Every morning at 9, a workflow I'm building for a home care agency opens their Google reviews and sorts the unanswered ones into two stacks.
4 and 5 stars with nothing sensitive in the text get a short thank-you draft. Anything rated 3 or below, or anything touching a hard subject, never gets a draft at all. It goes to an email with the reviewer's name, the rating, and the full text, and a person writes the reply.
That sorting rule took longer to settle than the build did. An AI is genuinely good at a 1 star review. It stays calm, it never gets defensive, it lands the right length on the first try. That's the problem. A family writing about their mother's care can tell when the calm came free.
So the workflow has 2 gates instead of 1, and the second one exists to catch what the first one misses. The build itself is done and sitting inactive, waiting on Google to approve the connection that lets software read and post reviews. The decision doesn't need Google's approval, and it's the part worth stealing this week.
───
🗞️ THIS WEEK IN AI
1. ChatGPT's new work agent wants all of your access, not some
On August 24, TechCrunch published a long look at ChatGPT Work, OpenAI's version of ChatGPT that carries out multi-step work across your email, documents, and messaging apps. It's on the $20 plan. Asked whether it might pull from a private DM without knowing that information was off limits, the engineer leading the desktop app answered "Yes." The reporter tried to give it read-only access to a drive, hit errors, and eventually got a dialog box telling him only complete access would work.
Why this matters: "It finished the file" and "it's ready to send" are not the same sentence. If a tool only works once you hand over the whole account, keep the send button with a person until you've watched it for a few weeks.
2. Claude went down again Monday, and took part of Copilot with it
Early on August 24, Anthropic logged elevated errors across Claude.ai, the API, Claude Code, and Claude Cowork, and people saw 529 Overloaded errors for hours. The same morning, GitHub reported degraded availability for one of the models inside Copilot and blamed "an issue with the upstream model provider," telling users to switch models to keep working. The affected model was one of Anthropic's.
Why this matters: One outage turned into 2 products failing, and most people using the second one had no idea they were downstream of the first. Two weeks ago I said to write down the manual version of any task that now runs through an AI tool. If Monday cost you a morning, that's the reminder.
3. A test model got out of its box, and a state attorney general noticed
Also on August 24, Alabama's attorney general subpoenaed OpenAIover an incident Reuters reported earlier. OpenAI has acknowledged that an unreleased cybersecurity model with no guardrails escaped an isolated test environment, connected to the internet, and hacked the AI dataset platform Hugging Face. Hugging Face was one of 4 victims of what was supposed to be an internal evaluation.
Why this matters: You are never going to run that model. You are going to run tools that can browse, send, and post. The small business version of the same failure is a reply that goes out under your name before anyone reads it.
───
🛠️ THIS WEEK'S TRICK
Draw 2 Gates Before Anything Reaches Send
Most "let AI handle our reviews" plans have one path: draft, post, done. That's how a polished apology goes out under your name at 9am. Here's how to split the work this week, even if you never build a workflow.
Step 1: Split your last 10 reviews into 2 piles
Open Google Business Profile and make 2 lists, on paper or in a sheet.
Pile A is 4 and 5 stars with nothing in the text about harm, money, or a named employee. These are thank-yous.
Pile B is 3 stars or below, plus anything you'd be embarrassed to get wrong in public. These stay yours.
If a review could go either way, it goes in pile B. The point of a gate is to fail closed.
Step 2: Let AI draft pile A only
Paste this into ChatGPT or Claude. Do not paste pile B.
Draft a Google review reply for this 4 or 5 star review.
Business: [what you do, in one line]
Review text: [paste the review]
Reviewer first name: [name or "the reviewer"]
Rules:
- 2 to 4 sentences
- Thank them for a specific detail they mentioned
- No discounts, no excuses, no promises you can't keep
- If the review mentions harm, money trouble, a medical issue, or a named
staff member, do not draft a reply. Write "this one needs a person" and stop.
Give me the reply only, ready to paste.
That last rule is gate 2. Gate 1 was mechanical, based on the star rating. Gate 2 asks the model to stop itself on a review that slipped through. Either gate trips and a person writes it.
You still paste the result into Google yourself. The model does not post.
Step 3: Write pile B yourself, even if it's shorter and later
A clumsy human reply on a 1 star reads better than a perfect machine one. People can tell. A family writing about their mother's care wants to know a person read it.
Put one 15 minute block on the calendar this week for pile B. If you skip it, the unanswered 1 star keeps reading as true.
What this works best for:
Google reviews on a local service business
Any inbox where thank-yous pile up on top of the hard messages
The first month with any AI tool that offers to "handle replies"
One honest limitation: a prompt can't force a connected tool to respect your gates. If a product only works with full posting access, don't connect it yet. Draft in a chat window and hit send yourself.
───
🔧 TOOL OF THE WEEK
Google Business Profile (https://business.google.com)
This is the page your customers and the AI tools recommending you both read before anyone picks you. It's also the only place your reviews sit in one list, sorted by date, with the unanswered ones easy to spot.
Last week's issue was about whether ChatGPT can find you. This week is about what it finds when it opens your reviews, and the answer is usually a wall of 5 stars with no replies and one angry 1 star sitting there alone.
Price: Free.
Best for: Local service businesses that have Google reviews and a habit of not answering them.
Worth trying? Yes, this week, on your last 10 reviews. That's the whole test. If sorting pile A takes you more than 15 minutes, your process is still too heavy for a workflow to save.
───
💡 PROMPT OF THE WEEK
Use this after you split the piles, to turn the sort you just did by hand into a rule you can reuse.
I need a simple routing rule for review replies.
My business: [what you do]
Who answers hard reviews today: [me / an employee / nobody]
What I will never let a draft cover: [harm, refunds, named staff,
anything rated 3 or below]
Write a 2 gate rule I can follow by hand this week:
1. Gate 1, mechanical: which reviews skip AI entirely
2. Gate 2, on the draft: when a thank-you draft should still go to a
person instead of going out
Keep it under 120 words. Write it as instructions to me, in second person.
If a review is ambiguous, it fails the gate. Do not invent a third pile.
Tip: tape the rule it gives you above your keyboard for one week and note every time you break it. If you keep making exceptions, the exceptions are the real rule, and those are what you'd be automating.
───
👋 THAT'S A WRAP
The hard part was never getting an AI to write a reply. It's deciding which replies are allowed to leave without you.
This week, open your last 10 Google reviews and split them into 2 piles. Let a chat window draft the thank-yous. You write the rest. Automate the thank-yous, not the apologies.
Forward this to someone who's about to connect their reviews to a tool because the demo looked clean.
See you next Friday.