Lesson objective

Write the authority statement that opens every automation design, stating what the agent may do alone, what it may only propose, and what it may never touch.

Plain-language explanation

Why this matters now

Once a firm accepts that the application, not the model, decides what happens, the next question is where that decision is recorded. In most firms the honest answer is that it is not. It is spread across a vendor's setup screens, a project ticket, a prompt someone edited last month and the memory of the person who built the integration.

That works until the first time somebody asks. A client asks whether a machine sent the letter. An insurer asks what the firm's automations are permitted to do. A partner asks why a matter was closed on a Saturday. A new automation lead inherits a dozen workflows and has to decide which of them to trust. Each of those questions needs a single document that states the authority, written before the automation ran and kept up to date since.

Security standard: the OWASP GenAI Security Project published the OWASP Top 10 for Agentic Applications for 2026 on 9 December 2025. Its entries include tool misuse and identity and privilege abuse, both of which are failures of authority rather than failures of model quality. The OWASP page itself shows only the cover and a download link, and the entry list relied on here was confirmed from a secondary summary, so read the PDF before quoting it. OWASP material is a voluntary standard, not law, and following it neither creates nor discharges a professional duty. What it contributes is a vocabulary for asking, of every agent, what it is permitted to do and whether its actions require confirmation.

A question to sit with

If your automation did something unexpected tomorrow, could you produce a document, dated before it happened, showing whether that action was within its authority? Or would you be reconstructing its authority from its behaviour?

What goes wrong

1. Authority is implied by capability. Nobody wrote down what the agent may do, so the de facto answer is whatever it can do. Every expansion of capability silently becomes an expansion of authority.

2. Authority is stated by purpose rather than by action. "The agent helps with intake" tells nobody whether it may send a message, create a contact, schedule a call or decline an inquiry. Purposes do not constrain; lists of actions do.

3. The never-list is missing. Teams write down what an agent should do and forget to write down what it must never do. Without a stated prohibition, a later change that adds a forbidden action looks like an ordinary improvement.

4. Scope is unbounded. The agent may "update matters". Which matters? Which fields? At what volume? A statement without scope authorises the agent to update every record in the system at once.

5. Nobody owns it. The statement exists but has no named owner and no review date. It was accurate on the day it was written and has drifted ever since.

6. The statement and the configuration disagree. The document says propose-only; the application can act. When they disagree, the configuration is what happens, and the document becomes evidence that the firm knew better.

The practical answer

Every automation design opens with an authority statement. It is short, it comes first, and a design that lacks it has not been designed. It has three lists and a set of attributes for every action in them.

List Meaning Default control
May do alone The agent performs the action with no person beforehand Logged; reviewed after the fact; reversible wherever possible
May propose The agent prepares the action and a named person approves or rejects it Approval gate with the approver identified by role
May never The agent may not perform or propose the action, and the application must be unable to execute it Permission removed at the executor, not merely discouraged in the prompt

Anything not on one of the three lists is on the never-list by default. That single sentence, written into the statement, is what stops capability from expanding into authority unnoticed.

Describe each permitted action with the same attributes.

Attribute The question it answers Example entry
Action What exactly happens, as a verb and an object Create a draft calendar entry
Trigger What starts it A new document classified as a court order arrives in the matter workspace
Scope Which records, fields and matters, and how many at once Open litigation matters in one office; one entry per trigger
Data What information it reads and writes Reads the order text; writes date, description and source link
Systems Where the effect lands Practice management calendar, draft layer only
Reversibility How it is undone and what is lost Draft deleted without loss
Approver Who approves, by role, for "may propose" actions Docketing clerk, then responsible attorney
Stop condition When the agent must halt and hand to a person Date unreadable, more than one possible trigger date, or matter not found
Evidence What record shows it happened Action log entry with source document identifier

Add volume and rate limits. An action that is harmless once can be catastrophic a thousand times. State a maximum per run and per day, and require that the automation stops, rather than continues, when a limit is reached.

Tie the statement to the configuration. For every "may do alone" and "may propose" action, name the tool and the permission that implements it. For every "may never" item, name how the application is prevented from executing it. A reviewer should be able to lay the statement beside the configuration and match every line.

Give it an owner, a version and a review trigger. The owner is a named person accountable for the automation. The version changes whenever a list changes. The review is triggered by any new tool, any new connector, any change of model or vendor, and at least once a year.

Write it in plain language. A partner, an insurer or a client should be able to read the three lists without a technical glossary. The attribute table can be technical; the lists should not be.

Remember this

  • Every automation design opens with an authority statement of three lists: may do alone, may propose, may never. Anything unlisted is prohibited by default.
  • Describe each permitted action by trigger, scope, data, systems, reversibility, approver, stop condition and evidence, and set volume limits that stop the run.
  • A "may never" item must be enforced at the executor by removing the permission, not by discouraging it in the prompt.
  • Tie every line to the tool and permission that implements it, so the statement and the configuration can be checked against each other.
  • Give the statement a named owner, a version and review triggers, and have a licensed attorney sign it off for the practice area.

Legal example

The fictional firm Pellbrook & Stone LLP has run a document-classification automation in its litigation group for eight months. It began as a tagging aid. When the new automation lead, Tamsin Oyelaran, asks for its authority statement, there is none, so she writes one by observation and finds the problem immediately. The automation now does four things: it tags incoming documents with a suggested type, renames files to the firm's naming convention, moves documents it classifies as correspondence into a correspondence folder, and, since a change two months earlier, creates a task for the responsible associate when it sees what it believes is a court order. Nobody decided that the automation should move documents or create tasks. Each was added as an improvement. She drafts the statement properly. May do alone: tag a document with a suggested type, reversible without loss, in open litigation matters only, with a limit of five hundred documents per day after which the run stops. May propose: renaming and moving documents, approved by the matter paralegal; creating a task for a suspected court order, approved by the docketing clerk. May never: delete a document, change a document's security settings, send anything outside the firm, create a calendar entry, or act in a closed matter. She adds the sentence that anything not listed is prohibited. Then she lays the statement beside the configuration and finds that the automation can still delete documents, a permission inherited from the service account it was installed under. That permission is removed the same day. The responsible litigation partner, Hollis Marchetti, signs the statement off as version 1.0, with a review due in twelve months or on any new tool or connector. Three months later a vendor update adds an option to file documents directly into the court e-filing queue. Because that action is not on any list, it is prohibited by default, and it stays switched off without anyone having to argue the point.

Responsible-use warning

Never write an authority statement that describes a purpose instead of listing actions. "Assists with document handling" authorises everything and constrains nothing, and every later change will look like it fits. Equally, never let a statement say "may never" about an action the application can still perform. When the document and the configuration disagree, the configuration is what happens, and the document becomes evidence that the firm knew the risk and did not close it.

Human-review requirement

A licensed attorney for the practice area signs off the authority statement before the automation touches a client matter, and signs off each new version. The reviewer confirms that the three lists are stated as actions rather than purposes, that the statement says anything unlisted is prohibited, that every "may propose" action names an approver by role, that stop conditions and volume limits are stated, that every "may never" item is enforced at the executor, and that the statement has been checked line by line against the configuration rather than accepted from the project team.

Practical exercise

Write an authority statement for a fictional automation.

  1. Invent a firm and an automation that handles inbound correspondence in a practice area of your choice.
  2. Write the one-sentence purpose. Then set it aside: it will not appear in the authority lists.
  3. List at least three actions for "may do alone", three for "may propose" and five for "may never".
  4. Add the sentence stating that anything not listed is prohibited by default.
  5. For each permitted action, complete the attributes: action, trigger, scope, data, systems, reversibility, approver where relevant, stop condition and evidence.
  6. Set a volume limit for each "may do alone" action and state what the automation does when the limit is reached.
  7. For each "may never" item, write how the application is prevented from executing it. If your answer is the prompt, rewrite it.
  8. Name an owner by role, a version number, and at least three review triggers.
  9. Imagine a vendor update adds a new capability. Write the paragraph explaining why it is prohibited until the statement changes, and who must approve that change.
  10. Lay your statement beside an imagined configuration and write down one line where they might disagree, and how you would detect it.

Knowledge check

  1. What are the three lists in an authority statement, and what is the default for anything not on them?

    Answer guidance: May do alone, may propose, and may never. Anything not listed is prohibited by default, which prevents new capabilities from becoming new authority unnoticed.

  2. Why is a statement of purpose not enough to bound an agent?

    Answer guidance: A purpose such as "helps with intake" does not say whether the agent may send, create, schedule or decline anything, so every action can be argued to fit it. Only a list of specific actions with scope constrains what the agent may do.

  3. How must a "may never" item be enforced?

    Answer guidance: At the executor, by ensuring the application does not hold the permission or cannot perform the operation. Discouraging it in the prompt is not enforcement, because the application would still execute the action if asked.

  4. Why does an authority statement need volume limits?

    Answer guidance: Because an action that is harmless once can cause serious damage when repeated at scale by a fault or a bad trigger. A stated maximum per run and per day, with the automation stopping when it is reached, limits the blast radius.

Sources and further reading