Client data and AI: what leaves your account and what stays

Connecting an ad account to an AI raises a legitimate question that rarely gets a direct answer: what exactly leaves my account, and where does it go? This article answers that at the operational level, without legalese and without alarmism.

What gets authorized, in practice

The connection happens through OAuth, the same mechanism you use to sign into a service with your Google account. The flow is:

1. You click connect and are taken to the platform itself, Meta or Google.
2. You sign in with your credentials, on their page, and see a screen listing what is being requested.
3. You approve, and the tool receives a scoped access token.

Three things follow from that: your password never passes through the tool; the scope is declared on the authorization screen; and access is revocable at any time, in the platform's own app settings.

Read and write are separate permissions

This distinction matters more than it appears.

Read. The tool queries spend, results, campaigns, metrics. It changes nothing.

Write. The tool can change budgets, pause campaigns, create items. It changes your account.

They are different scopes, requested separately. The practical recommendation: start with read only. Spend a few weeks trusting what you read before granting permission to change anything. Nothing is lost in that time — write stays available whenever you want it.

The question that separates tools

There are two designs, with quite different implications:

Neither is wrong. Historical reporting tools need to store, because the platform does not return everything forever. The point is knowing which one you are buying, and the question is direct: "do you store my account's data or query it live?"

What to ask before authorizing any tool

Five questions that fit in one email and that any serious vendor answers without hedging:

1. Which scopes do you request? Read, write, or both?
2. Is the data stored? If so, where, for how long and with what encryption?
3. Is the data used to train models? The answer needs to be explicit.
4. How do I revoke? And what happens to what was already stored.
5. Who at your company can see my account data?

If any of those produces a vague answer, that is already the answer.

Personal data: the line not to cross

Campaign metrics are not personal data. Spend, impressions, clicks and cost per result are aggregate numbers, and running them through an AI raises no individual privacy question.

What does raise one is different: a lead list with names, phone numbers and emails. Pasting that into a chat is a problem regardless of how good the tool is. Two practical reasons: you probably have no legal basis for that processing, and the data subject did not consent to that destination.

The rule that resolves nearly every case: aggregate metrics, yes; identifiable personal data, no. If you need to analyze leads, work with counts and rates, not with the list.

The agency side

Anyone managing someone else's account has an extra layer. Technically, you already administer the account and can authorize. Contractually, that is another conversation.

What avoids trouble:

• A clause in the agreement stating that analysis tools may be connected to the accounts, reading metrics.
• A list of connected tools, kept current, that the client can request at any time.
• Revocation at offboarding. When the contract ends, access goes with it. Worth being on the exit checklist.

The broader structure for agency relationships is in client reporting for small agencies.

The risk nobody mentions

The most common risk in AI and marketing is not a leak. It is undue confidence in a wrong number.

An AI reports a broken figure with precision and conviction. If the pixel fires twice, it will say conversions doubled, with the same confidence it would say anything else. If the date range is wrong, the comparison comes out wrong and sounds perfect.

The counterweight is the usual one: numbers going into an invoice, a contract or a results meeting get verified at the source. The conversation is for investigating fast, not for having the last word.

GDPR, CCPA and the proportionate answer

The short version: less than fear-marketing suggests, and more than most agencies actually handle.

Aggregate campaign metrics are not personal data. There is no identifiable subject in "US$ 12,000 spent and 340 leads". Connecting an ad account to a read-only tool does not, by itself, create personal data processing.

What does create it are three concrete situations:

• An exported lead list. Names, phones, emails. Here there is a subject, processing and responsibility. Do not paste it into a chat.
• Custom audiences by upload. Uploading a customer base to a platform is processing, and it needs a legal basis and notice to the subject.
• Agency-held client data. You are the processor, the client is the controller. The contract needs to say so, and the list of sub-processors — tools included — needs to be known.

None of those three is solved by picking the right tool. They are solved by contract and by process, and the tool enters as one line on a list. Which is why the clause in the agreement is worth more than any badge on a vendor's page.

A practical note: most ad platforms have been sub-processors in your operation for years. Adding a read-only analysis tool to the list is a change of degree, not of kind — provided it reads aggregate metrics, which is the design described above.

A short checklist before you connect anything

1. Read-only first. Always. Add write later, deliberately.
2. One account first. Connect a single account, run a week, compare what the tool says against Ads Manager.
3. Check the app list on the platform. Meta and Google both show every app with access. Review it quarterly and revoke what you no longer use.
4. Know the storage answer. Live query or stored copy, in writing.
5. Tell the client. Before, not after.

None of these steps is expensive, and together they turn a decision that feels risky into an ordinary operational choice.

The wider view of working with AI in paid media is in AI for media buyers, and the reporting structure on the paid media reporting page.