Progressive autonomy: the trust ladder for AI in compliance

How much control do you actually want to hand an AI agent on day one? For most compliance and fraud teams, the honest answer is: not much, yet. That instinct is correct, and it's exactly what progressive autonomy is designed for.

‍

Think about how autopilot entered aviation, or how self-driving cars are classified by levels of driving automation. Nobody went from fully manual to fully autonomous in one step. Control was handed over in stages, and each stage was earned. Progressive autonomy applies the same idea to risk operations. Rather than treating autonomy as a single on/off switch, it breaks AI deployment into five deliberate levels, a trust ladder your team climbs one rung at a time, per queue, with evidence at every step. Here's what each rung actually looks like, and why it isn't the same question as how sophisticated the underlying agent is.

‍

Watch the walkthrough

Kunal Datta, Unit21's CPO, walks through the five levels, why teams set them queue by queue, and how to QA the top rung.

‍

This isn't the same question as agent sophistication

We've written before about the four levels of agentic AI in financial crime, which classify how capable an AI system is: whether it's still a copilot, genuinely agentic with human oversight, or fully autonomous end to end. Progressive autonomy is a different axis. It assumes you already have a genuinely agentic system and asks a separate question: how much of that system's output are you, today, willing to act on without a human checking it first?

‍

A highly capable agent can sit at the lowest rung of the trust ladder in your queue for months, on purpose. That's not the agent underperforming. That's the process working as designed.

‍

The five levels, L1 to L5

The trust ladder has five levels:

‍

• L1, No AI: AI features are turned off for the queue. The analyst does everything manually.
• L2, Start Review (AI on standby): the AI agent is available, but an analyst has to trigger it. It won't run on its own. This is deliberate, to keep automation bias from creeping in before your team trusts the output.
• L3, Auto Review (AI ready): the agent runs automatically as alerts land. It packages the data, sources, and analysis into a summary, so the work is ready when an analyst opens the case. The analyst still makes the call.
• L4, Recommendation: the agent recommends a disposition, escalate or close, and lays out its reasoning based on your configuration. The analyst approves or overrides. Nothing is executed without them.
• L5, Auto Close: on defined, low-risk case types, like alerts the agent and your historical data agree are clear false positives, the agent closes the case itself, with a full supporting narrative and documentation.

‍

‍

Every level above L1 keeps a human in the loop somewhere. The ladder just changes where in the process that human sits.

‍

The two guardrails that make every rung defensible

Two mechanisms run underneath every level of the ladder, not just the top ones:

‍

Random sampling: pulling a sample of cases where the AI agent and analyst agreed, and a sample where they disagreed, for ongoing QA. This is what lets you catch drift before it becomes a pattern, not after. It matters most at L5: teams running auto-close should regularly sample closed cases and have a human review them, so precision stays high and anything closed that shouldn't have been gets caught early.

‍

Backtesting: running a newly configured agent against historical alerts before it ever touches a live queue, so you have evidence of how it performs before it has the chance to underperform on something real.

‍

These two guardrails are also why L5 Auto Close doesn't mean the agent is deciding on its own without a net. If you want the deeper argument for why false positive reduction has to be evidence-based rather than just faster, how to reduce false positives in AML transaction monitoring covers that ground well.

‍

How fast teams actually climb it

In practice, most teams follow a crawl, walk, run pattern. They start in AI assist mode at L2, Start Review, where the analyst prompts the agent and validates the output against their own judgment before trusting it on anything else. Once the output is consistently trusted, usually somewhere between one and two months in, teams move the same queue to L3, Auto Review: the agent runs automatically the moment an alert lands, so by the time an analyst opens it, the summary and evidence are already sitting there.

‍

That shift, from "let me ask the agent" to "the agent already ran," is where teams typically see time-to-investigate drop the most. It's not a bigger model. It's more trust, earned one rung at a time.

‍

Who decides how high you go

The ladder is configurable per queue, not platform-wide, and customers never start at the top. Different teams, and different processes within the same team, have different appetites for automation. A team might run L4 on a well-understood alert type with years of historical dispositions behind it, and L2 on a newer, less-tested one, at the same time. That's the design working as intended, not an inconsistency to fix.

‍

Human-in-the-loop is the default here, not a hedge added for regulators' sake. Your existing QA and QC processes still apply to an AI agent's output the same way they'd apply to a human analyst's. Interestingly, some teams are now running that relationship in both directions: using AI to check human analyst work for consistency and SOP adherence, not only the reverse. You own the outcome either way. The same human-review principle shows up beyond alert triage: with SAR narrative automation, the agent drafts the narrative and extracts the information, and a human reviews before anything is filed. If you're building the underlying tasks that feed a specific rung of the ladder, how custom AI agents are transforming fraud and AML operations is the practical next read.

‍

FAQs

‍

What are the five levels of progressive autonomy?

L1 No AI, L2 Start Review, L3 Auto Review, L4 Recommendation, and L5 Auto Close. Each level moves the human checkpoint later in the process, and teams set the level per queue rather than platform-wide.

‍

Does progressive autonomy mean the AI can't be trusted at first?

No. It means trust is built deliberately, with evidence, rather than assumed. Most agents are capable of more than L2 from day one; the ladder is about how much of that capability your team has verified against its own data.

‍

Can autonomy levels differ by queue?

Yes. A queue with years of historical dispositions to backtest against might reach L4 quickly, while a newer alert type stays at L2 or L3 until there's enough evidence to move it up.

‍

What stops an agent at L5 from closing something it shouldn't?

L5 Auto Close is scoped to defined, low-risk case types only, backed by backtesting before go-live and ongoing random sampling afterward. It's not a general autonomy switch.

‍

How do you QA cases the AI auto-closes?

Through random sampling of auto-closed cases, reviewed by a human to confirm precision, on top of backtesting the agent against historical alerts before it goes live.

‍

Who is accountable if an AI agent's disposition turns out to be wrong?

Your team is, in the same way it would be for a human analyst's decision. The audit trail shows exactly what the agent looked at and why, which makes that accountability easier to demonstrate, not harder.

‍

How long does it typically take to move from AI assist to auto-review?

Most teams see one to two months of AI assist use (L2, Start Review) before moving a queue to auto-review (L3), though this depends on alert volume and how much historical data is available to backtest against.

‍

Climb the ladder on your terms

Progressive autonomy isn't a compliance requirement bolted onto AI as an afterthought. It's the mechanism that lets a team move from skeptical to confident without ever having to take that leap on faith. See how Unit21's AI Agent supports this rollout, queue by queue.

‍