How Howe protects your data

Why it matters in a tool like Howe

Howe works with sensitive things: your listings, your prices, your orders and the credentials it uses to connect to Mercado Libre and Tiendanube. It also has an AI agent that can propose changes to all of that. So the question isn't only where the data is stored, but what each piece can do and who has to agree before anything changes.

This post goes through the answers in order: your data, your team, the agent and the layers that protect them. It covers what Howe does today; what doesn't exist yet isn't here.

Your data is yours and kept apart from everyone else's

Every business, isolated

Every business lives isolated from the others. It isn't just a rule written in the code: the database enforces it too, with row-level security. So a badly written query can't show another business's data, because the database doesn't hand it over. And automated tests check that one business can't use another's data.

Credentials, encrypted and kept away from the model

To read your listings and apply the changes you approve, Howe connects to Mercado Libre and Tiendanube with your authorization, through each one's official API and without asking for your password. The credentials that come out of that authorization are stored encrypted and never reach the browser or the AI model.

What we store and what we don't

We store only what's needed. We store nothing about your Tiendanube buyers: no names, no emails, no addresses. Howe doesn't train models on your business's data. If you ask to close your organization, its data is deleted; only what doesn't identify anyone stays in the audit log. And every night a backup copy is made, encrypted and kept off the server.

The detail of which personal data is handled, what for and for how long is in the privacy policy.

Everyone sees and does what's theirs

Every team member has a role, from whoever manages everything to whoever can only look. Permissions are checked on the server and not on the screen: hiding a button isn't enough as protection, so the real check happens where nobody can touch it.

Connecting a sales channel is the account owner's job. Signing in takes a password of at least 12 characters, failed attempts are limited and sessions expire.

The agent proposes, you approve

An AI agent that can touch prices needs limits that don't depend on its good judgment. In Howe they are these:

• It only proposes. The agent's write tools only add changes to a list that you approve. No change reaches your channels without going through the change engine.
• What it reads is data, not instructions. Supplier lists, messages and descriptions are treated as information. If a file said "ignore your rules", the agent reads it as text, not as an order.
• You choose how much it does on its own. It can ask you every time, follow your rules or apply everything. A price at which you'd lose money is never applied automatically, in any mode.
• Everything can be undone. Every change keeps how it was before: one click and it's back.
• There's a record. It notes who changed what and when, in a log nobody can rewrite.
• Approvals expire. If nobody attends to a pending change, the approval expires instead of staying open forever.

You approve from the web, from your phone or from WhatsApp. How the agent works inside is in the post An AI agent for your store.

Layers that assume the previous one can fail

The underlying idea is defense in depth: every layer assumes the previous one can fail. The code filters by business and the database requires it again. Permissions are checked on the server. Processes run with a database user that has no administrator privileges. And the app reaches the internet through a Cloudflare tunnel, with no open ports on our servers.

There's also a cap on AI usage per plan, so a mistake doesn't turn into a surprise.

The standards we work with

We take these standards and regulations as our reference:

• OWASP Top 10 and ASVS 5.0, level 2, for application security: authentication, sessions, access control, validating what comes in and error handling. Every finding of a review is reported with its OWASP code.
• OWASP for AI and agents: the Top 10 for applications with language models and the one for agentic applications. Everything the model reads is data and not instructions, and on your channels the only thing the agent can do is propose changes.
• Argentina's Personal Data Protection Law 25,326, as a reference for which personal data is stored, what for and how it's deleted. When Tiendanube reports that a customer's data has to be deleted or handed over, Howe processes it.
• Platform requirements: OAuth authorizations, official APIs and notifications that are verified before they're read. We follow Tiendanube's and Meta's privacy requirements.
• Review and testing: every change that touches access, a business's data, notifications, credentials or what the model reads goes through a security review before it ships.

These are working references: they set how we build Howe, they aren't third-party seals.

If you find something

If you find a flaw or something that doesn't look right, write to hola@howe.com.ar with what you saw and how to reproduce it. Test only with your own account and without accessing other people's data. A summary of all this, with the same facts, is on the security page.