Bounded Autonomy Trust Ladder
Increase agent autonomy only after each bounded batch earns trust
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 7
- Confidence
- 98%
Treat autonomy as something an AI employee earns through evidence. Begin in draft mode, where the agent prepares targets, messages, or work products for human review without taking external action. Once quality is acceptable, authorize a bounded live batch under the agent’s own clearly identifiable account. At the end of that batch, inspect outcome metrics and qualitative failures, then ask the agent to propose improvements. A human approves the next experiment, increases the batch size, maintains the current level, or returns the agent to draft mode. Each stage limits downside while producing the information needed for the next delegation decision. The approach builds calibrated trust through checkpoints rather than relying on either fear or enthusiasm.
Origin
Extracted from Marketing Against The Grain while the hosts discussed safely launching the demonstrated email agent.
Core principles
- 01Trust should follow observed performance
- 02Draft mode separates capability testing from external action
- 03Bounded batches limit downside and generate evidence
- 04Dedicated identities make responsibility visible
- 05Metrics and feedback determine the next autonomy level
How to run it
- 1
Start in Draft Mode
Ask the agent to prepare work without sending, publishing, or otherwise acting externally.
Pro tip Require a review document that combines proposed targets and outputs.
Watch out Do not confuse fluent drafts with proven operational reliability.
- 2
Review Representative Samples
Inspect both routine and edge-case outputs, including who or what the agent selected for action.
Pro tip Review targeting quality separately from content quality.
Watch out A polished message can still be harmful when sent to the wrong recipient.
- 3
Authorize a Bounded Batch
Allow a fixed number of external actions or a fixed operating period, then require a checkpoint.
Pro tip Choose a batch large enough to produce useful evidence but small enough to contain failure.
Watch out Avoid indefinite authorization during the first live run.
- 4
Use a Distinct Identity
Have the agent act through its own inbox or account so recipients and operators can distinguish its work.
Pro tip Make the handoff path to a human obvious.
Watch out Do not conceal autonomous authorship behind a human identity.
- 5
Review Performance
Measure outcomes after the bounded batch and inspect poor-quality cases, replies, and operational errors.
Pro tip Include both conversion metrics and qualitative review.
Watch out A single aggregate metric can hide serious edge cases.
- 6
Apply Feedback to the Next Batch
Ask the agent to diagnose performance, approve appropriate changes, and test them on another bounded batch.
Pro tip Change a limited number of variables so improvement can be attributed.
Watch out Do not let the agent silently rewrite its own boundaries.
- 7
Adjust Autonomy
Expand, maintain, reduce, or pause autonomy based on the accumulated evidence.
Pro tip Document why the autonomy level changed.
Watch out Trust should not increase merely because no one noticed a failure.
In the wild
A re-engagement agent first creates customer targets and email drafts for approval. After its work is sampled, it sends a fixed initial batch from its own inbox. It then reports performance, suggests improvements, and waits for approval before applying those changes to the next group.
→ The business gains automation while preserving explicit quality and risk checkpoints.
Common mistakes
Jumping Straight to Full Autonomy
Unlimited external action provides no safe checkpoint for detecting weak targeting, poor communication, or tool failures.
Reviewing Copy but Not Selection
Quality control must examine whom the agent chose to contact as well as what it drafted.
Expanding on Anecdotal Success
A few good examples are not enough evidence to remove meaningful oversight.
Is it for you?
Best for
External-facing workflows where mistakes carry reputational or commercial cost but automation has substantial upside.
Not ideal for
Irreversible or safety-critical actions whose downside cannot be contained by batching and review.
From the transcript
“it can do the draft, it can set up all the drafts, and you can send the emails until you get comfortable.”
“hey, email the first thousand, and then we'll review the results after a thousand cents to like get going and check in how it's performing.”
“if on the off chance it did go a little dry, because I think that's what people worry about, it it basically said something, you…”
From the episode
Build An Army Of AI Employees To Run Your Business (With Hermes AI)