Trust

How we earn trust

We are asking banks, public agencies and households to let us near the most sensitive data people have. That access has to be earned structurally, not promised in a paragraph.

In place
True of the system as it runs today.
By design
An architectural commitment we are building against, not yet fully built.

Regulated rails, licensed partners

In place

Account access runs through licensed PSD2 providers. We never ask anyone for banking credentials and we never scrape.

  • Data reaches us only through licensed open-banking providers and partners.
  • Access is read-only. Balansa cannot move money and is not built to.
  • No screen-scraping, no credential sharing, no browser automation against a bank.
  • The regulated route was the starting condition, not a later upgrade.

EU-resident models, no training on your data

In place

Inference runs on models hosted in the EU, and financial data is never used as training material.

  • Models are EU-hosted, and can be pinned to a specific region where a partner requires it.
  • No financial data is used to train or fine-tune any model, ours or a vendor’s.
  • Where a partner needs stronger guarantees, reasoning can run against a self-hosted model.
  • Nothing is sent to a general-purpose consumer AI service.

The model sees as little as possible

By design

Agents get tunnel vision on purpose: aggregated inputs, role-scoped visibility, and a deterministic path where reasoning is not actually needed.

  • Data is aggregated before it reaches a model. A pattern is enough; a line-by-line history rarely is.
  • Each agent role sees only the slice its job requires — a briefing agent and a fraud agent do not share a view.
  • Agents are constrained to the question in front of them and cannot roam beyond their instructions.
  • Once a signal is well understood, it stops being a prompt and becomes a plain deterministic rule — cheaper, auditable, and incapable of improvising.

Raw transactions stay with the customer

By design

We do not keep raw transactional data. It lives on the customer’s device or in their own cloud storage, and is reached only with just-in-time access.

  • Raw transaction records are never persisted by Balansa by default.
  • Where records are retained at all, it is with explicit consent and under the customer’s control.
  • Storage sits on the customer’s own device or their own cloud account — not in a Balansa data lake.
  • Access uses just-in-time tokens scoped to one specific purpose, and expires with it.

Every delegated step is consented and revocable

By design

Anything a customer hands to Balansa is verified, explicitly agreed to for a named purpose, and can be withdrawn at any moment.

  • Consent is granted per use case, not as a single blanket permission at sign-up.
  • Each delegated step is verified before it runs, and recorded so it can be reviewed.
  • Revocation is immediate and available at any time, without contacting support.
  • Withdrawing consent stops the processing it covered — it does not degrade into a lesser default.

We ship recipes, not data access

By design

Balansa supplies the workflow, the processing rules and the prompts. The data can stay somewhere we never reach.

  • What we distribute is a method: how to normalise, what to watch for, how to phrase the question.
  • That method can execute in the customer’s or partner’s environment, on data we never see.
  • What comes back to us is feedback and impact measurements, not financial records.
  • The measure we care about is whether households move onto a reversing trajectory — that number does not require us to read anyone’s statement.

The short version

The less Balansa can see, the more useful Balansa can be. Every design decision above trades our own access away for something a bank or an agency can actually approve. We would rather ship a method that runs on data we never touch than build a system that depends on holding it.

What we handle today, in concrete terms, is set out on the data and privacy page.