Intent Classification

Why it matters for outbound teams

Raw reply volume is not actionable on its own; a team needs to know which replies are worth a rep's time and which ones are noise. Accurate intent classification is what makes positive reply rate measurable and lets a team prioritize the handful of interested replies buried in a much larger pile of auto-replies and rejections. Manual classification does not scale past a small volume of replies, which is why most reply management tools automate it.

How intent classification works

There are three common ways to decide what a reply means.

• Manual. A person reads the reply and picks a label. This is accurate and flexible, and it does not scale past a small volume.
• Rules. Keywords and patterns trigger a label, for example "out of office" or "unsubscribe". Rules are fast and predictable, and they only catch what you wrote them for.
• Model based. A language model reads the whole message, including tone and context, and picks the closest label. It handles the many ways people phrase the same idea, such as "send me more info", "sounds interesting" and "can you share pricing?".

Most teams combine them: rules for obvious system messages, a model for the rest, and a person reviewing a sample. The output of the process is a label, covered in reply labeling.

What a good taxonomy looks like

Intent classes should be few, distinct and tied to actions. A practical set covers interested, meeting request, question, referral, not now, wrong person, not interested, unsubscribe, out of office and bounce. Add your own where your market needs it, such as "already has a vendor".

Master Inbox applies AI labeling to every email and LinkedIn reply as it lands, with thirteen default labels or your own taxonomy, running on your own OpenAI, Anthropic or OpenRouter key. The labels then trigger workflows, such as a Slack card with an approve button, a HubSpot sync, or stopping the sequence when a meeting is booked.

Keep the set stable once reps rely on it. Renaming labels every month breaks saved views and reports. Add new classes, retire unused ones, and note the date of each change so past reports still make sense.

A worked example

This is an illustrative example with round numbers, not measured data. Consider three replies to the same campaign.

• "Thanks, can you send over pricing and a case study?" is interested. It goes to a rep right away.
• "I am out until the 14th, back then." is out of office. It is parked and the sequence resumes.
• "Not me, try Sam in finance." is a referral. A new contact is created.

Keyword rules would catch the second one easily, and might miss the first, since it has no obvious trigger word. A model reads the intent in all three. Across 200 replies, that is the difference between a rep reading every message and reading only the few that matter. It also feeds positive reply rate, which depends on accurate labels.

Common mistakes

• Labels without definitions. Write a sentence for each class, with an example reply.
• Too many classes. Fine distinctions confuse both people and models. Merge classes that lead to the same action.
• Treating mixed replies as one thing. "Not now, but send info" is two intents. Choose a primary label and use a rule for the tiebreak.
• No review. Check a weekly sample and correct errors, so accuracy holds as language drifts.
• Ignoring other languages. If you sell across markets, test classification on real replies in each language you use.

Part of reply management is deciding who reviews edge cases.

Frequently asked questions

What is intent classification in cold email?

It is the process of deciding what a reply means, such as interested, not now, referral, out of office or unsubscribe. The result is a label that routes the reply to the right person or action and feeds metrics like positive reply rate.

Is intent classification accurate?

Accuracy depends on clear label definitions, good examples and regular review. Language models handle varied phrasing well, but they still make mistakes, so teams check a sample every week and correct errors to keep results dependable.

Can I define my own intent categories?

Yes, with most modern tools. Start with the common set, then add classes your market needs. Give each a short definition and a next action, and test the set on a few hundred past replies before using it on live traffic.

What is the difference between intent classification and sentiment analysis?

Sentiment measures tone, such as positive or negative. Intent identifies what the sender wants to happen next. A polite reply can say no, and a blunt one can ask for a meeting, so intent is the more useful signal for sales.

← All glossary terms