Most production AI failures are not model failures. They are permission failures. The agent did exactly what it was configured to do — and that configuration was wrong for the edge cases that arrived three weeks after launch.
Defining autonomy tiers before deployment is not a theoretical exercise. It is the decision that sets your rollback cost.
The Four Autonomy Tiers
Every agent deployment sits at one of these levels:
- Read-only. The agent observes data and logs findings. It takes no action. A human reads the output and decides what to do.
- Suggest-only. The agent produces a recommended action. A human approves or rejects it before anything executes.
- Execute-with-review. The agent acts immediately, but every action is logged and surfaced for human review within a defined window — say, 15 minutes. A human can reverse the action inside that window.
- Fully autonomous. The agent acts and the action is considered final. No review step. No reversal window.
Each tier has a legitimate use case. The problem is not autonomy itself. The problem is choosing the wrong tier for the wrong task, or drifting to a higher tier without a deliberate decision.
Why Autonomy Creep Happens
Teams almost always start at suggest-only. That is the right instinct. But suggest-only creates a queue. Someone has to process that queue. When the queue grows, the pressure to remove the review step grows with it.
The conversation sounds like this: "The agent has been right 94% of the time for six weeks. The review step is just slowing us down. Let's flip it to autonomous."
That logic is not wrong. It is incomplete. The 6% error rate was tolerable at suggest-only because a human caught every error before it executed. At fully autonomous, that same 6% executes without review. At 200 actions per day, that is 12 bad actions per day running unchecked.
Autonomy creep is a volume problem disguised as a confidence problem. The agent did not get more reliable. The consequences of its errors just got larger.
A Concrete Example: Lead Routing
A B2B sales team deploys a lead routing agent. The agent reads inbound form submissions and assigns each lead to a sales rep based on territory, deal size, and product line.
At Suggest-Only
The agent produces a routing recommendation. A sales ops manager reviews the queue twice a day and approves assignments.
Edge case: A lead comes in from a company that is already a customer — a potential upsell, not a new logo. The agent routes it to the new business rep. The manager catches it, reassigns it to the account manager. No damage.
Cost of the error: 30 seconds of manager time.
At Fully Autonomous
Six weeks later, the team removes the review step. Volume is 80 leads per day. The manager cannot review every assignment anyway.
The same edge case arrives. The agent routes the existing customer to the new business rep. The new business rep cold-calls a contact who has been a customer for two years. The account manager finds out. The customer is annoyed. The deal stalls.
Cost of the error: one damaged relationship, one stalled renewal, two hours of internal cleanup.
The agent did not change. The tier changed. The edge case was always there.
The Right Path
The correct move is not to stay at suggest-only forever. It is to move to execute-with-review first. The agent routes the lead immediately — no queue — but every assignment is visible in a review feed for 30 minutes. The manager skims the feed, not every item, just the flagged ones. Flags trigger on known edge-case patterns: existing customer domains, deal sizes above a threshold, leads from blocked territories.
That design captures 90% of the speed benefit of full autonomy while preserving a catch window for the cases that matter.
How to Set the Tier Before Deployment
Three questions determine the right starting tier:
- What is the cost of a single wrong action? If the answer is "annoying but reversible in under five minutes," execute-with-review is probably fine. If the answer is "a damaged customer relationship or a compliance event," start at suggest-only.
- What is the expected error rate on edge cases? Not average accuracy — edge case accuracy. Most agents perform well on the common case and poorly on the tail. Estimate the tail volume.
- Is there a reversal mechanism? Execute-with-review only works if the action is actually reversible inside the review window. Lead routing is reversible. Sent emails are not. Tier selection must account for reversibility.
Document the answers before the first deployment. Revisit them at 30 days with real production data. Move tiers deliberately, not under queue pressure.
Boring wins. An agent running at the right autonomy tier for 12 months beats an agent that demos at full autonomy and rolls back in week four.
If you are scoping an agent deployment and working through the tier decision, Start a conversation →