Noblle

Voice and text

What an AI agent can safely do in a factory, and what needs a human

A practical framework for deciding which factory actions an AI agent can take on its own, which need confirmation, and which should stay with a person.

By Noblle · · 8 min read

Short answer

An AI agent in a factory can safely answer questions and surface alerts on its own, take low-risk reversible actions with a quick confirmation, and prepare consequential decisions for a person to approve. It should always act within the speaker's permissions, record every change in an audit trail, and never commit money, materials or customer promises without explicit human confirmation.

Key takeaways

  • Sort actions by consequence and reversibility, not by how clever the agent is.
  • Reading and alerting are low risk. Writing is higher. Committing money or promises is highest.
  • Permissions should follow the person, not the agent. If you cannot do it in the system, you cannot do it by asking.
  • An audit trail turns an agent's action into something accountable and reviewable.

Start with consequence, not capability

Language models are good at understanding a request and bad at knowing which request is dangerous in your business. The safe design is to decide, in advance and in plain rules, what the agent may do, and to enforce it outside the model.

A useful way to think about it is three tiers: things it can do freely, things it can do after a confirmation, and things it can prepare but a person must decide.

Tier one: read and alert

Answering questions and pushing alerts changes nothing, so the risk is mainly wrong or leaked information.

  • Order status, stock levels, schedule, who is on which machine.
  • Warnings about a job at risk of running late, a material short against a committed order, a quote about to expire or an invoice overdue.
  • Summaries of the day, week or a particular customer.
  • Controls: only show data the person is allowed to see, and say so when the answer is uncertain.

Tier two: act with a confirmation

Routine changes that are easy to reverse are good candidates, as long as the agent repeats back what it is about to do and waits for a yes.

  • Rescheduling a job within existing capacity.
  • Marking a delivery received or a job stage complete.
  • Creating a draft order from a customer message for a person to review.
  • Controls: read the action back clearly, require a simple confirmation, log who asked and what changed, and make it reversible.

Tier three: prepare, then a human decides

Some decisions carry money, materials or reputation and should not be taken by an agent alone, however confident it sounds.

  • Accepting an order that commits materials and capacity.
  • Issuing a quote or promising a delivery date to a customer.
  • Raising or approving purchase orders above a threshold.
  • Changing prices, credit terms or invoices.
  • The agent's job here is to gather the facts, show the consequences and recommend. The person approves.

The guardrails that matter

Whatever the tiers, a few controls should be non-negotiable.

  • Permissions follow the person. The agent has no authority of its own beyond theirs.
  • Every action is logged with who, what, when and the original request.
  • Ambiguity triggers a question, not a guess. If two jobs match the description, it asks which.
  • There is a clear way to undo, and a human can always take over.
  • Access is reviewed regularly, especially for voice and messaging channels where devices are shared or lost.

Frequently asked questions

Can an AI agent be trusted to change data in an ERP?
Yes, within limits: respect user permissions, confirm consequential actions, log everything and keep high-impact decisions with a person.
What should an AI agent never do in a factory without approval?
Commit money, materials or delivery promises, such as accepting orders, issuing quotes, approving large purchases or changing prices and invoices.
How do you audit what an AI agent did?
Require a log of who asked, what was changed, when, and the original request, so every action can be traced and reversed if needed.

Related in the platform

See how this works in your factory

We configure the walkthrough to your own process before the call.