istGalaxy.
Trust Architecture

An agent that admits what it doesn't know — and proves what it does.

AI's best-known flaw is making things up. Galaxy solves it with three layers: answers rest on your knowledge, a reply that breaks its word never reaches your customer, and critical decisions always end with a human. Your customer sees a clean answer; you see what it rests on.

Every answer

rests on a search of your knowledge base; its source is recorded in the panel

100%

of replies caught with unsupported numbers are stopped before reaching the customer

Human

the final word on payments, refunds and anything above a cap — by design

layer 1 · knowledge

It speaks from your knowledge, not its imagination

Before every answer, the agent searches the knowledge you approved — not a preference, a mandatory step. The answer rests on that search; the document behind it sits next to the conversation in your panel. On topics outside your knowledge it says so, or hands over — it doesn't guess.

Mandatory search

Looking at your knowledge base first is a system rule, never left to the model's whim.

Visible grounding

Each answer in the panel shows the document it rests on; auditing is one click away.

Out of scope = honesty

On unknown topics it says 'I don't know'; it doesn't speculate.

Prices and official fees come from a fixed table — the model is never allowed to generate those numbers.

layer 2 · brakes

A reply that breaks its word never reaches the customer

Every reply passes brake checks before it goes out: did it say 'sending the payment link' without sending one? Did it quote an unsupported number? Promise a discount above your cap? A reply that trips a brake is stopped — fixed or handed to a human. Complaints and legal topics escalate to your team automatically.

Say-do checking

A reply that claims to have done something it didn't is never sent.

Number brake

Any figure without a recorded source stops the reply.

Discount cap

A promise above the cap you set never reaches the customer.

The brakes work identically on the phone — rules don't loosen when the channel changes.

layer 3 · human

Critical decisions end with you

Money movement, refunds, order confirmation and publishing all end with a human: the agent proposes, your operator approves. Take over any conversation at any moment; the agent goes quiet when you do. Changes to the agent pass rehearsal first — and roll back with one click.

Approval gates

Payments, refunds and orders are final only with human approval or the customer's own button.

Instant takeover

The moment your team joins, the agent steps back; it never talks over you.

Rehearsal + rollback

Changes pass rehearsal conversations first; the live version rolls back in one click.

Even its own improvement suggestions never publish without human approval.

The gap between us and a bare model

Put a language model in front of your customers as-is and it won't cite sources, it can make up numbers, it can drift off-topic — that is its nature. Galaxy wraps the same model in grounding, safety brakes and a nightly exam.

Answers grounded in a source

Bare model

cites nothing

Galaxy

every answer rests on a recorded source

Unsupported numbers reaching your customer

Bare model

nothing in the way

Galaxy

100% stopped (by design)

Resolution rate — our own operation (real measurement)

June

82.0%

July

82.8%

August

85.2%

The first two charts are illustrative, not measurements: they describe how a bare language model behaves without these layers. The third chart is a live measurement from our own operation (last three months).

One agent, every job the conversation demands

Your customer doesn't think in departments; neither does your agent. In the same conversation it answers the question, books the appointment, moves the order — without switching desks.

In support

Answers with sources, solves the issue, hands over with a summary when needed.

Channels

In sales

Gets to know the prospect, qualifies with your questions, writes the deal, sends the quote.

CRM & Sales

In e-commerce

Recommends products, recovers carts; orders are final only with the customer's own button.

E-commerce

Frequently asked

What if the AI makes something up?

A reply with an unsupported number is stopped before reaching the customer; on out-of-scope topics the agent says so and hands over. And every answer's source sits in the panel — auditing is one click.

Who feeds the agent's knowledge?

You do: your documents, your site, the answers you approve. You upload and update from the panel, and choose exactly which content the agent may use.

Will my customers know they're talking to AI?

That's your call: you decide from the panel how the agent introduces itself.

I'm worried about losing control.

It's designed for the opposite: every rule lives in your panel, every conversation can be taken over, every change can be rolled back, every critical decision ends with a human.

Trust is built with architecture, not adjectives.

We handle the setup in your first week; from day one you'll see the source behind every answer in your own panel.