What is a trust ladder for AI agents?
A trust ladder is the rule set that decides what an AI agent may do without a human tap, per kind of work, earned by a streak of clean approvals and lost on a rejection.
By Max Gillespie, founder of Oetzi
A trust ladder is the rule set that decides what an AI agent may do without a human, per kind of work. The bottom rung asks before everything. The middle rung does the work after a window unless told no. The top rung does it and reports afterward. An agent climbs a rung by earning a streak of approvals the person did not edit, and drops a rung when a piece of work is rejected with a reason. The point is the same as with a new colleague: approval workflows are how trust starts, not how it should end.
Why not just a setting?
A single "autopilot on or off" switch fails in both directions. Off, the person is a bottleneck for everything, including the things the agent has done well two hundred times. On, one bad post goes out under the company's name. A ladder is per kind of work, so LinkedIn posts can be on the top rung while replies to strangers still wait for a tap, and it moves in small steps with a reason attached to every move.
How does an agent earn a rung?
By a streak, and by asking. In Oetzi, eight untouched approvals of posts on a platform, or ten sent replies, or six accepted video ideas, earn a proposal: one text that says what it would like to do on its own and what "yes" changes. The agent never promotes itself. A rejection with a reason drops a rung and the reason becomes a standing lesson, and "ask me again" in any channel brings every tap back at once. The rungs and streaks are visible under Settings, so the arrangement is never a surprise (how Oetzi communicates).
What stays on the bottom rung forever?
Money. Anything that spends beyond the plan's included usage (what Oetzi charges) asks every time, whatever the streak, because a co-worker with a company card still files an expense. The ladder is about judgment on work the person has already paid for, not about spending more.
Questions
Is this the same as human-in-the-loop?
It is human-in-the-loop with a dial that the agent has to earn its way along. Most agent frameworks, including the OpenAI agents guidance on approvals, treat approval as a fixed gate. A ladder treats it as a starting position.
What happens to the middle rung in practice?
It is the workhorse. "Posting at 5pm unless you say no" gives a person one glance and a one-word veto, which is how most work actually moves between colleagues.
Oetzi is an AI growth hacker for founders. It finds the threads where people ask for a product like yours, checks whether the AI engines recommend you, and drafts replies, posts, and short video in your voice. You approve everything. 100 credits included, 10 free to try with no card, 50 bonus credits with the first payment.
- GEO vs AEO vs SEO: what is the difference?SEO earns a rank in search results, AEO earns the answer box, GEO earns a place in an AI engine's written answer. How the three differ and where a founder should spend effort.
- Share of model vs share of voiceShare of voice counts how much of a category's media mentions you. Share of model counts how often AI engines name you. How the two differ and which one to track.
- What is a "what do you use" thread?A what do you use thread asks a community which tool to pick. How these threads work, why the first hour decides the shortlist, and how to answer one well.
- What is a buying intent signal?A buying intent signal is a public post showing someone is actively choosing what to buy. What they look like, where they appear, and how to answer one well.
- What is a keyword alert?A keyword alert is a standing search that pings you when a word appears in a public post. What alerts do well, why people mute them, and how to make them work.