The AI Delegation Register

Use this when AI does any recurring part of the work on your project, and before anyone goes on leave, hands over or joins. One entry per task records who does it now, a person, an AI with a person checking, or an AI on its own, plus what a human reviews and where the set-up lives. It makes the project’s AI operating model visible to the team instead of living in one person’s accounts.

The frame

One entry per task, in plain text so it pastes into a shared document, a wiki page or a ticket.

Task: [the recurring piece of work, named the way the team names it]
Done by: [Agent on its own / AI drafts, person approves / Person]
Tool and set-up: [the tool, and where the prompt, skill or automation lives]
Produces: [what comes out and where it goes]
Human review: [who checks what before it has consequences, or "none" with the reason]
Config owner: [the person who can change or switch it off]
Runs alongside: [other agents or steps that share its input or output, or "none"]
Last reviewed: [date]
Review next: [date, or the change that should trigger a review]

Have an AI draft it

Most teams start with more automations running than anyone has written down. This prompt turns a rough list into entries and shows you the gaps.

I'm building an AI delegation register for my project. Below is a rough list of the places AI is used in our work: scheduled agents, saved prompts, skills, automations and anything else.

For each item, write one entry using these fields: Task, Done by, Tool and set-up, Produces, Human review, Config owner, Runs alongside, Last reviewed, Review next.

Rules:
- Done by must be one of: Agent on its own / AI drafts, person approves / Person.
- Where my list doesn't say, write "MISSING:" and the question I need to answer. Do not guess an owner or a review step.
- If two items share an input or feed each other, name each in the other's Runs alongside field.
- After the entries, list any task where Done by is "Agent on its own" and Human review is empty or vague.

My list:
[paste your notes here]

Before you paste details of real systems, people or clients into a third-party AI tool, check your organisation’s policy on client and commercially sensitive material. Role titles in place of names keep the draft useful.

Every “MISSING:” line is a decision only the team can make. Agree those before the register is shared.

Worked examples

The entries below are illustrative, for a twelve-month system replacement programme.

Weekly risk extraction

Task: Weekly risk extraction from project channels
Done by: AI drafts, person approves
Tool and set-up: Scheduled automation in the team's workflow tool, prompt stored in the shared skills folder
Produces: Candidate risk entries in the RAID log's review tab
Human review: PM reads every candidate for context and political sensitivity before it moves to the live log
Config owner: PM
Runs alongside: Status report drafting, which reads the live log the next morning
Last reviewed: 1 October
Review next: Quarterly, or when the channel list changes

Meeting actions summary

Task: Actions summary after the weekly delivery meeting
Done by: Agent on its own
Tool and set-up: Meeting assistant in the video call tool, team template
Produces: Action list posted to the project channel
Human review: None. A wrong or missing action is spotted in the channel and fixed within a day
Config owner: Project coordinator
Runs alongside: None
Last reviewed: 15 September
Review next: If a missed action causes a slip, move it to "AI drafts, person approves"

Budget variance commentary

Task: Monthly budget variance commentary
Done by: Person
Tool and set-up: AI used only to tidy wording, no saved set-up
Produces: Commentary section of the monthly finance pack
Human review: N/A
Config owner: Finance lead
Runs alongside: None
Last reviewed: 1 October
Review next: Quarterly capability review

The three levels

Agent on its own suits mechanical work where a wrong output is cheap and quick to correct: formatting, filling a template, extracting fields from a consistent input.

AI drafts, person approves suits work that is useful as a start but needs a person’s read of the context before it has any consequence.

Person covers work that depends on relationships, politics or project history nobody has written down. AI may help in the background, but the work is yours.

The register doesn’t aim to move everything to the first level. It records the current decision so the team can agree it and change it on purpose.

Runs alongside matters once several agents share inputs. A wrong assumption early in one can show up in all of them at once, and this field tells you which entries to check together.

Check it before you share it

Every answer should be yes.

• Does every task AI touches on the project have an entry, including the ones you set up for yourself?
• Does every “Agent on its own” entry give a reason why no review is needed?
• Could someone covering your leave switch off any automation using only the Config owner and Tool and set-up fields?
• Does every entry have a Review next date or trigger?
• Has someone else on the team read it and recognised the operating model?

When it does not apply

A one-off question to a chat assistant, with nothing saved and nothing scheduled, doesn’t need an entry. Start the register when a piece of AI use repeats, runs on a schedule or feeds something another person relies on.

What this feeds

Each entry is your project’s line in the register of an agent governance charter, which adds the risk questions and controls. Before you move a task to “Agent on its own”, run it through the Agent QA Checklist. When you hand the task over, the Outcome Brief records what done looks like.