AI reply management for cold email that keeps you in charge

What AI reply management for cold email does

A cold email sequence produces a mix of answers. A few are real conversations. The rest are out-of-office notices, bounces, ticket receipts, "wrong person" replies and unsubscribe requests. In a regular inbox they look alike until someone opens them, so a hot reply sits unread, an opt-out waits until Monday and a sequence keeps writing to someone on holiday.

AI reply management does three jobs, and FoxKnock's AI sales inbox does all three while leaving the sending to you (with one opt-in exception, covered below):

• Triage. Every reply arrives with a label, a confidence score, a one-sentence summary and a suggested next step, not just a subject line.
• Bookkeeping. An unsubscribe is suppressed, an out-of-office notice holds the sequence and a referral becomes a contact and a task.
• Drafting. A reply that needs an answer gets a draft next to the thread, written from your own notes.

That replaces skimming a Gmail label and updating a spreadsheet by hand.

How cold email reply classification works

Most replies that match something you sent are read by Claude, by Anthropic, together with the email they answer. The text of both is sent to Anthropic for this. Claude picks one of 13 classes.

Rules label a few kinds of mail instead: delivery failure and spam-complaint notices, most automatic replies that carry auto-response headers, and explicit unsubscribe requests from senders you have no contact for.

Here is how a few illustrative replies, not real ones, would be sorted:

• "How does this connect to our CRM?" is a Question, so it waits under Needs reply with a draft.
• "Can we talk Tuesday afternoon?" is a Meeting request, and hot replies notify your team.
• "Not this quarter, try me in January." is Not now: a short, polite draft for you to send.
• "I am out until next Monday." is Out of office: no draft, and the sequence holds.
• "Wrong person, but Dana Lee in operations handles this." is a Referral: Dana is added as a contact at the same company, with a task.

Every label from the AI carries a confidence score. Below 0.6 the reply is marked Needs review and appears under its own filter, so a shaky guess does not trigger the automatic actions below. Bounces are the exception: they are read from the delivery failure notice itself.

Corrections feed a classifier precision figure on the Analytics page: the share of AI labels nobody had to correct. For what each class means and how to answer it, see cold email reply classes.

What it does by itself, and what stays with you

A reply from a person also stops that contact's sequence at default settings. Beyond that, the inbox takes a short, fixed list of actions without asking. Each is recorded on the thread and needs a confidence of at least 0.6 unless noted:

• Out of office (on by default). Holds the sequence until one business day after the return date in the notice, or for a week when the notice gives none.
• Unsubscribe (on by default). Adds the address to the suppression list. If you switch this off, a task asks a person to confirm instead. See the built-in GDPR tools for how suppression works.
• Referral (no switch). When the reply names a colleague or gives an address, adds that person as a contact at the same company, or finds them in your CRM, and creates a task to reach out that names who made the introduction.
• Booking-link reply (off by default). Answers a meeting request with a short fixed message that holds only your own booking URL. It needs a saved booking URL, 0.9 confidence by default (you can change it), a request under an hour old, sending not paused, and at most one send per conversation per week. If a condition fails, nothing is sent and the reply waits for you with a draft. A request that is too old, or sending being paused, also creates a task.

The booking-link message is the only reply FoxKnock can send without your approval, and only if you switch it on. You send every other reply yourself. Outreach emails have their own gate: see approval-first outreach.

How bounces are handled

Bounces are verified first. A notice counts only when it looks like it came from a mail system, quotes the Message-ID of an email you sent and names the address you wrote to, so a forged "undelivered mail" message is not enough to suppress a contact.

A verified hard bounce marks the address invalid, suppresses it, stops its sequences and refunds the email-find credits FoxKnock charged to find it, if it charged any (10 credits at default prices, once). Imported addresses were never charged. Soft bounces are only logged.

Spam-complaint reports suppress the address, stop its sequences and count against mailbox health.

AI-drafted replies from your own notes

When a reply needs an answer, FoxKnock drafts one and shows it next to the thread. It writes from:

• your knowledge base: FAQs, objection notes, case studies, offer, policies and snippets
• your playbook brief and the recent thread
• when there is one, the latest research summary on that company

Policies are treated as hard rules, such as "never promise a go-live date".

A line under the editor lists the knowledge items it used. If your notes do not cover a question, the draft is instructed to say you will confirm and come back, not to guess. You edit it and send it from your own mailbox, and sending checks the suppression list and the kill switch again.

Unlike outreach emails, reply drafts are not run through the claim checker that blocks unsupported claims in cold emails. A thin knowledge base makes a thin draft, so read before you send.

How to handle cold email replies with an AI sales inbox

1. Connect a mailbox that can sync replies. Google Workspace or any SMTP and IMAP mail server connects with an app password (see cold email sending). Microsoft 365 is limited: Microsoft no longer accepts passwords (Basic authentication) for IMAP in Exchange Online and is retiring it for SMTP AUTH, so a Microsoft 365 mailbox may not connect or may not sync replies. FoxKnock does not offer Microsoft sign-in.
2. Fill the knowledge base. Onboarding proposes FAQs and "never claim" policies from your website. Treat them as a first draft, and add the objections you hear most.
3. Work Needs reply, then Needs review. Needs reply holds unanswered interested, meeting-request, question, objection and referral replies, plus other unread replies that are not automatic. Needs review holds the shaky labels. Correct wrong labels as you go.
4. Keep the booking-link reply off at first. Watch the meeting-request labels for a few weeks. If they hold up, switch it on in the Reply agent card under Settings, AI, where you also save your booking link.
5. Mark meetings yourself. Click Mark meeting booked on the reply. It creates the meeting, stops the sequence and fires the meeting.booked webhook. Cost per meeting in Analytics counts only meetings you mark.
6. Know what it reads. FoxKnock fetches new mail from the INBOX folder over IMAP in read-only mode and stores only mail that matches an email you sent or a contact in your CRM, plus explicit unsubscribe requests. Everything else is dropped without being stored.

What it costs at default prices

The AI inbox costs 0 credits at default prices: labels and suggested replies are free. A daily AI budget bounds them, shared with the workspace's other AI features: a dollar cap per UTC day on FoxKnock's own AI key. It defaults to $5 on Starter and $15 on Growth, and a workspace can lower it. With your own Anthropic key the budget no longer applies unless you set one.

When the budget runs out, a keyword fallback labels replies and marks them Needs review, no drafts are written until it resets at 00:00 UTC, and an explicit unsubscribe request is still honored.

The pricing page lists the AI inbox from the Starter plan ($49 a month, 5,000 credits) upward. Credit prices can change, so check it for current numbers.

To try FoxKnock on your own market, sign up and use the free credits to find and research companies first.