What an AI assistant must know before it acts

Answering is not acting

A generic AI assistant produces impressive answers. Ask it to write a payment reminder and it will do it well. But it does not know which customer is concerned, whether the invoice was disputed, whether a commercial gesture was already made, or who must sign. Its answer is brilliant and isolated.

Acting inside a business requires something else: knowing what you are talking about, having the right to talk about it, and knowing your limits. Without that, AI remains a writing tool. With it, AI becomes an agent.

Context: what the agent must know

Useful context comes down to a few elements, always the same, whatever the task.

• The object concerned: the customer, the contract, the project, the equipment, the ticket.
• Its history: previous exchanges, decisions already taken, exceptions granted.
• The company’s rules: reminder delays, approval thresholds, responsible people.
• The current state: what is in progress elsewhere, who is available, what has already been tried.

That context only exists if the information is connected. An agent plugged into tools that do not talk to each other rebuilds a partial version of reality, just like the teams before it. That is why a shared system is the prerequisite, and why the Business OS comes before the agent.

Allowed: what the agent has the right to see

Seeing everything is not a quality. An agent that accesses all of the company’s data is a risk, not an asset. The right principle is role-based access: the agent sees what its scope allows, like any member of the team.

The same applies to what is passed to AI models. What is necessary, not everything. A financial advisor does not need the full ledger to explain a variance. A reminder agent does not need the personnel files.

Confidence: knowing when to hold back

A useful agent knows its level of confidence. When the signal is clear and the rule established, it acts or prepares the action. When the signal is ambiguous, it asks. When it lacks data, it says so, instead of filling the gap.

This behaviour is verified rather than promised. An agent that cites its sources, states its confidence and leaves a record can be checked. An agent that asserts without showing cannot.

A good answer says what it knows, what it assumes and what it does not know.

Boundaries: what it never decides alone

An agent’s boundaries are management decisions, not technical settings. They come down to three safeguards we detailed in Automate without removing human judgment: a threshold, an approval, a record.

In practice, an agent can prepare a reminder, create a ticket, notify the right team or flag a variance. It does not trigger a payment, does not modify a contract and does not send a sensitive communication without approval. You stay in control, and the record proves it.

Murphy, the agent that keeps watch

In Neoo, that agent is called Murphy. It does not wait for your questions. It monitors your sales, invoices and tickets, detects delays, prepares reminders, flags anomalies and triggers authorised actions, according to your company’s rules.

The typical scenario unfolds in a few steps: a signal appears, Murphy puts it back in context, proposes a framed action, obtains approval when it is required, then executes and records. The same mechanism applies to technical incidents, described in See an incident before it becomes an emergency.

What Murphy knows, it knows because the company is connected. What it does, it does because you authorised it. That is the difference between an assistant that talks and an agent that works for you. You can see it in action on the Murphy page.