istGalaxy.
Trust

Safe AI is built as a partnership.

You draw the boundaries; the system stays inside them. Brakes, isolation, encryption and audit are our job — what may be said, what is off-limits and where work hands over to a human are your calls.

partnership

The division of duty is clear from day one

Safe deployment is not a one-sided promise but a clean split of responsibility. What we carry and what stays in your hands is explicit at every step.

We ensure

  • Brakes run on every answer — an unsupported claim never reaches a customer
  • Workspace isolation: no other customer can ever touch your data
  • Access keys stored encrypted; data encrypted in transit
  • Systems monitored continuously; nightly exams run every day
  • Every change written to the audit log

You define

  • What the agent knows and answers from — the knowledge base is yours
  • Forbidden topics and boundaries — you say where it stays silent
  • Procedures and human-handover points
  • Which systems are open as tools
  • Which scenarios are tested before release

secure by design

Safety isn't bolted on; it's built in

You describe, test and approve the agent's behaviour before a customer ever sees it. Two tools carry that work:

Procedures

Describe multi-step work with its limits: the agent branches on conditions, stops at sensitive moments, hands over to your team. No path you haven't defined is ever walked.

Procedures

Rehearsals and the release gate

Run rehearsals that mimic real conversations; an agent that fails the exam cannot ship. The live agent is re-examined every night.

Testing

on our own business

We built this architecture for ourselves first

We run Galaxy on our own customer traffic every day: it books our appointments and answers our questions. In the last 60 days, 82.7% of our conversations resolved without a human; median written reply time is 1.6 seconds.

Measured in our own operation's panel; definitions are open on the Analytics & Quality page.

security & privacy

Your data is yours; it stays that way

Every layer follows the same principle: the data belongs to you, we only process it, and every action leaves a trace.

Data rights built in

Export customer data with one click and erase it permanently on request. The right to be forgotten is a panel feature, not a paperwork process.

Encrypted storage

Your connection keys and access credentials are stored encrypted; data is encrypted in transit as well.

Layered isolation

Each workspace is isolated at four separate layers; one customer's request technically cannot reach another's data.

Audit log + signed messaging

Who changed what and when — recorded with before and after. Every notification from an external system passes signature verification.

Never used for training

Your data is not used to train AI models; the way we work with model providers rules it out contractually.

Access under your control

Role- and module-based permissions: who on your team sees what, who changes what — you define it.

reliability

You are not tied to a single provider

The system is built to absorb trouble before your customer notices it.

Model redundancy

The agent works with more than one AI provider; if one slows or fails, the conversation switches to the backup model automatically.

Continuous monitoring

Systems are watched around the clock; anything unusual lands in front of our team within minutes.

Brakes at every layer

Answers pass checks before reaching the customer; when certainty is missing, the agent stays silent and hands over.

our promises

Three promises, in writing

Not marketing lines — rules coded into the product:

We don't charge for handed-over conversations

Conversations handed to a human or marked as spam don't count against your quota and aren't billed. If the agent couldn't solve it, you don't pay for it.

Unsupported numbers never reach customers

Answers containing figures without a source are stopped by design — 100% of them. A human steps in instead.

Nothing ships untested

An agent change cannot face customers without passing the rehearsal exam; and what shipped can be rolled back with one click.

Frequently asked

Is my data used to train AI models?

No. Your data belongs to you and is not used for model training; our arrangements with providers rule it out contractually.

How do you prevent made-up answers?

Every answer must stand on recorded content in your knowledge base and passes brakes before reaching the customer. With no grounding, the answer is stopped and the conversation hands over — no guessing.

Can I export or erase my data?

Yes. Export and permanent erasure live inside the panel; the rights data-protection law grants don't require a separate request.

What happens if an AI provider goes down?

The conversation switches to the backup model automatically; your customer doesn't notice. Systems are also monitored around the clock.

You draw the lines. We keep it inside them.

In your setup week we define the forbidden topics, handover points and test scenarios together.