Study 01 · Framework

As AI becomes more capable, the interface should not simply accumulate more safeguards. The right safeguards change with the authority a person has actually granted.

Study 01 follows one relationship between a person and an agent, from approving every invoice to taking authority back. The framework asks how that lifecycle might carry to other actions an agent takes on someone’s behalf.

The model

  1. Capability, consequence and accumulated evidence shape what the agent may propose.Stage 2
  2. The person’s choice creates granted authority.Stage 3
  3. Granted authority and the situation decide what the agent may do.Stage 4

The lifecycle

The five stages of Study 01, and where each safeguard appears, changes or goes. Markers use the study’s grammar: dotted for what the agent proposes, yellow for what the person authors.

  1. Evidence builds

    Moment 1Moment 2

    The agent checks every invoice and Sam approves each one. Its record grows, while repeated approval becomes harder to justify.

    • IntroducedApproval of every invoice
  2. The agent proposes

    Moment 3

    It cites its record and asks to stop asking for routine invoices. It recommends a limit. Nothing changes until Sam decides.

    • IntroducedA preview of what the authority would have done
  3. The person grants

    Moment 3

    Sam sets the terms in one readable sentence: who, how much, when it pays, and when it pauses for review.

    • Deliberately retiredApproval of each routine invoice
  4. The agent acts

    Moment 4Moment 5

    Routine invoices become receipts. Rules decide what may be paid; judgment adds what it noticed. Routine receipts recede; exceptions expand for review.

    • AutomatedRoutine payments, each with a receipt
    • Escalated only in exceptionsA changed account, held for Sam
  5. Authority changes or returns

    The ending

    Tightening applies at once. Loosening asks first and names what changes. Turning off is one action with one confirmation, and authority pauses at its review date.

    • StrengthenedA lower limit, at once
    • ModifiedA higher limit, after naming what newly pays

Principles

  1. The agent proposes. Only the person grants.

    Moment 3
  2. Authority reads as one sentence, built from fields a person can check.

    Moment 3
  3. Rules decide what may happen. AI judgment adds what it noticed.

    Moment 5
  4. The model observes. The product writes the explanation and the next step.

    The evidence
  5. Reducing authority should require less friction than expanding it.

    The ending
  6. Routine recedes. Exceptions expand.

    Moments 4 and 5
  7. Authority never renews silently. It pauses at its review date.

    Moment 3

Beyond payments

The same lifecycle, sketched for other actions an agent might take. Sketches, not studies.

ActionThe agent might propose, once its record supports itExample hard boundary
Replying to messagesAnswer routine scheduling requests from known contacts without askingNever to a new recipient, and never with an attachment
Changing bookingsRebook within a time window and a price bandNothing non-refundable
Deleting informationClear duplicates and expired draftsNothing that can’t be recovered