An AI assistant that writes nothing without your confirmation

Short answer: by default NG Copilot in NG CRM only reads and answers. When an administrator enables action proposals, every write still goes through a card that shows what will be created or changed, old value against new, and waits for your click. More sensitive actions also ask you to retype the amount or enter a code from your authenticator app, and some need a colleague's approval. Only the “Action outcome” block — written by the server after it checks the database, not by the language model — can state that something was actually written.

This is not about distrusting AI. It is the same rule sales teams already apply to people: nobody changes someone else's opportunity without a trace, and nobody approves their own risky decision.

Why “AI that acts on its own” is the wrong goal for a CRM

A CRM is a record of promises: what the customer was offered, when, on what terms and who is accountable. An assistant that quietly edits those records causes damage you do not see right away. You see it later, when the forecast stops matching reality or a customer receives a quote nobody reviewed.

So NG CRM reverses the order. Copilot reads and answers first, and every answer carries its origin: From your data, From the help docs, From the internet or “from the model's general knowledge — unverified”. When the data is not there, it says so instead of offering a number that merely sounds right.

Four steps from a sentence to a write

1. Proposal. You ask in plain words — “schedule a call with this customer on Thursday”. Copilot builds a card listing what will be created or changed, with every value, old against new.
2. Confirmation. You click Confirm. A more sensitive action also asks you to retype the amount or number from the card, or enter a code from your authenticator app. A proposal you do not confirm in time expires.
3. A second person. Losing an opportunity, merging lead duplicates and emailing a quote to a customer wait for a colleague whose approval limit covers them. Once approved, you execute the action yourself.
4. Outcome. The server records the state: Executed, Awaiting your confirmation, Awaiting approval, Rejected or Failed. Only that block claims a write happened.

Quotes and campaigns that Copilot prepares are created as drafts. Nothing reaches a customer without approval.

Rules that do not depend on how good the model is

• Copilot cannot do more than you. It works with your permissions and data scope. Company policy and your own My AI settings can only narrow it, never widen it.
• The proposer never approves their own request, even if their limit formally allows it. If nobody else in the company can decide, Copilot says so before the request is created.
• No recorded limit means no authority. “Not configured” never means “unlimited” — the approval task cannot be resolved until a limit is set.
• Voice can request an action but not confirm it. Confirmation always happens on screen. Repeating a sentence does not create a second record, because each action carries its own key.
• Emergency controls. An administrator can instantly switch off all AI, only AI writes, or put the system into read-only mode. Each activation requires a reason that is permanently logged.

What this means for the team

Sales reps get speed where the work is mechanical: finding records, preparing drafts, scheduling activities. Managers keep control where decisions are made: losing a deal, merging customers, sending a quote. Auditors get a trail — who asked, who approved, when, and what was actually written.

Policies, answer sources and internet search are covered on the NG Copilot page. How permissions, limits and the audit trail work across the system is on the Security page.