Open a Maps listing with four hundred reviews and eleven owner replies, then scroll it the way a customer does. The impression is not "busy business". It is "nobody is home".
Replying to reviews used to be a small job you did on a Friday afternoon. Then review requests got automated, inflow went up, and the reply queue turned into an operational problem with a real cost attached. That is what AI drafting is actually solving. Whether it solves it well depends almost entirely on what happens between the draft and the publish button.
Why response rate became a business problem
The arithmetic is unkind. One location collecting twenty reviews a month is a twenty-minute weekly task. Forty locations collecting twenty reviews a month is eight hundred replies, and at three minutes each - reading the review, checking whether the complaint is real, writing something that is not identical to the last one - that is forty hours. A full working week, every month, on a job nobody wants to own.
So it gets skipped, and selectively. Five-star reviews get a one-line thanks. One-star reviews get avoided, because they require a decision. The worst possible pattern follows: the reviews that most need an answer sit unanswered, in public, on the part of the profile with the highest read rate.
Google surfaces owner replies directly under the review in both Maps and Search. A prospect comparing three businesses is reading them. Reply behaviour is one of very few signals on a listing that shows how a business handles being challenged, which is exactly what somebody choosing a contractor or a clinic is trying to work out.
What customers actually notice in a reply
Customers read replies in a vertical column, one after another. That is the thing template writing never survives. An email template is invisible because each recipient sees one copy. A review reply template is visible because the tenth reader sees ten copies, stacked.
- The opening words - if six consecutive replies start "Thank you for your feedback", the reader stops treating any of them as real.
- Whether the reply engaged with the content - a review complaining about a two-week wait for a callback, answered with "we are sorry you did not have a great experience", reads as not-read.
- Marketing language - taglines, service lists and offers inside a reply to an unhappy customer. Nothing destroys credibility faster.
- A name - replies signed by a person, even a first name and a role, read differently from replies signed by a brand.
Where AI drafting helps and where it fails
The honest split is not "AI is good" or "AI is risky". It is task by task.
It genuinely helps with the four and five star pile. Those replies need to be warm, varied, specific enough to prove the review was read, and produced quickly. A model given the review text and a voice brief does that well, and the failure mode is mild: a slightly generic sentence in a reply to a happy customer costs you almost nothing.
It also solves the multilingual problem. Replying in the language the review was written in is now a non-event rather than a translation ticket.
It fails, sometimes badly, in a small and predictable set of places:
- Facts it does not have - a model will cheerfully write "our team attempted to contact you three times" when nobody did. It has no access to your CRM and it fills the gap plausibly.
- Apologising for the wrong thing - reflexive apology to a complaint that is factually incorrect, which in some sectors is an admission you did not want in writing.
- Sarcasm and ambiguity - "brilliant, only waited two hours" gets classified as positive and gets a cheerful reply.
- Named staff - reviews that praise or blame an individual. Publishing an employee's name in a public reply is a decision a human makes.
- Anything with money or liability in it - refunds, compensation, clinical outcomes, legal matters. Not in public, and not auto-drafted.
A review-reply policy worth writing down
If you manage reviews for clients, this document is what stops the process degrading the moment the person who understood it goes on holiday. One page, agreed with the client, covering:
- Response times by rating - typically same day for one and two stars, two or three days for the rest. Pick numbers you can actually hit.
- Who approves what - four and five star replies can go out on junior approval; anything at three stars or below gets a second reader.
- What never appears in a public reply - discount offers, appointment details, order numbers, health information, or anything confirming the reviewer was a customer where that is itself sensitive.
- Whether staff names may be used - and whether the reviewer's first name may be used.
- The offline handoff - the exact contact route offered in a complaint reply, and who monitors it. Inviting someone to email an unmonitored address is worse than not inviting them.
- Escalation triggers - allegations of injury, discrimination, food safety or regulatory breach go to the client before anything is published. No exceptions, no drafting.
Handling negative reviews with AI assistance
The sequence that works is not "generate and send". It looks more like this.
Read the review twice and establish what actually happened before drafting anything. Most negative reviews contain a checkable claim: a date, a service, a name. Check it internally first. The single most valuable sentence in a complaint reply is the one that proves somebody looked.
Then draft. Give the model the review, the voice brief, and the verified facts you are willing to state in public. Ask for a reply that acknowledges the specific issue, says what has been done or will be done, and offers a route offline. Ask for it short.
Then cut. Generated complaint replies are almost always too long and too apologetic. Remove the second apology, remove the brand language, remove any sentence promising something nobody has confirmed. What survives is usually three or four sentences, which is the right length anyway.
Do not argue in public, even when you are right, and particularly when you are right. If a review is defamatory or comes from someone who was never a customer, that is a removal request to Google, not a rebuttal in the reply field. And when a complaint does get resolved, edit the published reply to say so. A reply ending "this has now been sorted" is worth considerably more to the next reader than the original was.
Where a tool fits in this
The operational shape you want is drafts generated automatically, humans approving them, and a record of who approved what. Ampli5 Pulse works that way: AI-assisted replies you review before they publish, every connected location in one queue rather than forty dashboards, and team roles so a junior can draft while someone senior signs off the sensitive ones.
What no tool solves is the policy above. A drafting model with no approval step and no written rules will fill a listing with pleasant, plausible, occasionally untrue replies faster than any human could. Decide the rules first, then automate the typing.