An AI agent that answers, acts and knows when to hand off
It replies from your own help center and docs, with a source on every answer. It looks things up and takes the actions you allow. When it isn't sure, your team gets the ticket with a short note.
Every reply follows your knowledge and your rules
A customer writes in. The agent finds the facts, checks what it's allowed to do, then answers or passes the ticket on. You can see every step it took.
Answers from your docs, with the source attached
Point it at your help center, your website, PDFs and solved tickets. It keeps up as they change.
- Searches by meaning and by keyword, then keeps the best few pieces
- Shows where each answer came from
- Checks every claim against a source before it sends
Write how you handle a case in plain English
A playbook is a set of steps for one kind of request, like a double charge or a lost package. Rules are short policies that apply when they fit.
- Your playbook is turned into steps and checked for gaps before it runs
- Rules like "never promise a ship date" apply only when they fit the moment
- Fixed wording for sensitive moments, like refunds or outages
- Versions with Draft, Testing and Live status
Takes action in your apps, only as far as you allow
Connect Stripe, Shopify, HubSpot, Linear and more with one click. You decide which actions run on their own and which wait for a person.
- Lookups only for a verified customer
- Customer details are filled in by code, so the model can't change them
- A one-time code by email before sensitive actions
Passes the hard ones to your team with a summary
The agent hands off when a customer asks for a person or seems upset. It also hands off when the answer is missing, after two failed tries, or when a fact check fails.
- A private note sums up the thread and what was tried
- The right team is assigned and the customer is told
- An out-of-office reply after hours
You set how far each action can go
Every action in a connected app has a risk level. Amount limits and daily caps sit on top, and each action is logged on the conversation.
Runs on its own
Low-risk steps like resending an invoice or sharing a tracking link. The agent does them right away.
The customer taps to confirm
Changes the customer should agree to, like booking a call or starting a return. Buttons appear in the chat.
A person approves first
Money and account changes, like refunds or cancellations. A card shows up in your side panel to approve or decline.
Help for your team on the tickets that need a person
Copilot works inside the reply box and the side panel. It drafts, explains and digs, and your team decides what gets sent.
Draft and improve
Write a first draft, fix the tone or make a reply shorter, right in the reply box.
Translate
Read the customer's message in your language and send your reply in theirs.
Ask Copilot
Ask about a customer in plain English: "Did we talk about pricing with them last month?"
Investigate
For hard bugs, it checks recent deploys, errors, team chat and similar tickets, then writes a likely cause with links.
Summaries
One click sums up a long thread, so whoever picks it up can start right away.
Suggested next step
Recommends a macro, a playbook or the team to route to.
It earns its way to Live
Start in Draft mode, where the agent writes and your team sends. A playbook goes Live only when its tests pass, and every change runs them again.
See how well it works, and stop it any time
Replay tests
Run your own past conversations through it before go-live. See solved, wrong and handed-off counts for each playbook.
Simulated customers
Practice runs with angry, vague or tricky customers, built from the cases where it went wrong before.
Golden tests
A set of must-pass cases for each playbook. Any change that breaks one stays out of Live.
Watch
Scores every conversation, by the agent or a person, against checks you write in plain language.
Monitors
Track a theme like "customers mentioning a competitor" with counts and alerts.
Explain any reply
See each step, rule, lookup, time and cost behind a reply, right in the side panel.
Guardrails
A fact check and a rule check on every reply. Attack attempts are caught, and personal data is masked in logs.
Kill switches
Turn the agent off for one inbox, for your whole team or everywhere, in one click.
Spending caps
Set a monthly limit. You get an alert well before you reach it.
Replies that fit the customer and the channel
Answers in the customer's language
The agent spots the language and replies in it. Your team sees both sides translated, so anyone can step in.
Written for where the customer is
Short for website chat. Formatted for email. Plain for SMS. On WhatsApp it stays within the 24-hour reply window.
Draft or Auto, per inbox
Pick Draft mode, where your team reviews each reply, or Auto mode, where the agent sends on its own. Set business hours for each inbox.
What's ready and what's coming
The inbox, channels, help center, automations and reports are live today. The AI agent is rolling out to early access teams in stages.
Common questions
Does it make things up?
It answers only from your knowledge and your connected apps. Before a reply goes out, a fact check confirms each claim has a source. If the check fails, the ticket goes to your team.
Can it issue refunds?
Yes, if you allow it. Most teams set refunds to need a person's approval, with an amount limit and a daily cap. Each refund is logged on the conversation, and a retry can never run it twice.
What happens when it doesn't know the answer?
It hands the ticket to the right team with a short note about what the customer needs and what was tried. The customer is told a person will reply.
Which AI model does it use?
Claude, by Anthropic. Quick tasks like sorting and checks use a small, fast model. Replies use a larger one, and hard investigations use the most capable one.
Can I try it on past tickets first?
Yes. Replay tests run your past conversations through the agent and show how many it would solve, hand off or get wrong. You can also start in Draft mode, where your team sends every reply.