AI, and what it will never do
A permission-bound colleague. Not an oracle, and not an autopilot.
What the assistant does today
Five behaviours, and this is all of them. Each one runs on request, from a screen you were already on, and none of them runs on a schedule or in the background.
- Draft with AI
- Drafts a work-item description or a comment from a short instruction. What comes back is text in the editor, which you keep, change or discard.
- Draft Article with AI
- The same thing for a Library article, including restructuring one you have already started.
- Timesheet Assistant
- Answers questions about your own time and finds the gaps in it before somebody else does.
- Week Review
- Summarises a week of time for the person who logged it, ahead of submitting.
- Approval Brief
- Prepares an approver's summary of what is waiting, so a queue is read rather than clicked through.
The assistant is off in every new organization and stays off until an owner turns it on. Each run records what it cost, against a budget the organization sets, and the whole module can be switched back off in one action without losing anything a person kept.
Three commitments
Optional
AI is an organization-level opt-in and it defaults to off. With it off, every feature works identically. No workflow depends on a language model being reachable.
Bounded
An agent never receives a database credential. It acts only through registered tools, and each tool calls the same domain service a human request calls, with the same authorization check.
Accountable
Every run records its prompt version, model version, tool calls with arguments and results, and its outcome. Every mutating action produces an audit event attributed to both the agent and the person on whose authority it ran.
An agent's reach is an intersection
Not a role it is granted, and not a fence it operates inside. What an agent can reach is the overlap of three sets: what the invoking person may see, what the agent definition declares it needs, and what the organization's policy allows.
So an agent is never more privileged than the person who asked it for something, and the check runs on every tool call rather than once at the start.
What agents do, in order of trust
Capability is earned in that order. Only the last group is consequential, and that is the group which stays human-approved.
| Capability | Examples | Consequential? |
|---|---|---|
| Retrieval | Search employees, read a profile, find missing timesheets, read team utilization. | No |
| Analysis | Summarize a week of time, flag entries worth a human look, prepare a utilization report. | No |
| Drafting | Draft timesheet reminders. Drafts persist as proposals and send nothing. | No |
| Approved action | Send the reminders, apply a proposed correction, approve a timesheet. | Yes, always human-approved |
What agents will not do
- Run arbitrary SQL, or receive a database credential of any kind.
- See data the person who invoked it could not see through the API.
- Take a consequential action without a recorded human approval.
- Be the only path to any product outcome.
A further category is reserved from agents permanently, whatever permissions are granted. Approving a timesheet automatically, scoring an employee, and sending a message on someone's behalf all sit in it.