Earned autonomy means an AI agent earns the right to act on its own from its own record, one level at a time. Each agent holds an autonomy level and a results score: the share of its counted results that went well. Every night its last 14 days are checked, and slipping work drops it one level; a clean record over 30 days nominates it for the next level, and a person decides.
The question every risk team has to answer
Before an AI agent goes live on a bank desk, someone has to decide how much it may do without asking. Read and draft only? Update a record? Send a letter to a customer? The answer is easy on day one: very little. The harder question is what changes after the agent has done good work for three months. A bank wants an answer it can defend to its board, its auditors and its regulator, and that answer has to rest on evidence.
Earned autonomy gives the decision a structure. The agent starts narrow. Its record builds up. A person widens its authority one step at a time, on that record, and the record keeps watching after the step is taken.
Step 1: a level for every agent
In Tuvel Control Plane, every staff agent holds one of five levels. Every action above its level waits for a person.
- L1, Observe. Reads and reports. Anything it would write, even a draft, waits for a person.
- L2, Assist. Reads and drafts on its own. Changes to records other people rely on wait.
- L3, Operate. Changes workspace records on its own. Anything that reaches outside the workspace waits.
- L4, Represent. Acts outside the workspace, such as emailing a contact. Publishing waits.
- L5, Lead. Publishes and hands work to agents below it. Deleting and governance changes always wait.
Bank agents start at L1 or L2. Of the 99 agents on Tuvel’s 14 desks, most start at L1: they read documents and the bank’s knowledge base, draft and propose, and people decide.
Step 2: a results score from the agent’s own record
A level says what an agent may do. A results score says how well its work has turned out. In Tuvel, each agent’s score is the share of its counted results that went well: an approval, a reply, a booking, a person’s rating. It shows once at least 20 results are counted, so a new agent has no score until there is enough to judge. It is refreshed every night, and a reviewer can recompute it on demand. It sits next to the agent’s level in the control plane.
The score informs people. It never moves a level on its own.
Step 3: a nightly check on four signals
Every night, Tuvel looks at each agent’s last 14 days through four signals that people and outcomes produce:
- Approval cards. More than 30% of its cards dismissed or edited, with at least 10 decided.
- Errors. More than 15% of its runs ending in an error the agent caused, with at least 20 runs.
- Ratings. Its work rated bad or wrong by 3 or more different people.
- Outcomes. 2 or more outcomes rated poor.
Any one of these drops the agent one level. It drops at most one level a night and never below L1. An error a model provider caused, such as an outage, is not counted against the agent, and the check never reads anything it wrote itself. After a drop, the window starts again, so the same cards cannot count twice.
Step 4: moving up, level by level
The same signals work in reverse. After 30 days at its level, with 25 or more decided cards, under 5% of them dismissed, no level drop and a results score of at least 60, the agent is nominated for the next level on a card. A person decides. The step is on the record, with who made it and why.
About one run in twenty is also sent back to a person to grade, so the record keeps a human check even when nothing goes wrong.
Going one level at a time keeps each decision small. Moving a credit appraisal agent from drafting notes to updating the file is a different question from letting it email a borrower, and each one gets its own review.
Step 5: some breaches drop a level at once
Autonomy that can only grow is a ratchet. Earned autonomy also works in the other direction, and some breaches do not wait for the night. Going over its budget stops an agent and drops it one level straight away. Asking for an action reserved for people, outside an approval card, drops it to L1 and pauses it. Any agent can also be paused in one click by the bank’s team: its next action is refused until someone resumes it.
Step 6: what never moves
Some actions stay with people at every level. Irreversible actions, such as deleting records or changing governance, always wait for a person. On bank desks, sanctions, payments, filings and anything a customer sees are a person’s decision, approved with one button on a card in the desk’s room.
For customers: two keys
Customer-facing agents need one more control. In Tuvel Advisor, the customer holds one key: the level of freedom they give the advisor. The bank holds the other: the highest level each customer segment may use, and caps per payment and per day. An action runs on its own only where both keys allow it, and commitments such as credit applications and payments to people always wait for the customer’s tap. We explain this in two keys for an AI money advisor.
What your risk team sees
Each agent has a panel in the control plane with its level, results score, budget, model, granted tools and named approvers. The live operations view shows what every agent is doing now and what it just decided, and lets the team pause, demote or override any agent. Every turn, tool call and decision is on the record, so a reviewer can see why an agent moved up, moved down or was stopped.
A checklist for earned autonomy
Put these questions to any agentic AI platform, and ask to see each answer working:
- Does each agent hold its own level, and what waits above it?
- Is there a score built from the agent’s own results, and how many results does it need before it shows?
- Which signals move a level, how often are they checked, and does the score alone ever move one?
- Who promotes an agent, and is the promotion on the record?
- What happens when an agent’s work slips, and how fast?
- Can one agent be paused or demoted without stopping the others?
- Which actions wait for a person at every level?
- For customer agents, who sets the limits: the customer, the bank, or both?
Earned autonomy is one of three ideas that make Tuvel different, with Multiplayer AI and the Control Plane. See what makes Tuvel different, and how banks govern AI agents more broadly.