The four automation levels, explained

Every vendor now says its product is AI-first. The question a guard company should ask is narrower: for each thing the software can do, does it do it on its own, and if so, how would I know? Automaata answers that with four levels. Every action the platform takes carries one of them, the level is set by what the action touches, and it does not change because the office is busy.

Level 0: drafts and files

Work that produces a document for a person to use. The incident report drafted from an officer's voice note. The daily activity report compiled from a shift's scans, tasks and photos. The reminder that a guard card expires in 21 days. The weekly brief in plain English. Nothing at Level 0 changes a schedule, a payment or a record of what happened; it changes how fast a person can read and act.

Two rules apply even here. AI-written text is labeled, and the original input is kept beside it, so a report that becomes evidence shows exactly what the officer said and what the assistant added.

Level 1: executes and logs

Routine actions that are reversible and fully recorded. Publishing a tour schedule that no one edited. Sending a routine DAR to the client portal when a shift ends. Filing a missed-checkpoint exception. Setting a scheduling block on the day a credential lapses. Each of these happens without a person, and each leaves an audit record: which agent, which model version, what it saw, what it did. If one is wrong, it can be undone, and the record shows it was undone.

Level 2: proposes, a person approves

Anything that changes who works, how much they earn or what a client pays. Shift assignments and overtime. Pay-run approval. Invoice release. Here the platform does the whole job up to the decision: the solver builds the roster and explains every assignment, the billing agent drafts the invoice with every hour tied to a shift, the cost of an overtime proposal is shown before anyone clicks. Then it stops, and the proposal waits in the approval inbox with its evidence.

This is the level the law cares about. California's rules on automated decision-making technology treat work assignments and compensation as significant decisions that require notice, a human review path and records. Level 2 is how a platform can automate the work without automating the decision.

Level 3: never automated

Discipline. Termination. Licensing determinations. These belong to a manager and a conversation. The platform can supply the records a manager needs, and it can refuse to schedule a guard whose card has lapsed, but it will not propose a disciplinary outcome, rank officers for termination, or decide whether someone meets a licensing standard. That is a product decision, not a limitation of the technology, and it is written into the data model so it cannot be switched on by a setting.

Why fix the line by consequences

The alternative is to let each company decide how much to automate, which sounds like flexibility and turns into risk. The office that is short-staffed on a Friday is exactly the office that would publish a solver's roster without looking, and exactly the office that will be asked, months later, why an officer with a lapsed card was on a post. Fixing the level by what an action touches means the busy night and the quiet night get the same safeguards.

It also makes the product honest to explain. "Automate everything" is true of the work. "Decide the important things yourself" is true of the decisions. Both can be true at once, and the four levels are how.