Once a company has a few AI projects behind it, a question comes up that sounds simple and isn't: how far should the system actually go? Can the AI make a decision on its own, or only prepare one? When does someone have to sign off, and when is monitoring enough?
The question usually gets asked too late. A system goes live, usage grows, and at some point somebody notices that nobody ever defined which decisions the AI can make on its own. This isn't an abstract governance problem. It's an operational risk with real consequences.
We use a five-level model for this. L0 is the digital starting point, followed by five levels, L1 through L5. They describe how deeply AI reaches into a process, from pure support to fully independent execution. With each level, the human role shifts from active decision-maker to supervisor and, ultimately, to an authority that intervenes only when necessary.
L0 describes the digital starting point. Data and workflows are captured and accessible. Every operational step, every decision, and all oversight still sit with people. AI plays no part at this level. Companies that haven't reached this point yet need to establish the digital foundation before introducing AI into the process.
At L1, the AI pulls together information a person would otherwise have to search for, prioritizes it, adds context, and presents it where the decision is made. An underwriter opens an application and sees the relevant risk signals. Comparable prior cases show up on a claims adjuster's screen. The decision stays entirely with the person.
L1 is the easiest place to start. No human decision disappears, so the risk stays manageable. Whether people actually use it depends on whether the information is more useful than what was available without AI.
At L2, the AI prepares more than information. It prepares an action: a draft response, a proposed decision, a routing suggestion. The person reviews it and either confirms or corrects.
The step from L1 to L2 carries more weight than it first appears. At L1 the person acts on their own initiative and the AI supplies the basis for the decision. At L2 the proposal already exists in finished form, and the person's job is to judge it. Judging a finished proposal is a different mental act from making a decision from scratch. The cognitive load shifts, and the risk profile shifts with it. Run L2 without tracking how often staff genuinely review a suggestion and how often they just click through, and you have no reliable read on process quality.
From L3 on, the AI acts independently. Defined process steps run without confirmation. Someone supervises the process by monitoring metrics, patterns, and outliers, and steps in when something deviates from expectations. Every individual case goes through without anyone watching that individual case.
That raises the requirements considerably. How would I know the system is deciding wrongly on a systematic basis? Which monitoring metrics actually tell me something? How do I escalate once a problem becomes visible? Those capabilities have to be in place before L3 goes into operation. The same applies to the scope of authority, which defines what the system may decide on its own and what remains reserved for human decision-making. That scope belongs in the system definition.
At L4, the process runs end to end automatically, inside clearly defined limits. The system decides, acts, and documents. People steer only when the system spots an exception and escalates it.
The word "limits" is doing the real work in that sentence. L4 works when cases that fall outside the defined scope are reliably identified and handed off. A routine claim with clear facts, processed from start to finish without an adjuster touching it, is L4. A system that escalates every unfamiliar case and sends most of the workload back to human adjusters is effectively operating at L2, just from the other direction.
L5 goes beyond predefined scenarios. The system can pursue a goal, adapt its actions to changing conditions, and execute the process without operational human involvement. People retain ultimate authority but are no longer part of day-to-day execution.
Most business processes don't meet the prerequisites for that yet: reliable system integration, rigorous testing, and continuous observability. In practice, agentic AI at L5 is still a way off.
The right degree of automation depends on the process itself. Every step up raises the demands on data, process standardization, governance, and the ability to run AI systems stably. A company can run L4 on standardized small auto claims and L1 on complex large losses at the same time. The most common mistake in practice is setting a high degree of automation as a strategic target before benefit, trust, and process maturity are there to support it.
The degree of automation is a process decision. Assess the characteristics of the process first: its variance, its data quality, its governance maturity. Then choose the level that fits, and you have a decision you can defend. Set an automation target first and then go looking for processes that fit it, and sooner or later you'll build systems that either can't be run in production or escalate more than they decide.