Claude is already open on my machine. Fernand's inbox is usually the tab I open when I remember, which is often later than I meant to.
I wanted Claude to help with that part of the day.
There are two different agents involved in support, and it's easy to start with the wrong one. I did. I turned on the chatbot and thought I was done. The next morning, the inbox was still a mess.
Fernand's AI is for your customers. It can answer in the chat widget, or help draft a reply in the inbox, using the information you've trained it on.
Claude is for you. It can read the inbox through the API, draft replies, tag conversations, and file a bug. It can leave the refund for you to decide.
They can work together. They do different jobs.
What I let it do
For the morning pass, I give Claude a key that can read conversations, create drafts, add tags, and change status. It can't send replies.
The instructions are simple on purpose: read what came in overnight. Draft replies when the help center already has the answer. Skip billing questions, refunds, and customers who are clearly upset. Tag conversations it has handled, so the reporting stays honest. Snooze anything waiting on the customer until tomorrow. Then give me a short summary I can read with my coffee.
I paste the prompt into Claude Code. Cursor and Codex can use the same instructions if that's what you work in.
Claude reads the docs we wrote for agents and the OpenAPI spec. If it needs a key, it asks. You can create one under Settings → Integrations → API keys.
I also tell it not to commit the key or print it back to me. I learned to include that instruction the hard way.
Give each job its own key
An agent with permission to reply can email a customer while you're in a meeting. That can look like the magic part of the demo. It's also the part that makes me want to close the laptop.
Drafts and private notes can't reach a customer. So the morning pass gets a key that can draft, but can't send. If I want another agent to file Linear issues, I give that job a separate key with the permissions it needs. If something starts acting strangely, I can revoke its key.
The permissions for each job are in AI agents. I wouldn't give one key access to everything just to save a minute. The morning pass doesn't need to send replies, and the bug filing job doesn't need to handle refunds.
I'd read the drafts for a week before considering automatic sending. The widget agent is a separate decision. I'd set up the docs before turning that on. Otherwise, it has to guess what your refund policy is, and it can sound very sure of itself.
Let it spot gaps in your docs
The inbox is also a useful way to see where your help center falls short. If the same question keeps coming in and there's no clear answer in your docs, Claude can flag the gap and suggest a new article.
It can draft the article, too. And if you give it the right API access, it can publish it to your help center through Fernand. I'd still review the first few. But the useful part is that the gap doesn't have to sit in your inbox as an unanswered question until someone remembers to write about it.
A normal morning
Five conversations came in overnight. A light day, but a pretty typical one.
One is SEO spam. Claude marks it as spam. I never have to open it.
Two are questions the help center already answers. Claude drafts a reply to each, in our voice, without promising a date or claiming it has issued a refund.
One is a bug I've heard about before. Claude leaves a private note, or files a Linear issue if I've given it that permission. When the issue closes, the conversation reopens so we remember to follow up with the customer. I've forgotten that step when working through the inbox myself. Claude doesn't.
The last one is about a refund. It gets tagged and stays in the inbox for me. Claude has sorted the pile. I still decide what to say.
What I leave alone
I don't ask Claude to make up a refund policy. If the help center doesn't say what the policy is, the draft shouldn't either. That's why billing stays out of the morning pass. I'm not planning to remove that instruction "just to see what happens."
I also wouldn't let an agent run through hundreds of conversations on a help desk that charges for each resolution. Fin charges $0.99 per resolution. Help Scout charges $0.75. I looked at the numbers in what AI resolution actually costs, and wrote about why I'm wary of that pricing in why per resolution pricing is a scam.
Fernand's API is included. It has a rate limit of 20 reads and 5 writes per minute, which is plenty for a pass every few minutes. Claude looking through the inbox doesn't create a second bill. I still find people are surprised by that. I understand why.
Two jobs, not one
If you already have an assistant in your product, Claude Code can help build a handoff into Fernand using your own components. We call that headless customer support. It's a different job for a different afternoon.
One agent builds the experience your customers see. The other helps with the conversations that come in. I wouldn't start both on a Tuesday.
Start with the morning pass. Give it a key that lets it draft, but not send. Read what it writes before sending anything. If the drafts are off, your docs may need work. In my experience, the prompt is rarely the first place to look.
Fernand is free to start, with 100 conversations a month and email included. Connect the inbox first. Then bring Claude in. It's hard for an agent to help with conversations it can't read.