A Practical First Automation: From Inbox to Booked Job

Step 1: Decide what counts as a booking request

Before any AI is involved, write down the kinds of email you get and what should happen to each. For example:

• New job request: someone wants a visit or a quote.
• Change to an existing job: a reschedule or a cancellation.
• A question: hours, service area, what you do and don’t do.
• Everything else: invoices, vendors, newsletters, spam.

Then list what you need to book a job: the customer’s name, a phone number, the address, what they need done and when they’re available. That list becomes the checklist the automation works through, and it tells you when something is missing.

Step 2: Let the assistant sort, not decide

The first AI step reads the email and fills in a structured form: which kind of email it is, the fields it found, and the fields it couldn’t find. It doesn’t reply to anyone and it doesn’t change anything. It just sorts.

Then ordinary code checks the result. Is there a phone number in a valid format? Is the address inside your service area? Did it find a date that’s actually in the future? If the answer is unclear, or the email doesn’t fit any of your categories, it goes to a person with a short note about why.

Two habits matter here:

• The sorting step has no tools that send, book or delete. An email is information from a stranger, and some emails are written to trick software (“ignore your instructions and...”). A step that can only sort can’t be talked into doing anything else.
• “I’m not sure” is a valid answer. The automation should be allowed to hand an email to you rather than guess.

Step 3: Look things up and draft, but don’t commit

For a new job request, the next step gathers what you’d normally check by hand:

• The calendar: which of the customer’s preferred times are actually open.
• The CRM or job system: whether this customer already exists, so you don’t create a duplicate.
• Your own documents: the right service description or the answer to their question.

With that, it prepares a draft job and a draft reply, such as “I can come out Thursday between 1 and 3. Does that work?” Nothing is saved or sent yet.

This is where MCP can come in. If you use an AI assistant that connects to your calendar, CRM and email through MCP servers, those connections do the looking up. The same flow can also be built as an ordinary workflow (in a tool like n8n, or in code) with one AI step for the sorting. MCP is useful when you want the assistant to work with the same tools interactively, too. Either way, the steps and the checks are the same.

Step 4: A person approves

You, or whoever handles scheduling, see the draft job and the draft reply side by side, with the original email. Approve, edit or reject. Reviewing a prepared draft is quicker than starting from scratch, because the reading, checking and typing are already done.

This is the step people want to skip, and it’s the one that keeps the whole thing trustworthy. The MCP specification makes the same point about AI tools in general: it says there should always be a human in the loop who can deny an action, with a confirmation before sensitive operations. Booking someone’s time and emailing a customer are sensitive operations. Once the automation has a long record of drafts you approve without changes, you can decide whether some simple cases can go through on their own. That’s a decision to make with evidence, not on day one.

Step 5: Book it, confirm it, check it

After approval, the automation does the clerical part:

• Creates the job in the CRM (or updates the existing customer).
• Holds the calendar slot.
• Sends the confirmation you approved.
• Reads the records back to make sure the job and the calendar event really exist and match.

It also needs to be safe to run twice. Email systems and networks retry things, and the same message can arrive again. The automation should recognise an email it already handled, so a retry never books the same customer twice.

Step 6: Watch it, and measure your own baseline

Every run should leave a record: what came in, how it was sorted, what was drafted, who approved it and what was created. Once a week, look at the emails that went to a person and why. That’s where the next improvement comes from, whether it’s a new category, a missing field or a clearer reply template.

To know whether it’s worth it, measure before you start. For a couple of weeks, note how many booking emails arrive and roughly how long each one takes by hand. Then compare with the same numbers afterwards, including the time spent on approvals and on the emails the automation handed back. I won’t quote a typical saving here, because there isn’t an honest one: it depends on your volume, how messy the requests are and how your tools are set up. Your own numbers are the only ones that count.

What not to automate yet

Some emails should stay with a person, at least at first:

• Pricing and quotes that depend on judgment or a site visit.
• Complaints and anything emotional. A quick, human reply matters more than a fast one.
• Refunds, payments and anything legal.
• Anything involving sensitive personal information, unless the right agreements and controls are in place.

The automation can still flag and route these. It just shouldn’t draft the answer.

What you need to make it work

• An email system with an API, such as Microsoft 365 or Google Workspace.
• A calendar the automation can read, and later hold slots in.
• A CRM or job system with an API, or at minimum a structured place to record jobs.
• Narrow permissions for each connection, as described in Connecting an AI Assistant to Your Business Tools Safely.
• Someone who owns it: a person who approves drafts, reviews the exceptions and knows how to switch it off.

The short version

Define the kinds of email and the fields you need. Let the AI sort and draft, never send or book on its own. Check its work with ordinary rules, keep a person approving every booking and reply, make it safe to run twice, and measure your own before and after.

If you’d like help mapping this for your business, with your inbox, your calendar and your job system, that’s exactly the kind of process I build on business automation. We start with how the work happens today and agree on what “better” means before anything gets built.

Sources

• Tools, Model Context Protocol specification, revision 2026-07-28: human in the loop, and confirmation for sensitive operations
• Model Context Protocol specification, revision 2026-07-28: user consent and control
• n8n: error handling, official documentation