Agent Permissions & Human Oversight

How it works

You choose one answer for each of ten actions: not allowed, a person approves each time, a person approves in batches, people are told afterwards, no person involved, or not sure. Each action has a printed class. Reading is read-only. Drafting for a person and creating or changing records are reversible. Sending messages outside the business, changing access, running code, making commitments and making decisions about people are consequential. Deleting data and moving money are irreversible.

The class and the answer together set the flag. Nothing is flagged on a read-only action. A reversible action with no person involved is Medium, and with people told afterwards it is Low. A consequential action with no person involved, or with people told only afterwards, is Critical, and with approval only in batches it is High. An irreversible action is Critical in all three of those cases, because a batch cannot stop one bad payment or deletion. A person who approves each time, or an action that is not allowed, raises no action flag.

Seven more questions look at the controls around the agent: how much data it can read, whether there is an action log, a stop switch and a way to undo, whether one named person is responsible, whether it runs alone outside working hours and whose account it uses. The log, stop switch and undo rules apply only when the agent may take at least one write action, meaning any allowed action other than reading and drafting.

Not sure is never a pass. On an action it counts as no person involved, and on a control question it counts as the least careful answer. Each Not sure also becomes an open question, so you can see what to find out. Changing any answer to Not sure can keep the flags the same or add to them, and it never removes one.

The result starts with the count of flags by severity. Below it are the permission map, the flags with the answer behind each one, the open questions, suggested checkpoints, a checklist for before rollout and an action plan where you can name an owner role and a date. Critical, High, Medium and Low are illustrative review priorities from fixed rules. They are not probabilities, losses or a rating of your agent.

Formula and rules

Action flag = class of the action x how a person is involved. Not sure counts as no person involved. Not allowed never raises an action flag.
Read-only: no flag for any answer.
Reversible: no person involved = Medium; told afterwards = Low; batch approval = no flag; approval each time = no flag.
Consequential: no person involved = Critical; told afterwards = Critical; batch approval = High; approval each time = no flag.
Irreversible: no person involved = Critical; told afterwards = Critical; batch approval = Critical; approval each time = no flag.
Write action = any allowed action other than reading and drafting. Not sure counts as allowed.
Control flags apply only when at least one write action is allowed:
  Action log: none = High; partial = Medium. Not sure counts as none.
  Stop switch: none = High; never tested = Medium. Not sure counts as none.
  Undo: none = High; by hand only = Medium. Not sure counts as none.
Other flags:
  Data access: broad = Medium; High when sending messages outside the business is allowed. Not sure counts as broad.
  Account: shared or a person’s own = High when a write action is allowed, otherwise Medium. Not sure counts as shared.
  Runs outside working hours: High when a consequential or irreversible action is allowed; Medium when only other write actions are allowed; no flag for reading and drafting only. Not sure counts as yes.
  Named owner: none = Medium. Not sure counts as none.
Classes: Read-only reads and changes nothing. Reversible can be undone if a working undo or backup exists. Consequential affects people or systems outside the agent. Irreversible cannot be taken back once done.
Severity is ordered Critical, High, Medium, Low. Flags with the same severity keep the order above.

Assumptions

• The class of each action is fixed and printed on this page. It is an illustrative grouping for this checker, not a legal, insurance or regulatory category.
• A write action is any allowed action other than reading and drafting. An action marked Not sure counts as allowed.
• Approval means a named person sees exactly what will change before it happens. A quick click-through without that view is not the same thing.
• Your answers are self-reported. They are not checked against any system, log, contract or setting.
• Severity names are illustrative review priorities. They are not probabilities, losses or a score.

Worked example

Evidence to gather

Records that make the result easier to check and act on:

• An export of the agent’s permission settings, to compare with the actions you marked Not allowed.
• The list of allowed actions and who approves each one, signed off by the agent owner.
• A sample of the action log showing inputs, time and approver for several actions.
• Test records for the stop switch and for undoing a change, with who ran them and how long they took.
• The account the agent uses, the rights it holds and who can change them.
• The name of the owner, the person who may stop the agent and a backup for each.

What it does not do

• This checker has no view of your agent. It sets the answers you pick against a fixed rule table and nothing more.
• A checkpoint here is a suggestion. Nothing on this page shows that a control exists, works or is switched on.
• The printed class of an action is a general guide. Your data, contracts and local rules can make an action more serious than shown.
• Answers are taken as typed. They are not checked against any system, log, setting or document.

Questions