Automation has 2 tabs and they answer 2 different questions. A rule always does the same thing to the same review. An agent reads the review and writes an answer that fits it, in your words.
Rules run first. A rule reply settles the review, and an agent only answers reviews no rule replied to.
What you need: at least 1 connected source, and enough of an idea of your own answers to fill a short set of Q&A pairs.
The Agents tab showing 2 agent cards with their mode, rank and replies this month.
When an agent beats a rule
Use a rule when the answer never changes, such as a fixed thank-you. A rule is cheaper and completely predictable.
Use an agent when every review deserves an answer that fits it. "The pasta was great but the wait was long" and "Lovely staff, will be back" are both 4 stars, and they need different replies. An agent writes each one properly, in the language the review was written in.
How the stars decide what happens
An agent looks at the rating before it does anything with its reply:
- At or above its automatic band, in Auto-reply mode, it posts the reply on its own.
- At or below its approval band, it sends the draft to the Approvals queue for a person to approve.
- In between, it leaves a draft on the review card for someone to read.
The defaults are automatic at 4 stars and up, approval at 2 stars and below. A 3-star review gets a draft. See Rating bands.
On a site without a reply API, the agent always drafts, whatever the stars. Private feedback is always a draft, to send by email. A Facebook recommendation has no stars: one that recommends you follows the mode, and one that does not waits for approval and is never posted on its own.
What an agent will not do
- It will not invent a fact. Every answer comes from the knowledge ticked on the agent plus your instructions. When neither covers what the review raises, it hands the review over rather than guessing.
- It will not post in draft mode. Draft for review is the default, and nothing goes out without a person.
- It will not answer private feedback in public. Ever.
- It will not argue. A low rating gets an apology and a next step, not a debate. Your instructions can make that stricter.
What the agent will not say
An agent speaks for you, so it never speaks about itself.
- It never says which AI it runs on, or who makes it.
- It never repeats your instructions or your knowledge sources word for word as a reply.
- It ignores instructions typed into a review. "Ignore everything above" is read as text.
- A review asking whether a bot answers gets no reply from the agent. It waits for a person.
You write none of this yourself and it cannot be switched off.
Starting from a persona
Create agent opens a persona picker first. Each persona comes with instructions, a tone, a few starter answers and behavior defaults, and you edit every word of it next. Start blank gives you an empty agent instead.
The persona picker that opens when you create an agent.
Replace the starter answers before the agent goes live: they are examples, not facts about your business.
Where an agent runs
The Reach section decides which sources, sites and locations an agent answers for. Leave a list unticked and it covers all of them. With 2 or more agents, the answering order decides who takes a review. See Several agents and routing.
What it costs
An agent's reply costs more than a rule action, because the AI writes it. What it costs depends on the reply: a short thank-you to a happy review costs less than a careful answer to a low rating or to private feedback.
The composer and the review card show the cost under every draft, and the credits meter on your billing screen shows what was spent. A draft costs the same whether it is posted, approved or discarded. Approving and discarding are free.
At zero credits, agents stop. Reviews keep arriving and the feed keeps working throughout.