Protect Your Data When Using AI Tools
The practical rule is simple: don't paste anything into an AI tool that you wouldn't be comfortable having stored on a server you don't control and read by someone who doesn't work with you. That includes personal data about clients and employees (ID numbers, salaries, medical records), credentials and access keys, contracts with confidentiality clauses, proprietary source code, and any financial or commercial information that gives you an edge over the competition, such as your cost structure or your bid prices before you submit them. The risk isn't that the AI will "steal" the information: it's that the data stays in the history, gets used to train future models, or is exposed in a security incident at the provider. The good news is that almost all of this can be managed with three things: your account settings, the contract you sign, and a clear rule for your team.
The difference between retention and training: two distinct risks
A lot of people lump them together, and that leads to bad decisions. It's worth separating them.
Retention is how long the provider stores what you typed. Almost all of them store something: your history so you can go back to past conversations, and temporary copies to detect abuse of the platform. Even if you delete a chat, a copy may remain for a window of several days in review or backup systems. That isn't necessarily bad, but it's something you need to know before pasting an entire contract into a chat.
Training is whether your text is used to improve the model. This one is more sensitive for one reason: once information has gone into a training process, it can't be "pulled back out." You can ask for your account and conversations to be deleted, but a model that's already been trained can't be untrained on request. That's why the question "is this used for training?" has to be asked beforehand, not afterward.
Generally speaking, free or consumer plans tend to come with training turned on by default and an option to disable it, while enterprise plans and API access usually exclude it by contract. But this changes often and varies by provider, so don't rely on what someone told you or what you read in an article from last year: the only valid answer is the one you read today on the official page of the tool you use, in its privacy document or its business terms. Download that version as a PDF and note the date you checked.
What data should stay out of any AI chat
- Sensitive personal data belonging to others: ID numbers, home addresses, account numbers, medical diagnoses, HR files, performance reviews with full names.
- Credentials: passwords, API keys, tokens, database connection strings, certificates. If you've ever pasted a key into a chat, rotate it today; don't assume deleting the message is enough.
- Documents under a confidentiality agreement: if you signed an NDA with a client, uploading that material to a third-party tool can be a breach even if the AI never "publishes" it.
- Non-public financial information: financial statements, actual margins, cost structure, projections.
- Active competitive strategy: prices you're about to bid in a tender, terms of an open negotiation, acquisition plans.
- Proprietary source code and internal architecture: especially if it includes business logic that sets you apart.
And one category people forget: meeting recordings and transcripts. Tools that join the call and generate minutes capture everything that was said, including the informal chatter before and after the official topic. Check where those transcripts are stored and who can see them.
What you can use without much concern
This article isn't meant to make you stop using AI. Most day-to-day work can be done with information that isn't sensitive: drafting an email from general bullet points, summarizing a public document, generating content ideas, checking spelling and style, translating marketing material, explaining a technical term, outlining the structure of an internal procedure.
For things that are sensitive, there's a trick that works well and costs nothing: anonymize before you paste. Instead of "Client Inversiones Zapote S.A., corporate ID 3-101-XXXXXX, owes ₡4.2 million since March," write "A client with an overdue balance of several months." The AI will still help you draft the collection notice, and you add the real details afterward, in your own document.
How to review the settings of the tool you already use, today
You can do this in the next twenty minutes, with no technical help:
- Go into your account settings and look for a section called "Data controls" or "Privacy." That's usually where you'll find the "improve the model for everyone" toggle or something similar. If it's on and you handle client information, turn it off.
- Check whether your team signs in with personal accounts (Gmail, their own email) or with company domain accounts. If they're personal, the company has no visibility or control over anything they upload, and the day that person leaves, the history leaves with them.
- Look at which integrations you've authorized. If an AI tool has read access to your Google Drive or your email, it doesn't matter what your employees paste: it already has access to everything. Revoke anything that isn't actively in use.
- Export your conversation history and skim the last two months. It's the fastest way to find out what your team is actually pasting. There are usually surprises.
- Find the provider's privacy document and business terms page on their website. Note two things: whether they use your data for training and how long they retain it. Save a PDF with the date you checked, because these policies change.
Which clauses to request in writing from an AI provider
If you're going to contract a service that processes your business information — not just a chat, but an agent that works with your documents — ask for these four things to be included in the contract or in a data processing addendum:
- No training: an explicit statement that your information is not used to train or fine-tune models, whether their own or third parties'.
- Retention period and deletion: how long they keep the data, what happens when the contract ends, and how many days it takes to complete deletion, including backups.
- Subprocessors: which third parties touch your data (the model provider, the cloud it runs on, the monitoring service) and which countries they're hosted in.
- Incident notification: how many hours it takes them to notify you of a breach, and who they notify.
If a provider can't answer this in writing within a couple of days, that's a signal. Not because they're dishonest, but because they probably haven't worked it out.
When you need to register a database with PRODHAB
In Costa Rica the framework is Law No. 8968, Protection of the Individual with Regard to the Processing of Personal Data, and its regulation, Executive Decree No. 37554-JP, overseen by the Agency for the Protection of Residents' Data (PRODHAB, the national data protection authority).
It's worth clearing up a common misconception here: not every database has to be registered. Under articles 5, 21 and 24 of Law 8968, registration with PRODHAB applies to databases that are sold or distributed for profit and to those held by public entities. A private company's internal database — your client list, your payroll, your CRM — doesn't require prior registration. That doesn't exempt you from the law's other obligations: informed consent, a declared purpose, security, and respect for the data subject's rights all still apply.
If your situation does fall under mandatory registration, article 47 of the regulation and PRODHAB's guidance require, in summary:
- Identification of the controller: name, personal or corporate ID number, and exact address of the party responsible for the database and of the processor.
- Purpose and use: why the data is collected and processed, described clearly.
- Types of data: which categories of personal data are collected, without including the data itself.
- Collection and updating procedure: how the data is collected with informed consent and how often it is updated.
- Recipients and interconnections: whether the data is expected to be transferred or shared with third parties, and with whom.
- Security measures: the technical, physical and administrative controls against alteration, loss or unauthorized access.
- ARCO rights: the channels through which a person can exercise access, rectification, cancellation and objection.
- Proof of fee payment: the annual registration fee set by PRODHAB. Check the current amount on the agency's official website before filing the application, since it is updated by resolution.
What the law requires when your data leaves Costa Rica
Almost nobody connects this part with AI, and they should: if the provider processes your information on servers outside the country — which is the norm — you are making an international data transfer. Article 9 of Law 8968 and articles 29 through 32 of the regulation require three things.
First, prior, explicit informed consent from the data subject, telling them who the recipient is and for what purpose their data is being transferred. Second, an adequate level of protection: the transfer is only lawful if the destination country or the receiving entity guarantees a level equal to or higher than what Costa Rican law requires. Third, if that destination doesn't have an adequacy declaration from PRODHAB, you need contractual safeguards — standard clauses — signed between the party exporting the data and the party receiving it, formally requiring the recipient to maintain the standards of confidentiality, custody and security.
The exceptions to prior consent are few and narrow: international treaties and instruments in force that Costa Rica has ratified, international judicial cooperation, and reasons of public health or urgent medical care for the data subject. None of those covers "I hired a tool and it's more convenient this way." In practice, before connecting an AI system to your CRM, it's worth reviewing two things with a lawyer: how the consent your clients sign is worded, and whether your contract with the provider includes transfer clauses. It's cheaper than fixing it after a complaint.
What no provider can guarantee you
Let's be clear about the limits. No one can guarantee that a cloud service will never suffer a security incident; what you can demand is that they notify you quickly and have reasonable controls in place. No one can guarantee that a data use policy won't change next year; that's why it's worth reviewing it periodically and keeping a copy of the version you accepted. And no one can reverse training that's already happened: if your information went in, it went in. That's why the important decision is made before you paste the first document, not after the first scare.
Getting this in order doesn't take weeks of consulting: it takes an afternoon to review settings, a one-page set of rules for your team, and an honest conversation with each provider. When we design Axiomify AI agents for a company, that conversation is part of the kickoff: what data goes in, where it's stored, for how long, in which country, and who can see it, written into the contract before the first line of work. If you're already using AI without having those answers, start with the reviews in this article; they're free and they'll tell you a lot about where you stand.
This guide is general information. Timelines, costs and requirements can change; for a specific case, review it with the right person before deciding.