NearSync Help

AI

NearSync AI Three layers, one configuration, and where the AI actually turns up in your day.

The AI is not a feature in one place. It runs in three layers, and this section is where all three are configured.

The three layers

The assistant. Somebody asks a question and gets an answer. It knows which HQ you are in, so the same question means something different in Finance than it does in Sales.

Agents. Nothing is asked. Something runs on a schedule, looks at your data, and files work when it finds something. Nobody has to remember.

AI inside the modules. Drafting a reply, summarising a meeting, classifying a conversation. It does not announce itself as AI, and it is the layer most people use without noticing.

Where you actually meet it

The AI is used in the surfaces it belongs to, and each is documented there rather than here.

Where What it does Read
The cockpit Ask, research, and request changes Talking to the assistant
Command bar Ask a question from anywhere Asking from search
Out loud Speak instead of typing Talking out loud
Sync Chat The assistant in a channel AI in Sync Chat
WhatsApp Answering customers AI assistant settings
Meetings Notes, summary and actions What Scribe does

If you are trying to use the AI rather than configure it, those are the pages you want.

What this section owns

Configuration and governance. Nine things, one per surface.

How it is set up. Where the intelligence comes from, and whether you supply it.

Choosing your model. Which model answers, and what changes when you switch.

Teaching the assistant. Giving it your own material so answers are about your organisation.

Persona and guardrails. How it sounds, and what it will not discuss.

AI agents. The scheduled ones that work without being asked.

The prompt library. Good questions, saved and shared.

Comparing models. Testing two or three against the same prompt before committing.

Watching AI usage. What is being used, how fast, and what is failing.

Credits and what uses them. What a credit is and what spends one.

What a question actually costs you

Worth knowing before you roll it out, because it changes how you talk about it.

Asking the assistant something is not free, and it is not expensive either. It draws on an allowance, and a normal person asking normal questions all day does not get close to the edge of it. The things that consume noticeably more are the ones that run without a person: agents on a tight schedule, and anything processing long recordings.

That is the whole shape of it, and Credits and what uses them has the detail. The practical consequence is that you should not ration questions. Telling a team to use the assistant sparingly produces an organisation that never learns what it is good for, and saves very little.

Managed or your own keys

This is the fork that determines most of what you see.

Managed. The intelligence is provided. There is nothing to connect, no keys to hold, no vendor account to open, and no bill from anybody but us. Providers show as active with an allowance, and that is the whole administrative burden.

Your own keys. You hold accounts with the model vendors and supply the keys. You pay them directly, you see the true cost, and you decide entirely which vendors are involved.

Self-hosted and on-premise deployments are always the second kind, because there is no shared infrastructure to pool. Everybody else is managed by default, and it is the right choice unless you have a specific reason otherwise. See How it is set up.

What the AI can see

It answers from your organisation's data, and it respects who is asking.

Somebody who cannot see salaries does not get salary figures by asking the assistant instead of opening the page. The permissions are the same permissions; the assistant is not a way around them, and there is no separate AI permission model to keep in step.

That is worth telling people explicitly, because the usual worry about an assistant over company data is exactly this, and the usual answer elsewhere is less reassuring.

What it is not

It is not a decision maker. It drafts, summarises, finds and suggests. Sending, approving and committing are acts a person takes.

It is not a replacement for the numbers. For anything that has to be right, the governed measures in Analytics are the answer, and the assistant will point you at them rather than compute its own version.

It is not always on for everything. Each surface is separately switchable, so a tenant can run the assistant and no agents, or agents and no chat assistant.

One thing to decide early

Who owns the AI. Not who set it up, but who keeps the knowledge current and reviews what the agents produce. Without a name, it works for a quarter and drifts after.

Rolling it out

Two failures are common and both are avoidable.

Switching everything on at once. People meet the assistant, three agents, AI in chat and AI on WhatsApp in the same week, form one impression, and it is usually "this is noisy". Turn on the assistant, let it settle, then add agents.

Setting it up and announcing it. An assistant with no knowledge of your organisation answers the first questions generically, and generic answers are how a team concludes the AI is not for them. That impression is expensive to reverse. Load your material first.

Where to start

Setting it up for the first time, read How it is set up, then Persona and guardrails, then Teaching the assistant. That order matters: an assistant with your material and your tone is a different product from one without.

Just want to use it, go to Talking to the assistant.

5 minUpdated 28 July 2026

Did this answer your question?

No, ask a person