Laurent Rochetta Book a call

The four locks: delegate to AI without losing control

To hand work to AI and sleep soundly, you must be able to bound it, approve what matters, trace everything and stop it. These are my four design criteria. For each one: what it is, what it prevents, how I check it, and its limit.

  1. Scope

    A written list of what the AI may read, write and use: your data, your tools, your messaging channels.

    What it prevents
    It reading another client’s data, or going through a channel you never opened.
    How I check it
    In the CRM I run, the assistant’s rights are set organisation by organisation, messaging is closed by default, and sessions are isolated from one organisation to the next.
    Its limit
    A scope says what the AI can touch, not whether it does it well. That is the next lock’s job.

  2. Approval

    The AI prepares a proposal. A named person approves it. Only then does the action run, within a limited window.

    What it prevents
    An important amount, document or message going out without anyone having seen it.
    How I check it
    In the CRM, nothing is written until the proposal is approved. A proposal that is not applied expires. The approver’s rights are checked again at the moment of action.
    Its limit
    Approving without reading is worthless. So the system has to make reading easy: source cited, differences visible.

  3. Traceability

    Everything the system reads, proposes, has approved and runs is recorded: who, when, from which source.

    What it prevents
    “We don’t know why it did that”.
    How I check it
    An amount is shown only if the figure and its label appear word for word in the cited excerpt. Otherwise, the system declines to conclude.
    Its limit
    A full log is useless if nobody reads it. We decide together who reads it, and how often.

  4. Reversibility

    Being able to switch the system off, undo an action and take over operations without depending on me.

    What it prevents
    Being held hostage by a supplier, a person or a platform.
    How I check it
    Sensitive functions are switched off with a single configuration switch. The assistant’s actions can be undone.
    Its limit
    You can’t take back what has already gone out: a message sent, a payment made. That is why scope and approval protect those actions first.

One example, start to finish

  1. A request comes in: for example, a funding file to prepare.
  2. The assistant reads the documents it is allowed to read and prepares a proposal, with the excerpts that justify it.
  3. A person opens the proposal and sees what changes, and why. They approve or reject it, giving a reason.
  4. The action runs in the agreed window, with the approver’s rights.
  5. It is written to the log: it can be read back and, if needed, undone.

Where this method comes from

It comes from running a system in production, not from a consultancy. My way of working with coding agents is published as open source: BMAD+, MIT licence.

BMAD+ on GitHub

Take your time back

One conversation is enough to know what AI can do in your place. Nothing to prepare: describe a task that takes too much of your time, and I’ll ask the questions.

Book a call