Ask ten operations leaders how mature their organization is with AI-driven automation and you'll get ten different, often overly optimistic, answers. Part of the confusion is that "using AI" covers everything from a single ChatGPT prompt pasted into a spreadsheet to a fully autonomous agent making unsupervised decisions at scale. Those are wildly different states of maturity, and conflating them leads to poor planning, because the next investment that makes sense at one level can be actively counterproductive at another.
We use a four-level model to help teams honestly locate where they actually sit, which is usually one or two levels earlier than they initially assume.
At this level, workflows are automated with deterministic, rules-based logic: if this field matches this value, take this action. There's no model making judgment calls, just explicit logic a developer wrote in advance. This level is often underrated. A huge amount of genuine operational value comes from simply eliminating manual, repetitive steps with straightforward scripting, before any AI model is involved at all. Organizations that skip this level and jump straight to AI agents often end up automating a mess instead of a process, because they never did the basic work of standardizing the rules-based parts first.
Here, AI models start participating, but only in an advisory capacity. A model might score a lead, suggest a response, or flag an anomaly, but a human reviews and acts on every single output. This is where most organizations that describe themselves as "using AI heavily" actually sit. It's a legitimate and often high-value stage, since it lets a team validate model quality against real outcomes with a human safety net on every decision, but it caps the efficiency gain, because a human is still touching every instance.
At this level, the model acts autonomously on a defined subset of cases, typically the high-confidence, low-risk ones, while routing everything else to a human. This is the level where genuine efficiency gains start to compound, because the human is now only reviewing the ambiguous or high-stakes fraction of volume rather than all of it. Getting here safely requires the verifiability groundwork we've written about elsewhere: a reliable way to know which cases are high-confidence enough to act on without review, and a clear, fast escalation path for everything else.
At the top of the model, agents operate independently across the full range of cases within a defined scope, with human oversight shifting from reviewing individual decisions to monitoring aggregate performance and handling genuine edge cases. Very few processes in most organizations are actually ready for this level, and that's not a failure, it's appropriate. Level 4 makes sense only for workflows with enough historical volume, low enough downside risk per instance, and strong enough monitoring infrastructure that drift gets caught quickly.
In our experience, most operations teams that believe they're at level 3 or 4 are actually solidly at level 2, with a few processes that have graduated partway into level 3. That's not a criticism, it's the natural and often correct pace of building trust in a system before removing human review from more of it. The mistake isn't being at level 2, it's not realizing it, and therefore either underinvesting in the verification infrastructure that would let you safely advance, or overinvesting in autonomy features for processes that aren't ready for them yet.
The right question isn't "how do we get to full autonomy," it's "what would need to be true for us to safely trust this specific process one level further than we do today." Answered honestly, process by process, that question does more to advance real maturity than any roadmap built around the destination alone.
.webp)