By James Proctor, Co-Founder and Managing Director, The Inteq Group
AI agent autonomy operates across five working levels: recommend, where the agent proposes and a person acts; draft for approval, where the agent prepares the action and a person releases it; act and hold, where the agent completes the action but it takes effect only on human release; act and report, where the agent acts independently and surfaces what it did; and act within limits, where the agent operates silently inside defined boundaries. Every level is legitimate. The design skill is matching each class of decisions in a process to the right level, rather than running the entire process at one level chosen by temperament.
What Distinguishes the Five Levels?
Each level answers two questions differently: when does a human see the action, and can the action take effect without one? Recommend and draft-for-approval put a person before every effect; they differ in how much work the agent has already done. Act-and-hold moves the person to a release gate: the work is complete, the effect waits. Act-and-report removes the gate but preserves visibility, which makes it the workhorse level for routine consequential work. Act-within-limits removes per-action visibility entirely and replaces it with boundary monitoring, which is exactly right for high-volume, low-consequence decisions and reckless anywhere else. Reading the levels as a progression of trust misses the point. They are a menu of control designs, each buying a different mix of speed and oversight.
How Do You Choose a Level for a Decision Class?
Two inputs decide it: the consequence of the decision class being wrong, and the evidence you hold about performance on it. High consequence pins work at the human-gated levels regardless of evidence. Low consequence with demonstrated performance belongs at the autonomous levels, where the value actually accumulates. The honest tension arrives in the middle, moderate consequence with growing evidence, and the discipline that resolves it is explicit progression: define, in advance, what evidence advances a decision class from one level to the next, and who signs the advancement. Field service operations show the full menu in one process. In elevator and HVAC service dispatch, routine preventive maintenance scheduling runs at act-within-limits, parts ordering against contract runs at act-and-report, emergency callout prioritization runs at act-and-hold because a mis-ranked emergency carries real safety exposure, and anything committing the company contractually stays at recommend. One process, one agent platform, four deliberate positions.
“Recommend-only is not a safety posture. It is a deferral with a burn rate.”
Why Do Organizations Get Stuck at the Bottom?
Because the bottom is politically safe, and I will say what the steering decks will not: most pilots park at recommend-only not because analysis put them there but because nobody wants to own the memo that grants an agent authority. Recommend feels prudent. It generates no incidents, no hard conversations about consequence, and no accountability. It also generates a fraction of the value while carrying most of the cost, and the gap never appears on anyone's report, because forgone acceleration has no line item. Prudence is the alibi. Avoidance is the behavior. If a decision class has run at recommend for two quarters with the humans accepting the recommendation unchanged ninety-plus percent of the time, you are not being careful. You are paying people to click approve, and calling the clicking governance.
Designing level assignments and progression criteria, and building the analytical bench that can defend them, is a focus of our agentic AI training courses, because level selection is a skill your teams will exercise on every process they touch.
