The Spectrum of Agency

The Wrong Question: How Autonomous Should It Be?

In early 2024, two hardware startups launched within months of each other. Both promised the same vision: an AI-powered personal assistant that understands you, anticipates your needs, and handles tasks on your behalf. Both raised hundreds of millions of dollars. Both bet on maximum autonomy as the selling point.

Rabbit R1 shipped a dedicated device with a "Large Action Model" that would navigate apps and complete tasks independently. Humane AI Pin shipped a screenless wearable that projected information onto your palm. Both flopped spectacularly. The deeper failure was a product design error: they asked users to adopt a new form factor AND trust maximum AI autonomy simultaneously. Two trust leaps at once. Neither had been earned.

Meanwhile, Apple shipped AirPods with incremental AI capabilities. Not an agent. Just progressively smarter features, each earned through reliability at the previous level. Users barely noticed the autonomy gradient increasing.

The lesson isn't that autonomous AI doesn't work. It's that autonomy is a spectrum, not a switch -- and the PM's job is to find the optimal position on that spectrum, not the maximum one.

Five Rungs from Suggest-Only to Self-Directed

The spectrum has seven levels, each defined by a structural question: who decides what happens next, and how much human involvement does that decision require?

The dividing line falls between Level 4 and Level 5. Below it, a human is in the decision loop for every consequential action. Above it, the model makes decisions and the human's role shifts from approving to monitoring. That transition -- human-in-the-loop to human-on-the-loop -- is where complexity, risk, and potential value all inflect sharply upward.

The same product can operate at different levels simultaneously. GitHub Copilot operates at Level 1 (inline autocomplete), Level 3 (chat), and Level 6 (Copilot Workspace). The PM's question isn't "what level is our product?" but "what level is each capability within our product -- and is that the right level?"

Choosing the Right Rung for the Right Workflow

The economics shift at each level. Levels 1-3 have predictable, low token costs. Level 5 is the inflection point: 5-15 model calls per interaction, variable costs, and errors that propagate before anyone catches them. Levels 6-7 are genuinely unpredictable without careful engineering.

Regulated industries face a hard ceiling. Healthcare, financial services, and government face mandatory human review for consequential decisions. The spectrum isn't one level per product -- it's one level per decision type.

Remember This

1. The consequential boundary is between Level 4 (copilot) and Level 5 (agent). That single-level transition changes your testing, liability, cost structure, and governance requirements simultaneously.

2. One product can and should operate at multiple levels simultaneously. The PM's job is calibrating autonomy per capability, per user segment, per risk profile.

3. Trust is earned by proving reliability at the current level before increasing autonomy. The product that starts at Level 5 fights human psychology. The product that starts at Level 1 and earns Level 5 uses it.