harder to read.² The annexes are often where the answers about countries and safeguards sit.

Ask the vendor for:

  • the current list of sub-processors
  • the country where each one handles data
  • how the vendor tells customers when the list changes

If a sub-processor is outside the EU, note which section of the DPA covers that transfer. If you cannot find the relevant section, ask the vendor for a written explanation.

Security and what happens after a breach

A DPA should name the security measures the vendor must take, such as encryption, access controls, staff training and regular security audits.¹ Encryption means scrambling data so that only people with the right key can read it. Access controls decide which of the vendor's own staff can see your data at all.

You cannot audit a large vendor yourself. What you can check is whether the DPA names concrete measures, or only promises "appropriate" security without saying what that means.

Then read the breach clause. Under GDPR, the processor must tell the controller without undue delay after it becomes aware of a personal data breach, and the controller must notify the supervisory authority no later than 72 hours after becoming aware of it.³ Whether a particular breach has to be reported is a question to settle with a lawyer, so write down when you first heard about it.

That clock is tight for a small team. Find out which email address receives the vendor's notices and who checks it. It may still be the person who signed up, who could now be on holiday or have left the company. Appoint a responsible person and a deputy, put their addresses in the vendor account's contact settings, and make sure both know what to do when a notice arrives.

What happens when the vendor withdraws a model

The model is the part of the tool that actually writes the drafts. Anthropic and OpenAI publish notice periods for retiring certain models.³ When that happens, your tool moves to a newer one, and drafts can change in tone, length or language quality. Look for model-retirement information separately from the DPA.

Notice periods differ between vendors. Anthropic gives at least 60 days' notice before it retires a publicly released model.³ OpenAI gives at least six months for generally available models, a shorter period for specialised variants and a much shorter one for preview models.³ If your team relies on a preview model, plan for short notice.

For a business in Latvia, Lithuania or Estonia, the main risk is language. A model that writes good Latvian, Lithuanian or Estonian replies may be replaced by one that writes them differently. After any change, check a sample of drafts in each language before staff trust them again.

Keep a small set of made-up customer messages for this check, often called synthetic data. Use them to check the quality of the new model's replies without entering real customer data.

How to do a first check of AI tool data processing terms in an hour

Rushed reviews are where mistakes happen. Time pressure and the lack of a common method are frequent causes of errors in DPA reviews,² and tight deadlines raise the error rate significantly.² The routine below gives one person a fixed method and the time to use it.

Appoint one reader and give them time

Choose one person. Plan about an hour per tool in which they can check the documents without interruptions. It helps if this is not the colleague who wants to start using the tool tomorrow.

Collect the current documents

Download the terms of service, the DPA, the sub-processor list and the page about model retirement. Save dated copies in one folder. Read the legally binding version of the terms, not just a summary.

Read the four areas

Training; storage, deletion and export; sub-processors and transfers; security, breach notice and model retirement. Note the section number for each answer you find.

Stop at a blocker

If the vendor trains on your data with no opt-out or offers no DPA, stop there.³ Reading the remaining clauses will not change that answer. If you handle health data, ask a lawyer which extra agreements apply before you go further.

Write down unanswered questions and ask in writing

Keep one shared page per tool: name, plan, date, answers found and unanswered questions. Email the unanswered questions to the vendor and keep the reply with the documents.

Until the answers arrive, staff can test the tool with made-up messages only. If you are still deciding where to start, our guide to choosing your first AI automation helps. You can also review how Lindevo works with a client and ask who approves what during a build.

Info

This is a reading checklist, not legal advice. If you handle health data, or if you cannot tell where your data goes outside the EU, ask a lawyer before anyone uses the tool.

Frequently Asked Questions

Do I need a data processing agreement to use ChatGPT or Claude at work?

A DPA sets out the rights and responsibilities of a data controller and a data processor concerning personal data.¹ Having no DPA is one of the answers that block a tool outright.³ Check which plan the DPA covers before anyone starts using the tool.

Does an AI tool use my customers' messages to train its model?

It depends on the vendor and the plan, so read the training clause for the plan your team uses. OpenAI, for example, says data sent through its API is not used for training unless you opt in.³ If training is switched on and there is no opt-out, do not use the tool.³

What happens to our data when we stop using an AI tool?

The DPA should say whether the vendor returns or deletes your data when the contract ends.¹ Anthropic's DPA, for example, says it returns data on request and deletes it within 30 days, with listed exceptions.³ Also check that you can export your data in a format another tool can use.¹

Who is responsible if an AI vendor has a data breach?

Under GDPR rules, if the vendor is the processor and your business is the controller, the vendor must notify your business without undue delay after becoming aware of a personal data breach, and your business must notify the supervisory authority no later than 72 hours after becoming aware of it.³ Check with a lawyer whether a particular breach has to be reported. Appoint a responsible person who receives the vendor's notice and acts on it, plus a deputy for when they are away.

Are the terms available in Latvian?

Check which languages the documents are available in. Ask the vendor whether a translation exists, and read the version that is legally binding. If a clause is unclear, ask the vendor in writing or ask a lawyer before you use the tool.

Sources (3)

Have a process you'd like to talk through?

Book a call (opens in a new tab)