GUIDE
AI can draft your review replies for under a cent each, but the bad reviews still need your facts
Yes, let AI draft your review responses. On your own Claude or ChatGPT key, 100 replies cost well under a dollar. The catch is the angry ones, where AI will invent apologies, promises, and details unless you hand it the real facts first.
A reply costs a fraction of a cent, so price is not the real question
AI can write your review responses, and on your own Claude or ChatGPT API key each one costs well under a cent. The hard part is the quality of the reply. An AI reply is only as good as the facts you give it, and the reviews that matter most are the ones where it is most likely to make something up.
Here is the math, so you can check it against your own numbers. A typical reply job has three parts. First, your standing instructions: tone, business name, sign-off, and the things it must never say. Call that about 400 words. Second, the review itself, around 100 words. Third, the draft that comes back, around 100 words.
OpenAI's rule of thumb is that one token is about three-quarters of a word. That makes the job roughly 700 tokens in and 150 tokens out. Mid-size models have been listed at around $3 per million input tokens and $15 per million output tokens. Prices change, so open the Anthropic pricing page or the OpenAI pricing page and plug in today's figure.
At those example prices, 700 input tokens cost about $0.0021 and 150 output tokens cost about $0.00225. Together that is under half a cent per reply. A photographer who gets 20 reviews a month would spend about 9 cents. A busy shop with 100 reviews would spend about 44 cents. A multi-chair clinic with 300 reviews would spend about $1.32.
So when a tool charges a monthly fee for "AI review replies," the AI is not what you are paying for. You are paying for the inbox that pulls reviews in, the alert when a one-star lands, and maybe a dashboard across locations. Those can be worth money. Just know which thing you are buying. If you do use your own key, give this job its own key so you can see exactly what it spends. We covered how in giving every tool its own API key.
AI writing your reply is fine, but AI writing your reviews is not
People mix these two up, and only one of them can get you fined.
Replying to a review is your business speaking in public under its own name. Whether you typed the words or a model drafted them, it is your statement, and you are responsible for it. That is the whole rule. Read it before you post.
Writing reviews is a different matter. In 2024 the Federal Trade Commission finalized a rule that bans fake reviews and testimonials. That explicitly includes AI-generated reviews that misrepresent who wrote them or what they experienced. The same rule covers buying reviews and suppressing negative ones, and it allows civil penalties.
The gray area to stay out of is this. Some tools and some well-meaning advice suggest using AI to "help" a happy customer by drafting a review they can paste. Do not do it. The moment your software writes the customer's words, you are closer to the thing the rule bans than to anything worth the risk. Ask for the review, make the link easy to find, and let people write their own three clumsy sentences. Those read as real because they are.
Where AI review replies go wrong
Most bad AI replies fail in one of five predictable ways. All five are fixable once you know to look.
Every reply sounds the same
Ask for a reply to 40 five-star reviews one at a time and you will get 40 versions of "Thank you so much for your kind words! We're thrilled you had a great experience." A customer scrolling your profile sees the pattern in seconds, and it reads like nobody is home. The fix is to paste five reviews at once and ask for five replies that each mention something specific from their review and do not share an opening line.
It invents facts to sound sorry
This one is the most dangerous. Give a model a complaint and nothing else, and it fills the gap with plausible apologies: "We've spoken with the technician involved," "We've updated our process," "We'd be happy to offer a refund." None of that happened. Now it is in public, under your name, and the customer will hold you to the refund.
It confirms things you should not confirm
A clinic, a therapist, a dentist, a lawyer, or even a photographer who shot a private event should never confirm that the reviewer was a client, let alone what for. AI does not know that. It will helpfully write "We're sorry your crown didn't fit properly after your visit on the 12th." The US health regulator has fined practices over exactly this. A dental practice paid $10,000 in 2019 after its replies to online reviews disclosed patient information.
It argues, at length
If you are angry when you write the prompt, the model mirrors you. You get four polite-sounding paragraphs that are really a rebuttal. Nobody reading a one-star reply wants a rebuttal. They want to see that you are reasonable.
It misreads the review
Sarcastic five-star reviews, reviews meant for a different business, and reviews that mention two staff members all trip it up. You catch these by reading the draft, which is the one step you cannot automate away.
Give the AI your facts and your rules, not just the review
The prompt that works has two parts. One is a standing set of rules you save once. The other is a short line of facts you add for any review that is not simply happy.
The standing rules, saved once:
You draft public replies to customer reviews for [business name], a [what you do] in [town]. Write in plain, warm, first-person-plural English. Keep replies under 80 words. Sign off as [your first name]. Mention one specific thing from the review. Never confirm that the reviewer is or was a customer, patient, or client unless they said so themselves, and never repeat any detail about their job, visit, or health. Never offer a refund, discount, or free service. Never claim we have done something unless it appears in the FACTS line. If a review is negative, thank them, own what is in the FACTS line, and invite them to contact [name] directly at [phone or email]. If there is no FACTS line on a negative review, reply only with: NEED FACTS.
That last instruction does the heavy lifting. It forces the model to stop and ask you, instead of guessing.
Here is a worked example. Say you run a body shop. A customer named Dana leaves two stars: "Took three weeks to fix my bumper and nobody called to tell me why." You paste the review and add one line:
FACTS: Parts were backordered from the manufacturer for 12 days. We should have called her and didn't. Our service manager is Luis, 555-0142.
The draft that comes back will be close to this: "Dana, you're right, and I'm sorry. The part was backordered for almost two weeks, and we should have called you as soon as we knew. That's on us, and we're changing how we update customers when parts are delayed. If anything about the repair isn't right, please call Luis directly at 555-0142. — Mike"
Read it before you post. Notice the line about changing how you update customers. It crept in even though it was not in the facts. If you are not actually changing anything, cut it. That edit is the job, and it takes ten seconds.
You can run this in a free chat window if you only get a few reviews a week. A saved prompt and a paste is enough, which is the same idea as using a chat window and your best past example for proposals. The API key starts to matter when the drafting lives inside a tool you already use every day, or when you want to see the cost per job.
Which reviews to hand to AI and which to write yourself
A simple split covers almost every business. Let AI draft the positive and neutral reviews in batches, and skim before posting. Write the negative and sensitive ones yourself, or at minimum supply the facts and rewrite the draft line by line.
The reasoning is about what is at stake. A bland reply to a five-star review costs you almost nothing. A wrong reply to a one-star review is the thing future customers actually read. People scanning your profile skip past the praise and go looking for how you handle a complaint.
A few categories should never get an AI-first draft:
- Anything touching health or treatment. Clinics, dentists, physios, therapists, med spas. Use one fixed, human-approved reply that does not confirm the person was a patient and invites them to call.
- Injury, damage, or safety claims. A contractor accused of damaging a floor, or a shop accused of an unsafe repair. Anything you write may be quoted later. Keep it short and take it offline.
- Money disputes. If a refund or a chargeback is in play, the reply must not promise or deny anything. Say you'd like to sort it out directly.
- Reviews that look fake or meant for someone else. Do not let AI apologize for something that never happened. Reply briefly that you have no record of the issue and ask them to get in touch, then use Google's own review tools to flag it if it breaks policy.
When a paid review tool is the better buy
Sometimes the right answer is to pay for a review platform. If that is you, do it without guilt. The AI inside those platforms is not what earns the fee. These are the situations where the fee is worth it:
- You have several locations. Logging into five Google profiles every week is the real cost, and a single inbox fixes that.
- Reviews come in on many sites. Google, Facebook, Yelp, and an industry site like Healthgrades or Houzz each have their own login and alerts.
- Staff reply, and you want to approve first. An approval step for front-desk replies is worth paying for in a clinic.
- You want text-message review requests. Asking by SMS right after a job tends to get more replies than email. Dedicated platforms do this well.
Tools like Birdeye and Podium are built around exactly that list. Get quotes and compare them against how many reviews you actually receive. If you have one location and mostly Google reviews, though, the free Google Business Profile app already sends you a notification for each new review. A saved prompt on your own key covers the drafting. That combination costs nothing per month beyond pennies of usage.
To be plain about where we fit: Fynl is not a review-monitoring tool, and it does not send SMS yet. Its AI runs on your own Claude or ChatGPT key, so the same key you use for drafting emails and quotes can draft review replies at the cost above. Posting the reply still happens in Google. If you need multi-site monitoring or text requests, a dedicated review platform is the better fit.
A fifteen-minute weekly routine that keeps replies human
Replying to every review within a few days matters more than the polish of any single reply. A routine you will actually keep beats a perfect one you abandon in March.
Once a week, open your reviews and sort them into happy and not happy. Paste all the happy ones into your saved prompt together and ask for one distinct reply each. Read them, fix any that misread the review, and post. That part takes about five minutes for a normal week.
For each unhappy review, write the FACTS line first. Write what actually happened, what you will own, and who they should contact. Then get the draft and edit it. Cut anything that apologizes for things that did not happen or promises things you have not decided. If you find yourself writing the same FACTS line three weeks running, the review is not the problem. The process is. That pattern is the most useful thing your reviews will tell you, and no tool will spot it for you unless you are the one reading them.
Questions people also ask
Will Google penalize my business for AI-written review replies?
We know of no Google policy that bans AI-drafted replies. Google's guidance is about what a reply says: it should be relevant, respectful, and free of personal information. A reply that breaks those rules is a problem whether a person or a model wrote it.
Should I set replies to post automatically?
Not for negative reviews, ever. Automatic posting is possible through Google's Business Profile API, but the savings are a few minutes a week. The risk is a public reply that invents a refund or confirms a patient's visit. Auto-drafting is fine. Auto-posting is where it goes wrong.
Is the free ChatGPT or Claude chat window good enough?
For a handful of reviews a week, yes. Save your rules prompt and paste reviews in. An API key is worth it when the drafting lives inside a tool you already use, or when you want to see the cost per job. Both OpenAI and Anthropic say they do not train on API data by default, but check the settings on any consumer chat account you paste customer reviews into.
How short should a review reply be?
Under 80 words covers almost every case. A happy review needs two or three sentences that mention something specific. An unhappy one needs an acknowledgment, what you own, and a direct contact. Long replies to complaints read as arguments.
Sources
More on ai you actually own
An AI receptionist trial should test what lands in your calendar, not how human the voice sounds
Make ten scripted calls yourself, then check four places after each one: your calendar, your phone alerts, the
AI YOU ACTUALLY OWNThe best free AI proposal generator is a chat window and your last winning proposal
The only AI proposal generator that is free and good is the free version of ChatGPT or Claude, given your real
AI YOU ACTUALLY OWNAI writes a proposal for a few cents, so pay for the software that gets it signed
An AI proposal generator writes the words. Proposal software sends them, gets them signed, and collects the de
One portal instead of five subscriptions
CRM, calendar, quotes, invoices, website and email — with the AI running on your own Claude or ChatGPT API key, so there is no markup on it.
See pricingMore posts