The AI marked an $8,680 wedding as booked. The couple was still "leaning toward you."

"We're leaning toward you guys. Let me talk to my fiance."

That's what a bride said on a call with a small caterer. The AI logged it in the customer sheet as Booked. So nobody asked for the $2,604 deposit.

AI Prompts, Reviewed takes one prompt or workflow from How I AI, Claire Vo's show where guests share their screens and walk through how they really use AI. We run it on a small-business job, score it, and tell you what we'd change.

Credit: Based on "Automate CRM Updates with Claude Projects and Zapier MCPs," a workflow from How I AI (hosted by Claire Vo, written up by the ChatPRD team), shown by Reid Robinson, product manager at Zapier, in the episode "Reid Robinson's Zapier Workflows for CRM Automation, Meeting Prep, and Feedback Loops." Original workflow: https://www.chatprd.ai/how-i-ai/workflows/automate-crm-updates-with-claude-projects-and-zapier-mcps . Full episode: https://www.chatprd.ai/how-i-ai/zapier-workflows-for-crm-automation-meeting-prep

What the original does

Reid gives Claude a locked-down set of Zapier tools that can only touch his CRM. A Claude Project holds the rules: find the person by email, check for a meeting that day, update if it exists, create only when nothing matches. He drops in meeting notes, and Claude fills in next steps and deal details, then reports every change.

The test

Composite (made-up) data and simulated drafts for a catering company: the owner plus 3 staff. Their "CRM" is a Google Sheet with 212 contacts and a Stage column. Test clock: Sunday, Oct 11, 2026.

We gave the AI notes from 6 calls in one week. The right result was 5 updates to existing customers and 1 new contact.

We ran it the way most owners would: the step list, with no written rules for each field. The AI did 3 updates and 3 new contacts. Two problems:

1. It made duplicates. An office manager who orders weekly lunches emailed from her personal address. No email match, so the AI created a new "lead." Her real record holds 9 orders at $480 this year, 9 x $480 = $4,320. The new record showed $0 and tagged her as a cold prospect. The second came from "Kathryn" spelled "Katherine."
2. It guessed the Stage. The bride's wedding is Saturday, May 15, 2027: 140 guests at $62 a plate, 140 x $62 = $8,680. She said "leaning toward you." The AI picked Booked. The owner's follow-ups only go to Quoted rows, so the 30% deposit request, 0.30 x $8,680 = $2,604, never went out.

The page's own notes name the fix: "Document each field, allowed values, and examples in the Claude Project instructions."

The rerun

We added one line per field. Booked means a signed contract and a deposit received. Quoted means a price was sent. Lookup order: email, then phone, then name. If more than one record could match, stop and ask.

The result: 4 updates, 1 new contact, and 1 question. "Kathryn or Katherine? Two possible records, no email in the notes. Which one?" The wedding stayed at Quoted, with a deposit follow-up set for Tuesday.

Booked means the deposit cleared. Your AI won't know that unless you write it down.

Verdict: 7/10

Strengths: search before you create, and report every change. The "ask when it's unclear" rule saved us from a third duplicate.

Weakness: Zapier, a Claude Project, and HubSpot or Coda is a lot for a 4-person shop. And the field rules it leans on are the one file small owners never wrote.

Founder tweak: before you connect anything, write one sentence per Stage that ties it to money. If a stage can't be checked against a contract or a payment, the AI will guess.

Copy the MCAFax version

This is our own rewrite, not the original prompt. Copy the original from Reid Robinson on How I AI: https://www.chatprd.ai/how-i-ai/workflows/automate-crm-updates-with-claude-projects-and-zapier-mcps

You update my customer list after calls. My list is a spreadsheet (or CRM) with these columns: Name, Email, Phone, Stage, Deal amount, Next step, Next step date, Source.
Stage rules:
Lead = they asked but I sent no price.
Quoted = I sent a price.
Booked = signed contract AND deposit received.
Lost = they said no or went quiet 30+ days.
For each call note I paste:
Step 1. Pull out the person, the date, and any amount.
Step 2. Look for an existing row by email, then phone, then name.
Step 3. If one row matches, update it. If none match, make a new row. If more than one could match, stop and ask me.
Step 4. Set Stage only from the rules above. If the note doesn't prove a stage, keep the old one and tell me why.
Step 5. Recompute every amount and show the math.
Step 6. List every row you changed or created, field by field, quoting the note as the source.
Never delete a row. Never mark Booked without a deposit in the note.
My call note: [example: 140-guest wedding, $62 a plate, May 15 2027, "leaning toward you"]

Paste into Claude, ChatGPT, Gemini, or Copilot.

Before you let AI touch your customer list

1. Write one line for each Stage. Tie Booked to money received.
2. Make a copy of the sheet before the first run.
3. Run the MCAFax version on 5 old calls you already logged by hand.
4. Count updates, new rows, and questions. Compare with what you did.
5. Search the sheet for duplicate names once a week.
6. Keep the follow-ups moving: https://mcafax.com/notebook/prompts/lead-follow-up-automation (more prompts at https://mcafax.com/notebook/prompts)

SBA on managing your finances: https://www.sba.gov/business-guide/manage-your-business/manage-your-finances