The Handoff rating
A four-level rating for how much of a job you can hand to an AI agent, and where a human must step in. It rates the job, not the agent: the same agent can be fine for a Handoff 1 job and need close supervision on a Handoff 4 job.
Handoff 1 Read-only research
The agent reads and reports. It changes nothing outside the chat.
Typical checkpoint: check the sources before you act on what it found.
Playbooks rated 1
- Sort this morning's inbox into Act, Reply, Read later and Ignore
- Shortlist three suppliers for something your business needs
- One-page snapshot of a competitor's offer and prices
- Find grants or funding schemes your business might qualify for
- Summarise an official guidance page in plain English
- Research a new lead before your first call
- Compile a weekly numbers report from your spreadsheets
- Flag stock that needs reordering
- Compare travel options for a work trip
Handoff 2 Drafts for review
The agent prepares drafts or files. A human sends, files or publishes.
Typical checkpoint: read and edit the draft; you send, file or publish it.
Playbooks rated 2
- Draft replies to routine emails for you to check and send
- Propose meeting times from your calendar in a reply draft
- Draft follow-ups for quotes with no reply after seven days
- Turn call notes into a first-draft proposal
- Draft answers to common customer questions from your own FAQ
- Draft replies to online reviews
- Turn a folder of receipts into an expense spreadsheet
- Suggest categories for bank transactions you export
- Turn your rough notes into a blog post draft
- Plan a week of social posts from one article
- Draft your monthly customer newsletter
- Turn meeting notes into actions and owners
Handoff 3 Acts, but reversible
The agent changes things that can be undone: labels, moves, scheduled items, record edits.
Typical checkpoint: approve the plan or list of changes before it acts, and keep a way to undo.
Playbooks rated 3
Handoff 4 Acts irreversibly, or with money or people
Payments, messages to customers, cancellations, public posts. A human approves each one.
Typical checkpoint: approve each individual action, or do the final step yourself.
How we choose a rating
- What can the agent change? If nothing outside the chat, it is a 1. If it only produces drafts or new files, it is a 2.
- Can the change be undone? Labels, folder moves, scheduled (not yet published) posts and record edits with version history can be undone: a 3.
- Does it touch money, people or public? Payments, refunds, messages to customers, cancellations and public posts cannot be taken back: a 4.
- Rate the riskiest step. A job is rated by its riskiest step, even if most of it is research.
Always approve these yourself
- Any payment, refund, transfer or change to bank details.
- Any message sent in your name to a customer, supplier or the public.
- Deleting files, emails or records.
- Cancelling, signing up for or changing an account or contract.
- Sharing personal or customer data with anyone new.
Several agents now ask before some of these actions by default (see Which agent?), but defaults change and can be switched off. The checkpoint in each playbook is the one to keep, whatever your agent's settings say.
What the rating is not
It is not a test result, a security certification or legal advice. Playbooks are reviewed against official product documentation and dated; they are not claimed to have been tested on every agent.