SECURITY & DATA

Your data stays yours, and it stays in your AWS region.

Northra reads commercially sensitive material: contract terms, margins, customer history. That only works if the handling of it is boring, explainable and verifiable. This page describes how it is built — and, at the end, what we do not claim yet.

Send us your security questionnaire What we do not claim →

RUNS ON: AWS, in the region you choose
MODEL TRAINING: None on your data
WRITES BACK: Nothing, by default

Where your data isSingle boundary

1 · Stored

Inside AWS, in your chosen region

Encrypted at rest with AWS KMS. EU customers can be deployed to an EU region so data does not leave that jurisdiction.

2 · Processed

Inside the same AWS boundary

Including the AI models. Prompts and business data are not sent to a third-party model API outside AWS.

3 · Separated

One isolated tenant per customer

Your data is not pooled with another customer's, and is never used to train a shared model.

One cloud, one boundary, one region.INTENDED ARCHITECTURE

This describes the intended architecture of a deployment, not a certified state. What we do not claim is set out below.

The short version: Northra reads, it does not take.Your systems stay the source of truth. Northra holds a working copy inside your AWS boundary so it can answer questions across sources, and it proposes actions for a person to approve. It does not write back into your ERP or CRM, and it does not change a price.

If you switch it off, your systems are exactly as they were.

Where it runs

Everything happens within AWS.

There is no second cloud, no separate analytics vendor and no external model API in the data path. That is a deliberate design choice: one boundary is far easier for your security team to reason about — and to audit — than four.

Your region, your jurisdiction

The workspace is deployed to the AWS region you choose. For European customers that means an EU region, so data and processing stay inside that jurisdiction.

Encrypted in transit and at rest

TLS on every connection; encryption at rest using AWS KMS. Keys and their rotation are managed within your deployment, not held in some side system.

The models run there too

AI inference happens inside AWS. Your prompts, documents and figures are not sent out to a public model endpoint, and are not used to improve anyone's foundation model.

One tenant per customer

Isolated storage and isolated processing. Nothing about your book is visible to another customer, and no shared model is trained on it.

Least-privilege connections

Northra connects with scoped, read-only service accounts wherever the source system supports them. It asks for the fields it needs, not for everything.

Exit is a first-class feature

Your data is exportable, and deletable, on request. A pilot that does not continue should leave nothing behind.

What it does with your data

Read, interpret, propose. Not act.

Northra does

Read from the sources you connect, with the scope your admins set.
Hold a working copy inside your AWS boundary so questions can cross sources.
Cite every output back to the record it came from.
Propose actions with an owner, a deadline and the evidence attached.
Keep an audit trail of what was read, by whom, and when.

Northra does not

Write back into your ERP or CRM, or change a price, by default.
Send your data to a third-party model provider outside AWS.
Use your data to train models that any other customer touches.
Pool your material with another customer's in shared storage.
Act on anything without a named person approving it.

Who sees what

Your administrators own the boundaries.

In a commercial system the sensitive question is rarely "is it encrypted" — it is "can a regional sales manager see another region's margins". That is a configuration your people control, not one we set.

Control	How it works	Owned by
Sign-in	Single sign-on through your existing identity provider. No separate Northra password to manage or leak.	Your IT
Roles	Role-based access. Each seat sees its slice — a region, a business unit, a product line — and not the rest.	Your admins
Source scope	Which systems are connected, and which fields within them. Narrow scopes are normal and encouraged.	Your admins
Audit trail	What was read, what was produced, who approved it, and when. Exportable as a record rather than a screenshot.	You, exportable
Deletion	Export and deletion on request, including at the end of a pilot.	You, on request

AI governance

An answer you cannot trace is not an answer.

The failure mode of an AI system in a commercial setting is not that it goes rogue. It is that it is confidently wrong, and nobody can tell — because the reasoning is not visible. Northra is built against that specific failure.

Every output is cited

A figure on a screen links to the record it came from — the index publication, the contract clause, the order line. If it cannot be cited, it is not shown as fact.

Data quality is visible, not assumed

Every feed shows its last refresh and its quality checks. Records that fail those checks are quarantined and never reach a briefing, rather than quietly degrading one.

A person approves before anything counts

Northra proposes; a named person decides. The approval, and the reasoning behind it, is part of the record — which is also what makes the next decision better informed.

Built with the European rules in mind

GDPR, the EU AI Act and NIS2 shape how the system is designed: purpose-scoped data, human oversight of consequential decisions, traceability, and an exportable record. We say "in mind" deliberately — see below.

Being straight with you

What we do not claim.

Northra is an early-stage company. Enterprise security reviews go badly when a vendor overstates its position early and has to walk it back later, so here is the honest picture.

We are not certified yet

We do not hold SOC 2 or ISO 27001, and we will not imply that we do. If a certificate is a hard gate for you today, we are probably too early — and it is better that we both know that now.

"Compliant" is not ours to declare

GDPR, the EU AI Act and NIS2 shape the design, and we will sign a data-processing agreement. But compliance is a property of your deployment and your processes, not a badge a vendor awards itself.

No uptime promises we cannot keep

We are not going to put an availability figure on a page to win a meeting. Service levels belong in a contract, agreed against a real deployment.

A pilot can start narrow

If the sensitive systems are the blocker, start without them. A pilot on market data, published information and a limited slice of your own book still shows whether the reasoning holds up.

Your security review

What we will give your team.

Your questionnaire, answered

Send the standard one your procurement uses. We complete it properly, and we mark the gaps as gaps.

An architecture walkthrough

A session with your security and IT people covering the AWS deployment, the data flows, the access model and the retention.

A data-handling note per pilot

Written before a pilot starts: exactly which sources, which fields, which region, who has access, how long it is kept and how it is deleted.

Next step

Bring your hardest security question first.

We would rather spend the first conversation on the data-handling note than on a demo. If it does not clear your review, nothing else matters.

Talk to us How a pilot works