Lesson objective

Distinguish software that suggests from software that acts, and locate exactly where in the tool-call loop a law firm decides what an agent can cause.

Plain-language explanation

Why this matters now

The earlier levels of this programme mostly dealt with software that produces words for a person to read. A research assistant drafts a memo, a summariser condenses a deposition, a drafting tool proposes a clause. A person reads the output and decides what happens next. The risk lived in the words, and the control was reading them carefully.

This level deals with software that does things. It files a document into the firm's document management system, creates a calendar entry, sends an acknowledgement to a prospective client, opens a matter, updates a task, or posts a note to a matter record. The risk is no longer only in what the output says. It is in what has already happened by the time anyone reads it.

Vendor documentation: the mechanism is worth understanding precisely, because every control in this level attaches to it. Anthropic's documentation on tool use with Claude describes a round trip: the model returns a structured request to call a named tool with particular inputs, and the host application executes that call. OpenAI's function calling guide describes the same loop from a second vendor: the model returns a call, the application runs it and sends the output back to the model. The two vendors agree on the point that matters most for a law firm. The model never runs anything itself. The application does. Vendor documentation changes, so read the current pages before relying on any detail.

Open standard documentation: the Model Context Protocol project describes MCP as an open-source standard for connecting AI applications to external systems, "like a USB-C port for AI applications", through which an application can reach data sources, tools and workflows. The documentation is a living document; the version relied on here is dated 28 July 2026, as published on 14 September 2026. What it means in practice is that connecting an assistant to another system is increasingly a configuration step rather than an engineering project. That is convenient, and it is exactly why a firm has to decide an agent's authority before anyone connects anything.

A question to sit with

If the model never executes anything, and the application executes only what it has been set up to execute, then somebody in your firm has already decided what this agent can cause. Who was it, and where is that decision written down?

What goes wrong

1. "It is only a chatbot." A team describes an automation by its interface rather than its reach. The chat window looks harmless. Behind it sit connectors that can write to the practice management system and send email from a shared mailbox. Nobody reviewing the project asked what the application would execute, because nobody thought of it as software that acts.

2. Reach is granted by connection, not by decision. A connector is added so an assistant can look up matter numbers. The connector also exposes creating, updating and closing matters, because that is what the target system offers. Nobody chose to give the assistant those operations; they arrived with the connection.

3. The suggestion quietly becomes the action. A workflow drafts a reply to a client for a person to approve. Months later, to save time, someone changes it to send automatically when the draft "looks routine". The software is identical. Its authority has changed completely, and no one recorded the change as a decision.

4. The boundary is described, not enforced. A project document says the agent "only drafts". The application nonetheless holds a send permission, and the only thing standing between a draft and a sent message is an instruction in a prompt. An instruction to a model is not a control; the permission the application holds is.

5. The agent borrows a person's reach. The automation runs under a partner's account because that was quickest to set up. It can now do everything that partner can do, in every matter that partner can see. Automation identities and least privilege are the subject of Module 8; the point here is that reach is set by whose credentials the executor holds.

The practical answer

Define an agent by what it can cause, not by which model it uses. The same model can sit behind a harmless drafting aid and behind an automation that sends mail to opposing counsel. Classify the deployment, not the model. This programme uses a six-rung ladder.

Rung What the software does Who sees it before it takes effect Law firm example
0 Answers or suggests The person asking, always Explains a clause on request
1 Drafts for a person to use A person, who decides whether to use it Drafts a status letter into a review queue
2 Proposes a specific action and waits A person, who approves or rejects that action Proposes a calendar entry for approval
3 Acts inside the firm, reversibly Nobody beforehand; anyone can undo it Tags a document with a suggested type
4 Acts inside the firm with lasting effect Nobody beforehand; undoing it takes work or loses data Closes a task, merges two contact records
5 Acts on the outside world Nobody beforehand; it cannot be recalled Sends an email, submits a form, files, pays

Rungs 0 to 2 are tools that suggest. Rungs 3 to 5 are agents that act. The line between rung 2 and rung 3 is the most important line in this level, because above it the firm learns what happened after it happened.

Find the executor. For every automation, list what the application will actually run when the model asks. Do this from the configuration and the connector documentation, not from the project description.

Tool the model can call Operations exposed System touched Read or write Reversible Reaches outside the firm
matter_lookup search, read Practice management Read Not applicable No
matter_update create, update, close Practice management Write Partly No
send_message send Shared mailbox Write No Yes

Any row where "Read or write" says write, "Reversible" says no or partly, or the last column says yes is a row that needs an authority decision in writing, which is the subject of the next lesson.

Separate three questions that are often blurred.

  1. Can the model ask for this action? That depends on which tools are described to it. It is a design fact.
  2. Will the application carry it out? That depends on the permissions the executor holds and the checks the application applies. It is an engineering fact.
  3. Should it happen without a person? That is a professional and management decision. It belongs to the firm, and a licensed attorney owns it for anything touching client work.

A deployment is under control only when the answer to the second question never exceeds the answer to the third. If the application would execute something the firm has not decided it should, the prompt is the only barrier, and a prompt is not a barrier.

Remember this

  • The model returns a request; the application executes it. The firm's decision about authority lives in what the application is permitted to run.
  • Classify a deployment by what it can cause, using the ladder from suggesting to acting on the outside world; the line between proposing and acting is the one that matters most.
  • A connector brings every operation the target system exposes unless someone restricts it, so reach arrives by connection unless it is decided.
  • An instruction in a prompt is not a control. If the application holds a permission, assume it can be used.
  • Before connecting anything, answer three separate questions: can the model ask, will the application execute, and should it happen without a person.

Legal example

The fictional firm Pellbrook & Stone LLP deploys an assistant for its employment group. The project brief describes it as a drafting helper: associates ask it to prepare first drafts of routine letters, which they review before sending. The firm's new automation lead, Tamsin Oyelaran, is asked to sign off the deployment and does something the brief did not anticipate. She ignores the description and opens the configuration. The assistant has four tools. Two are read-only: a matter search and a document search in the firm's DMS. The third, "save_draft", writes a draft into the matter workspace, which she places on rung 3 because a draft can be deleted without loss. The fourth is "send_letter". It was added during a demonstration so the vendor could show a letter going out end to end, and nobody removed it. It sends from the group's shared mailbox without any approval step. Nothing in the prompt tells the model to use it, and the associates have never seen it used. That is not the point. The application would execute it if the model asked. Tamsin places "send_letter" on rung 5, notes that no authority decision covers it, and refuses sign-off until it is either removed or replaced by a rung 2 tool that places a proposed message into an approval queue. The team removes it. She then checks the matter search connector and finds it also exposes an "archive matter" operation, which arrived with the connection. That is restricted as well. Her sign-off memo records the four questions she asked of each tool: the operations exposed, the system touched, whether it writes, and whether its effect reaches outside the firm. The memo is two pages long and becomes the firm's template for every later deployment.

Responsible-use warning

Never accept a description of an automation in place of its configuration. "It only drafts" is a statement about intent; the permissions held by the application are a statement about capability, and only capability matters when something goes wrong. Equally, never treat an instruction in a system prompt as the thing that keeps an agent below the line between proposing and acting. If the application would execute a send, file or update on request, assume that one day it will be asked to, whether through error, a changed prompt or hostile input.

Human-review requirement

A licensed attorney who owns the practice area signs off the rung assigned to every tool that touches client matters, because placing a tool above the line between proposing and acting is a decision to let work happen before a lawyer sees it. The reviewer confirms that the tool list was taken from the configuration rather than the project description, that every write operation and every operation reaching outside the firm has a recorded authority decision, that nothing unused was left connected, and that the firm's own jurisdiction's rules on supervision and competence were read rather than assumed.

Practical exercise

Place a fictional firm's automation on the agency ladder.

  1. Invent a firm and one automation in a practice group of your choice. Write the one-sentence description a project team would give it.
  2. List at least five tools that automation can call. For each, write the operations the connector exposes, including any that came with the connection but were not requested.
  3. For each tool, record the system touched, whether it reads or writes, whether its effect is reversible, and whether it reaches outside the firm.
  4. Assign each tool a rung from 0 to 5 and write one sentence justifying the rung.
  5. Mark every tool on rung 3 or above. For each, write whether an authority decision currently exists and who took it.
  6. Find the tool whose rung is furthest from the project team's description. Write the sentence you would put in a sign-off memo about it.
  7. For each tool, answer the three questions separately: can the model ask, will the application execute, and should it happen without a person.
  8. Identify any case where the answer to the second question exceeds the answer to the third, and write the change that closes the gap without relying on the prompt.
  9. Write two sentences for the sign-off memo stating which tools you would remove before go-live and who must approve any tool you keep on rung 3 or above.

Knowledge check

  1. In the tool-call loop described by both Anthropic and OpenAI documentation, what does the model do and what does the application do?

    Answer guidance: The model returns a structured request to call a tool with particular inputs. The host application decides whether to execute that call, runs it, and sends the result back. The model never runs anything itself, so an agent's reach is whatever the application is permitted to execute.

  2. Why is the line between rung 2 and rung 3 of the ladder the most important one?

    Answer guidance: Below it a person sees and approves a specific action before it takes effect. Above it the software acts first and the firm learns what happened afterwards, so control shifts from prevention to detection and repair.

  3. Why is an instruction in a prompt not an adequate way to keep an agent from sending email?

    Answer guidance: Because the instruction only shapes what the model is likely to request. If the application still holds a send permission, a changed prompt, a model error or hostile input can produce a send request that the application will execute. The control is removing or gating the permission.

  4. What are the three questions to separate before connecting a tool?

    Answer guidance: Can the model ask for this action, which is a design fact; will the application carry it out, which is an engineering fact; and should it happen without a person, which is a professional and management decision owned by the firm. The second answer must never exceed the third.

Sources and further reading