I want Phanes

Data security and independence from external AI vendors

Most "AI-powered" tools send your campaign data to someone else's model. Phanes has its own language model, Aether, running on our infrastructure. Your data does not have to travel anywhere to get a good answer.

Phanes team7 min read
7 days of PRO free · no card

An ad account holds more than numbers: a company's strategy, margins implied by its bids, sales seasonality and budgets competitors would love to see. When a tool analyzes it through an external AI model, that data goes to another company's servers, under that company's terms.

Phanes took a different approach. A deterministic rules engine makes the decisions, and Aether handles the language: Phanes's own narrow language model, trained only for Phanes tasks. It runs on Phanes's own compute infrastructure and does not pass account data to external model vendors.

Why independence from an AI vendor matters

An external model is convenient, but it is also a dependency. The vendor can change the model version, data processing terms, limits or pricing. Every such change affects the product built on it and the data flowing into it. The differences:

A tool on an external AI model
Phanes with Aether
Campaign data goes to the model vendor's servers.
Data stays on Phanes infrastructure.
The model version can change without notice, and answer quality with it.
A new version ships only after winning on a frozen test set.
The vendor sets the data processing rules.
Phanes sets the rules: anonymization, privacy audit, facts outside the weights.
A general model knows a little of everything, including things a campaign does not need.
A narrow model trained only on Phanes tasks, in Polish and English.
You have to check the numbers in the answer yourself.
The model is penalized in training for any number outside the input data.
Cost and availability depend on someone else's pricing and limits.
The model and its throughput are under our control.

Where your data flows

Here is how one recommendation travels from account data to the sentence you see on the card. Every step happens in the European Economic Area.

  1. Reading the data

    Phanes reads account metrics through the official Google and Meta APIs. Access tokens are encrypted with AES-256-GCM.

  2. Rules engine

    The deterministic engine computes the measurement verdict, priced waste and recommendations. This is where decisions are made.

  3. Knowledge base

    Phanes knowledge excerpts are retrieved for the answer through vector search (RAG).

  4. Aether

    Our own model turns the finished numbers and the rule into readable sentences. It runs on Phanes infrastructure.

  5. A card for you

    You see the recommendation with the rule, the numbers and a button. The change reaches the account only after your approval.

Account data, engine, knowledge base and language model in one closed Phanes loop.

Our own model, our own rules

Aether is not a copy of a large general model. It is a narrow, bilingual model built on open weights (Bielik and Qwen families) and tuned for Phanes tasks. Four principles set it apart from the typical "AI feature" in a tool:

Own infrastructure

Training and answers run on Phanes's compute infrastructure. Generation and training run in parallel, and the model trains in full precision.

Facts outside the weights

The model learns form, discipline and style. It receives facts and rules from the knowledge base at question time, so account data does not stay in the model.

Anonymization before training

No company names, people, addresses or account identifiers go into training. The dataset passes the same privacy audit as the rest of Phanes.

Verifiable rewards

The model learns with GRPO only from answers that pass automatic verifiers: numbers from the data, a valid schema, clean language, a citation of an existing rule.

What your company gains

Fewer parties with access to data

Campaign data is not passed to an AI model vendor and does not leave the European Economic Area. That is easier to document in your processing register and data processing agreement, with no transfers outside the EEA.

Predictable answers

The model version changes only after passing a release gate. This month's report is as dependable as last month's.

Continuity

A change in an external AI vendor's pricing, limits or policy does not change how Phanes describes your campaigns.

GDPR Article 28 requires processors to provide sufficient guarantees, and Article 32 explicitly lists encryption. The fewer companies touch the data, the simpler those guarantees are to demonstrate.

Resilience and quality control

Owning the model also means owning its quality. That is why Aether has mechanisms that usually stay with the vendor:

  • A frozen Phanes evaluation set, separately in Polish and English, with tests for the verifiers themselves.
  • A target of 100% of answers without numbers outside the data, checked before every release.
  • Rollback to the previous checkpoint when a new version does worse in any category.
  • Hard validation on the Phanes side before an answer is shown to the user.

Data security starts with asking how many companies can see it. In Phanes your campaign data stays in one loop: your account, our engine and our model.

Does Phanes send my campaign data to OpenAI, Google or another AI vendor?

No. Descriptions, chat answers and classifications come from Aether, Phanes's own language model running on Phanes infrastructure in the European Economic Area. Phanes only reads account data from the Google and Meta APIs and executes approved changes.

Does the model learn from my account's data?

The model learns answer form and discipline from anonymized examples, without company names, people or account identifiers. Facts about your account do not go into the model's weights; they are only used at answer time.

What if a new model version is worse?

It will not ship. Every version must beat the previous one on a frozen test set with no regression in any category; otherwise the previous checkpoint stays.

Can Aether change anything in the account by itself?

No. The model has no write tool. A deterministic rules engine makes decisions, and you approve every change after API validation and a diff preview.

Sources (3)