Start here: where agents actually help

An agent is useful when the work repeats, the cost of a mistake is visible, and a person still checks the result. It is a poor fit when the work is a judgement, a payment, or a promise. This page is the map. The playbooks are the instructions.

Hand it over when the job is a draft

Sorting, extracting, and first drafts are the honest wins.

• An inbox full of newsletters, with a few real customers buried in it. Labels first, then a draft.
• The same customer question every week, once the answer is written down. Replies from a FAQ.
• A pile of receipt photos that need typing, not interpreting. Checked rows.
• One true note about finished work that should exist as three short posts. Social repurposing.
• A first list of suppliers, where every price is opened in a browser afterwards. Supplier research.

Keep it when the job is a decision

Sending the email, taking the payment, filing the return, quoting a price, and promising a day are still yours. The playbooks say this in different words because it is the whole method. An automation that can mark an invoice paid, or a phone agent that can offer Thursday, will eventually do the wrong one.

Turn on the boring tool first

Before you add an agent, check whether software you already pay for does the job.

• Accounting packages already remind people about overdue invoices. Use that before Make.
• Google Calendar already offers real empty slots. Use that before a scheduling agent.
• A form and a sheet already remember enquiries. Two steps, on the free plan.
• Your own voicemail already catches a missed call. An AI receptionist is only worth it if a UK number exists and the test calls are better than a text-back.

A sensible order for a one-person business

1. Labels on the inbox, so you can see the real work.
2. A booking link, so you stop negotiating times.
3. The invoice reminder inside the accounts tool.
4. One draft habit: replies, receipts, or posts. Not all three in a week.
5. A Monday note, once the other lists exist to summarise. The digest.

If the question is one of these

The chores people ask for by name each have a page. The test log on every one of them still says pending.

• A phone that answers when you cannot: Smith.ai for a small business, the call bundle, or Lindy if you will configure it. The receptionist comparison does not pick a winner.
• A draft inside Gmail: Gemini, if the button is there. Otherwise the ChatGPT inbox. If the mail is in Outlook, use rules and a draft you paste. The email comparison is the short version.
• A booking link: Calendly's free plan, or the Google Calendar version. One diary writer.
• An estimate: trade quote drafts, after you have opened the supplier pages.
• A public reply: Google reviews, posted by you.
• A call turned into tasks: Fathom notes, with the other person told. Google Meet notes only if an eligible plan already shows the control.
• A PDF copied into a sheet: data entry. Receipt photos are filed on their own page before anyone codes them.
• A queue for posts: Buffer, after the note is true. A sheet of dates is the version where you still press publish.
• A page of answers, then a widget: write the FAQ first. Tidio comes after that page exists.
• A company you are about to contact: a lead brief from pages you opened. Supplier prices stay on the Grok page.
• A Friday email to a client: the weekly report, from figures you already have. The Monday digest is your own reminder, not theirs.
• A Drive that is one long list: the folder layout. You move the files.
• A first automation: Zapier against n8n, if the job is a form or a workflow rather than a phone.

If you want the short version of that choice, the finder asks for the task, the budget, and how technical you want to be.

How to read a playbook

The first box is the answer. If it is not your situation, stop. The steps are written so a search engine or an answer engine can quote them, and so you can follow them without a video. The UK notes cover pounds, VAT, where personal data goes, and HMRC only when the rules change what you should build. The test log says whether anyone has run the steps. On this version of the site, that status is pending. We would rather publish a gap than a fake saving.

What we will not claim

No page here says a setup saves a number of hours. We have not measured that. When a guide is run on a real account, the changelog will record what broke, not a percentage. The methodology is the longer version of that promise.