Reviews are read twice. Once by the person who wrote one, and then for years by everyone who is deciding whether to come to you. The second reader looks at something else entirely: not the complaint, but how you answered it. A calm, specific reply under a two star review sells better than ten five star reviews with silence underneath.
That is why owners look for a way to answer all of them, and why most of them give up after a month. Writing twenty replies a week is dull work that needs a steady voice at the exact moment you have no patience left. An AI agent is good at exactly that, as long as it is set up to write like a business and not like a template.
Why most AI review replies read like a bot
A reply generator that only sees the review text has nothing to work with, so it fills the space with manners. You get the same four moves every time, and a reader scrolling your page sees the pattern in seconds:
- The stock opener. “Thank you for your valuable feedback” on the first review, the second and the ninth.
- No detail. The guest named a dish, a room number, a treatment, a delivery time, and the reply mentions none of them.
- An apology with no next step. “We are sorry for the experience” and nothing a person can actually do.
- One voice for every mood. The same cheerful register for a happy regular and for someone who waited an hour.
None of this is a model problem. It is a missing input problem: no brand voice, no facts about what happened, no rule about who decides.
What the reviews and reputation role is
In SVI, answering reviews is not a feature bolted onto a chatbot. It is one of the 19 roles you can put in a team in the agent builder: reviews and reputation. The role comes with ready rules and works next to the other agents, so it shares the same picture of your business as the one taking bookings or answering questions in Direct.
That shared picture is the whole difference. The agent knows your menu, your services, your hours and what you have confirmed as true, because it answers only from your own data. It writes in the language the review was written in, across the seven languages of the interface, and it does not pretend to be a person: if asked, it says plainly that it is an AI assistant.
How the agent learns your voice
Tone is not a slider from formal to friendly. It is built from your own material, the same way a new employee would learn it by reading your best replies:
- Your past replies. The ones you were happy with become the pattern: how you greet, how long you write, whether you sign with a name.
- Your rules in plain words. “Never argue about the bill in public”, “always offer a callback”, “no exclamation marks” go into the role as written, with no code.
- The best model for the job. The system holds 89 of the best AI models, and each role runs on the one strongest for its kind of task. Writing a careful reply to an angry guest is not the same task as calculating a price.
- Variety on purpose. The agent is told not to reuse openers, so a reader going down your page does not see the same sentence five times. If you have seen this in a restaurant feed, the same logic is described in AI agents for restaurants.
Negative reviews: acknowledge, do not argue
This is where tools get businesses in trouble, so the role has hard limits. A complaint is answered for the audience, not won against the author:
- Acknowledge without confessing. The agent recognises the experience and does not agree to facts nobody has checked yet.
- No point by point rebuttal. You can disagree without calling a customer a liar, and in public you do it in one sentence.
- Nothing invented. No prices, no promises, no discount or refund the owner has not approved. If the data is not there, the agent does not guess.
- No private details in public. One customer's data never appears in a reply another customer can read, so the order, the diagnosis and the address stay out of it.
The practical shape of a good public reply is short: acknowledge, one specific fact, one concrete next step, and a way to continue privately. Everything else belongs in the conversation that follows.
When a person steps in
Some reviews should never be answered by software alone, and the role is built to recognise them and stop. When the agent is unsure, it hands the case to a person with a short summary of what happened and a draft it would have sent:
- Anything about money, safety or health. A refund claim, an allergy, an injury, a lost item of value.
- Anything legal or about a named employee. A threat to go to a regulator, an accusation against a specific member of staff.
- Anything that smells like a fake or a competitor. A review with no visit behind it is a decision for you, not for an agent.
You also decide the publishing rule for everything else. In test mode the agents prepare replies and publish nothing at all. After launch, publishing follows the rule you set: every reply on your yes, or by the approved pattern with the hard cases still waiting for you.
Where the reviews come from and where the replies go
Reviews do not arrive in one place. They land on maps, in your social channels, in email and in private messages that never become a public review but hurt just as much. You connect the channels you use, and the role works across them with the same rules and the same voice.
It also stops being a separate chore. In the marketing team solution the same role sits next to social media, company copy, a designer and an analyst, and the time it takes you to answer reviews and comments becomes one of the things measured before launch and after.
Build the role into a team of your own
Reviews are rarely the only thing an owner wants covered, and a lonely review bot is not worth the setup. The point of a company of agents is that the one answering a complaint and the one taking the booking are part of the same team.
You can start from a ready solution on the solutions page, where reviews and reputation is already part of the restaurant, property, reception, cargo and marketing teams. Or you describe your business to Mai and she assembles the team from the roles that fit, including a role of your own that does not exist in the list.
Test it on your own reviews before you pay
No demo will tell you whether the tone is right for your business. Your own reviews will, which is why every team starts the same three steps:
- Install. The team goes into your account and your channels are connected. This step is on us.
- Test mode. The agents answer real incoming reviews and messages, nothing reaches the public, and you read every reply that would have been sent, with a report. On us as well.
- Launch. When the replies sound like you, you switch it on. Payment starts only here.
How to start
Open the agent builder to see the roles and put the reviews and reputation agent in a team, or start with a conversation with Mai: tell her what your business is and how you answer reviews today, and she will set the role up and run the test on your own reviews before anything is published.