Evaluating an AI provider: questions to ask before signing

Does the provider train on the data you send, where does it store it, how long does it keep it, and who can access it? Questions before adopting an AI service.

Topic
AI
Author
ARLITECH expert team
Published

Adopting an AI service often starts with a trial account: a few employees try it, find it useful, and the question soon becomes when everyone gets access. Before the contract is signed, though, it pays to go through a short evaluation, because AI services raise several questions that a typical software purchase never asks.

For the terms of use of the tools employees work with, the article on sensitive data that ends up in prompts gives a short checklist. This one is about preparing the organizational contract: what to ask the provider, which document to look for the answer in, and how to turn the answers into a decision.

Why an AI service differs from an ordinary SaaS

An invoicing or CRM cloud service stores data in predefined fields, which makes it fairly clear what scope of data reaches the provider. The first difference is that scope: in a service built on a large language model (LLM), users type free text and upload any file they like. What reaches the provider depends on what the employee happens to be working on that day.

The second difference is that submitted content can be used in several ways. Beyond producing the response, the provider may store the conversation for the history, use it to improve the model, or retain it for abuse monitoring. Training is especially sensitive, because a specific piece of data cannot be reliably removed from a trained model afterwards.

The third difference is how the service behaves. The model does not always give the same answer to the same question, and when the provider updates the model, the quality and style of the output can change too, while the interface and the contract stay the same.

The fourth is that the service is often assembled from the work of several companies. Behind the interface of a specialist application, another company's model may be running in a third company's cloud. Features beyond chat, such as file storage, connectors to email or document repositories, and web search, each open another data path. The specific questions raised by tools that carry out actions on their own are covered in the article on AI agent permissions.

Data use: training, retention, human review

The terms for the programming interface (API) are often in a separate document from those of both the consumer and the business plans. The evaluation should always look at the document that applies to the service actually ordered and that the contract incorporates by reference. A promise on a product page or in a blog post is not enough on its own, because the provider can rewrite it at any time.

Training

A phrase like “to improve our services” can mean model training, fine-tuning, internal quality measurement, or human labeling. Ask:

  • whether the provider uses prompts, uploaded files, and outputs to train any of its models,
  • if not, whether this is set out in the contract or is only an administrator toggle whose default may change later,
  • whether feedback (for example, rating a response) sends the entire conversation, and whether the opt-out covers that too,
  • what the provider means by “aggregated” or “anonymized” usage data, and what it uses it for.

Retention

Ask about the retention period for each data type, because a single service has several stores: the conversation history, the uploaded files, the searchable index built from the files (which a RAG-based feature answers from), the logs kept for abuse monitoring, and the backups. Ask whether deletion by a user removes the data from all of them, and how long it takes; what happens to the data when the contract ends; and whether there is a setting in which the provider stores nothing once the response is complete. Where such a setting exists, it usually comes with conditions, and some features, such as the history, are not available with it.

Human review

Many providers automatically screen submitted content for signs of abuse, and flagged conversations may be reviewed by staff. The same can happen when a bug report is investigated. The contract should make clear who may carry out the review (the provider's own staff or a subcontractor), from which country, under what confidentiality obligation, and whether the organization can ask to be excluded from this process.

Data location and subcontractors

The question “where is our data” needs two separate answers. One is where it is stored: which region holds the history, the files, and the indexes. The other is where it is processed: where the model that produces the response runs. The two can differ, because the provider may route requests to another region for capacity reasons. If the contract says European data storage, check whether that also covers processing.

If data leaves the European Economic Area, Chapter V of the GDPR requires an adequacy decision (such as the Commission's 2023 decision on the EU-U.S. Data Privacy Framework, for U.S. organizations certified under the framework) or appropriate safeguards, typically the standard contractual clauses adopted by Commission Decision (EU) 2021/914. According to the European Data Protection Board's Guidelines 05/2021, remote access from a third country, such as a support team viewing the data, may also count as a transfer, even if the data stays stored in Europe.

Sub-processors

Under Article 28 of the GDPR, a processor may engage another processor only with the controller's prior written authorization. With a general authorization, it must give advance notice of intended changes so that the controller can object, and it must pass its own obligations on to the subcontractor. In practice this means a public list of sub-processors and a notification process. In the evaluation, check:

  • who is on the list, and what role each one has (hosting, model, support, content filtering),
  • how, and how far in advance, notice of a new sub-processor arrives, and what to do if the organization does not accept it,
  • if the service is built on another company's model, on what terms the data reaches that company: whether it trains on it, how long it keeps it, and whether this also appears in the contract.

Encryption belongs here too: TLS in transit and encrypted storage media at rest are the expected minimum. Some services let the organization keep the key used for storage in its own key management system, so revoking the key makes the stored data encrypted with it unreadable to the provider as well. This does not necessarily extend to data being processed, or to caches, logs, and indexes protected with a different key.

Access management, single sign-on, and logging

An organizational AI service has to fit into the existing identity management. If users sign in with a separate password, the organization's conditional access rules do not apply, and multi-factor authentication and disabling the accounts of departing staff become manual work that sooner or later gets skipped.

Sign-in and accounts

  • Single sign-on (SSO). Does the service support SAML 2.0 or OpenID Connect with the organization's identity provider, and can password sign-in be turned off alongside it?
  • Your own domain. Can the organization verify its email domain, and can it take over accounts that employees created earlier on their own initiative with their work address?
  • Automatic provisioning. Does the service follow directory changes through SCIM or some other method, so that a departing employee's account is removed together with the directory entry (part of the access lifecycle)?

Roles and features

With roles, the question is who can turn what on. Who can install a connector, who can create a public share link for a conversation, who can share a custom assistant with the organization, and can each of these be blocked separately? With connectors, it matters whether they reach the document repository with the user's own permissions or through a broadly privileged service account that sees more than the person asking. For programmatic access, check who can create an API key, whether it can be restricted to a project and to specific permissions, and whether an expiry can be set.

Logging

Two kinds of logs are needed. The administrator log records sign-ins, settings changes, shares, and exports. Content access logs (who asked what) may be needed in an internal investigation or a legal dispute, but they contain personal data themselves, so access to them has to be kept narrow. Ask how long the provider keeps the logs, and whether they can be pulled through an API or another standard method so they can go into the organization's SIEM. If the log is visible only in the web interface, it is hard to correlate with events from other systems during an investigation.

Model changes and service levels

Providers update their models regularly. In the chat interface the user usually cannot choose a version, but the API typically lets you pin a specific model version, which the provider retires on a deprecation schedule announced in advance. Where the organization builds its own process on the model (document processing, classification, drafting text for clients), a model change can alter accuracy, format, and the behavior of the built-in filters.

In the evaluation, ask:

  • how far in advance the provider announces the retirement of a model or a change of the default model,
  • whether the new version can be tried before the switch,
  • whether there is a public changelog for models and features.

For your own processes, keep a test set of tasks with known good answers, including ones where the model has hallucinated before, and rerun it before a model change; testing is covered in more detail in the article on validating model output.

Availability and support

For the service level agreement, the availability percentage is not enough. Check how it is measured, whether a significantly slowed response counts as downtime, what you get when the commitment is breached, and how the request limits (requests per minute, daily quota) compare with the planned usage. For support, what counts is the response time and the channel for reporting issues, and for outages, a public status page.

Notification of security incidents is a separate point. Under Article 33 of the GDPR, the processor notifies the controller of a personal data breach without undue delay, but it is worth setting a more specific deadline and content in the contract. Finally, exit: can the history, the settings of custom assistants, and the logs be exported, and will the provider confirm deletion in writing after the contract ends?

Recording the answers and making the decision

Record the answers in a table, each with its source: the clause of the contract or the data processing addendum, or the title of the documentation and the date it was downloaded. “We don't know” is also a valid answer, as long as it is written down. A verbal promise or a sales presentation is not a source.

AreaAnswer to recordTypical source
TrainingWhether the content is used for training, and whether this is a contractual commitmentBusiness terms, data processing addendum
RetentionRetention period by data type, deletion at the end of the contractData processing addendum, documentation
Human reviewWho, from where, for what purpose, whether an exclusion can be requestedData processing documentation
LocationRegion of storage and processing, legal basis for transfersContract, standard contractual clauses
SubcontractorsList, roles, notice period, how to objectSub-processor list, data processing addendum
AccessSingle sign-on, provisioning, roles, API keysAdministrator documentation
LoggingLog types, retention, exportabilityAdministrator documentation
ModelsVersion pinning, retirement notice, changelogDeveloper documentation
Service levelHow availability is measured, limits, incident notification, exitService level agreement, contract

A common outcome is that the service can be approved for some categories of data and not for others, so it helps to put the organization's data classification next to the decision. Also record which features stay turned off (for example, public sharing or certain connectors), who gets access, and which gaps the organization accepts knowingly. That last one is a risk acceptance, and it should have an owner and a review date.

The evaluation is not a one-off: when to revisit it is covered in the article on AI governance roles. Article 21 of the NIS2 Directive lists supply chain security among the mandatory risk management measures. For financial entities, Article 30 of DORA sets out the mandatory elements of ICT service contracts, including the location of data processing, and for contracts supporting critical or important functions, an exit strategy as well. From August 2, 2025, Article 53 of the EU AI Act requires providers of general-purpose AI models to give documentation to providers that integrate the model into their own AI systems. This is a usable source if the organization integrates the model into its own system itself (for example, through the API), or if the provider of the service passes it on; an organization that only uses the service as a deployer is not entitled to it.

Related service: AI security controls deployment →

Contact

Ready to secure your critical systems?

Our experts will assess where your security stands and build a roadmap that fits your organization.

ARLITECH's office in Gödöllő

Visit us

You are welcome to visit our Gödöllő office in person.

Address
Centrum Office Building, 2100 Gödöllő, Dózsa György út 13, 3rd floor, office 309
Office hours
Mon–Fri, 9 a.m. to 5 p.m.
Get directions (opens in a new tab)