AI Email Triage: What It Handles and Hands Off

AI email triage rarely fails because the model can’t read an email. It fails because nobody decided which emails it may answer, which it must hand over, and who picks up what it hands over. Then one bad reply to an important client and the team stops trusting the whole setup. Getting this right is mostly a design problem, not a technology problem.

We build email triage and other Claude workflows through our AI integration service, so weigh that. Every fact below about Anthropic, the Gmail API and Microsoft Graph was checked against their own documentation in October 2026.

Short answer: AI email triage reads each incoming email, gives it a category and a priority, then drafts a reply, files it, or hands it to a person with a note on why. Let it handle routine email: status questions, acknowledgements, invoices, newsletters. Send complaints, contracts, key clients, requests that need action and anything the rules can’t place to a person, with no AI draft. Test those rules on your real inbox before going live. On Anthropic’s published prices, the model cost in our worked example is about $0.35 to $7 per 1,000 emails. The build and the rules are where the work is.

What the Vendors Document

As of October 2026, from Anthropic, Google and Microsoft documentation:

Sources: Anthropic pricing, Batch processing, Structured outputs, Anthropic on model training, Anthropic data retention, Gmail push notifications, Gmail API scopes, Microsoft Graph change notifications, Outlook categories.

What AI Email Triage Actually Does

AI email triage is not magic inbox management. It is a classification and routing layer that sits between your mailbox and your team.

At its core, it does three things: reads each incoming email, assigns it a category and a priority, and either drafts a reply or flags it for a person. The AI doesn’t know your calendar, your CRM or your client history unless you pass that context in with the email.

How It Connects to Gmail and Outlook

Triage that runs on its own needs to know when mail arrives. Gmail sends push notifications through Google Cloud Pub/Sub after your app calls watch on the mailbox. Google says delivery is typically within a few seconds but can occasionally be delayed or dropped, so a working build also checks the mailbox history on a schedule. The watch must be renewed at least every 7 days or the notifications stop (source). Outlook works the same way: your app subscribes to a user’s Inbox through Microsoft Graph and renews the subscription before it expires (source).

Writing back is where triage shows up for your team. In Gmail, one message can carry several labels (source), and AI-written replies can be saved as drafts until someone sends them (source). In Outlook, your code can apply categories from the user’s master list (source), set importance to low, normal or high, and create a draft reply for a person to edit or send (source).

Ask for the narrowest access that does the job. Google classes gmail.readonly, gmail.compose and gmail.modify as restricted scopes, and an app that stores or transmits restricted-scope data on its servers must go through a security assessment (source). Settle that early.

The Four Core Functions

Classification assigns each email to a category you define. For a B2B service business, the categories might be new enquiry, existing client request, vendor or supplier, invoice, newsletter or promotion, and internal. A person defines the categories. The AI matches against the list you give it. Anthropic’s ticket routing guide makes the same point: start from how your team handles these messages today (source).

Priority scoring ranks emails within categories. The AI can pick up signals such as “the client mentions a deadline”, “the email refers to a contract” or “this sender has written three times this week”, and weight priority accordingly. The scoring rules are still yours to define.

Draft generation writes replies for routine categories. A quote request gets an acknowledgement with an expected response time. An order status question gets a link to tracking. A support request gets a confirmation with a reference. Drafts save the most time when the category is clean and the emails are predictable. Ambiguous or multi-topic emails produce drafts that still need heavy editing.

Escalation flagging catches the emails the rules can’t place and routes them to a person, with a note explaining why. This is the most important function and the easiest one to skip.

For the output, ask Claude for a fixed structure: category, priority, a short reason and an optional draft. With structured outputs, the response is guaranteed to match the JSON schema you define, and listing the categories as an enum stops Claude inventing a new one (source).

What AI Handles Well

Routine volume is where AI triage earns its keep. These categories are predictable enough to automate, once you’ve tested them on your own email.

Order and booking status queries. “Where is my order?” and “Can I change my booking?” have a small set of answers. For an online store, the AI can identify the query type, pull out the order number and draft a reply that points the customer to their tracking page, or flag the email if there’s no order number. This works best with clean order data and consistent emails. Messy order data, frequent exceptions or emails that mix an order question with a complaint will need more human handling.

Acknowledgements for new enquiries. A prospective client sends an enquiry on a Friday evening. The system sends a well-toned acknowledgement, confirms receipt and says when they’ll hear back. A person picks up the enquiry on Monday. This keeps the lead warm. It doesn’t replace the real reply, and it only works if the tone suits your audience.

Invoice and vendor routing. Supplier invoices, renewal notices and vendor emails don’t need someone to read them first. They need to land in the right place. Triage can label them and route them to the right folder or person.

Newsletter and promotional filtering. Newsletters and promotional emails can be labelled and moved out of the main view without anyone reading them. Not glamorous, but it clears noise from the inbox your team actually works in.

What AI Must Pass to a Human

The boundary matters more than the capability. These categories should go to a person every time, with no AI draft attached.

Unhappy clients. An AI can pick up negative tone, but it can’t weigh your history with that client, what the relationship is worth or the right way to calm things down. Any email flagged as a complaint, a dispute or clearly unhappy goes to a person with no draft. A bad AI-drafted reply to a frustrated client does more harm than a slower human one.

Contractual or legal language. An email that mentions a contract, an SLA, liability or a formal notice needs a person to read it before anyone replies. This isn’t a technical limit. It’s a risk decision, and it belongs in the routing rules as a fixed rule.

Ambiguous high-value senders. If a sender is on a key accounts list you maintain, every email from them goes to a person, whatever it seems to be about. A casual check-in from a major client deserves a personal reply, not a template.

Requests that require system access. Triage can draft a reply to a password reset request, but it can’t reset the password. If the email needs someone to do something, not just answer, it routes to whoever owns that action.

Low-confidence classifications. Any email the rules can’t place, or that fits two categories at once, goes to a person with a note on what made it unclear. Don’t rely on the model’s own sense of how sure it is. Route on signals you can check, covered in the next section.

The Line Is a Design Decision, Not a Default

Triage tools don’t arrive with your escalation logic. The boundary between “AI handles” and “a person handles” has to be written as routing rules, and those rules have to be tested against your real email before going live.

A small agency with a handful of key client relationships needs a different boundary from an online store handling a steady stream of order questions. Same model, different rules, very different results.

This is why a quick no-code template rarely holds up on its own. The template handles the trigger-and-classify step. It doesn’t know your escalation rules, your key account list or what should happen when an email doesn’t fit.

Route on Rules You Can Check

Escalation rules work best when anyone on the team can check them. For example:

• The sender is on the key accounts list.
• The category is complaint, dispute, legal or contract.
• The email mentions a refund, cancellation or chargeback.
• A field the draft needs, such as an order number, is missing.
• The email doesn’t match any category, or matches two.
• The sender is writing for the third time about the same thread.

Each rule is a yes or no, and the escalation note says which one fired.

Anthropic’s ticket routing guide recommends evaluating a routing prompt against clear thresholds; its example sets 95% accuracy across 100 tests (source). Pick your own target before launch, and test on emails your team has already sorted by hand.

Real Cost and Real Output

What the API Costs

Model costs for email triage are low. Here’s our own worked example from Anthropic’s published prices (source). Assume each email uses about 2,000 input tokens (instructions, category list and the email itself) and 300 output tokens (the category, a reason and a short draft).

• Claude Haiku 5.5: 2,000 input tokens at $0.10 per million is $0.0002, and 300 output tokens at $0.50 per million is $0.00015. That’s about $0.00035 per email, or about $0.35 per 1,000 emails.
• Claude Sonnet 5.5: 2,000 input tokens at $2 per million is $0.004, and 300 output tokens at $10 per million is $0.003. That’s about $0.007 per email, or about $7 per 1,000 emails.

Your numbers depend on prompt length, email length and whether every email gets a draft. Anthropic’s ticket routing guide notes that Claude Haiku 5.5 suits routing because it’s the fastest and most cost-effective current model (source).

The Batch API halves the price, but batches can take up to 24 hours (source). That’s too slow for live triage. It suits one-off jobs, like classifying a year of old email to see what your inbox contains before you design the categories.

What the Build Involves

The build is the real cost, and it depends on your setup. A complete build connects to Gmail or Outlook, keeps the mailbox subscription renewed, applies your routing rules, writes labels or categories and saves drafts, and logs every escalation somewhere your team will see it, such as a shared Slack channel. After launch, the ongoing work is reviewing the escalation log and updating the rules when new edge cases appear.

Workflow tools that charge per task can be a sensible start. At higher email volumes, compare their per-task pricing with the cost of a direct API integration before you commit.

Building It Right Versus Building It Fast

For an online store handling order support, the categories might be: order status, return request, product question, shipping complaint and other. The rule for “shipping complaint” sends the email straight to a person with no draft, because a drafted reply that cites the wrong order turns a delivery problem into a trust problem. The rule for “order status” creates a draft that pulls the order number from the email and fills in a tracking link template.

That level of detail doesn’t come from a template. It comes from mapping your real email history before writing any configuration. If you want to know what your inbox actually contains before scoping a build, an audit against real data is the right first step.

What Happens to the Email Content

Email is full of personal data: names, addresses, order details, sometimes more. A few facts help. Anthropic says it doesn’t use inputs or outputs from its commercial products, including the API, to train models by default (source). For API users, it deletes inputs and outputs from its backend within 30 days, unless you use a service with longer retention such as the Files API, you have a zero data retention agreement, or it needs to keep them for Usage Policy enforcement or by law (source). Its Data Processing Addendum, with Standard Contractual Clauses, is part of its Commercial Terms. If you use Claude through another company’s platform, that platform’s terms apply instead (source).

Your own obligations remain. Under GDPR, personal data should be adequate, relevant and limited to what is necessary for the purpose, the data minimisation principle in Article 5(1)(c) (source). In practice, send Claude the email and the context the rules need, not the whole mailbox or the full customer record. That lowers your privacy exposure and your token costs.

For a complete integration that classifies, routes, drafts, escalates and logs, we scope the build before any commitment. If it’s built on your own API key, the prompts and routing rules are yours. Talk to us about what this would involve for your setup.

Related Posts

• Human review checkpoints in AI automation: what the person picking up an escalated email needs to see
• AI customer support escalation logic: designing the handoff to a human
• AI confidence calibration: why a model’s own certainty is a weak routing signal
• Google Workspace AI integration: the Gmail and Workspace workflows worth automating first

Frequently Asked Questions

What is the difference between AI email triage and an AI email assistant?

An AI email assistant is something you use by hand: you ask it to summarise or draft, and it responds. AI email triage is automated. It runs on every incoming email without anyone starting it. Triage is a workflow layer. An assistant is a productivity tool. They can work side by side, but they solve different problems.

How accurate is AI email classification in practice?

It depends on your categories and your email, so test it rather than trust a general figure. Clear, distinct categories classify well. The usual failures are emails that span two categories or use unusual phrasing. Before launch, run the system on emails your team has already sorted, measure how often it agrees against a target you set in advance, and keep an escalation path for the emails the rules can’t place.

Can AI triage handle emails in multiple languages?

Often, yes, but test your languages. Anthropic publishes Claude’s scores relative to English on a translated general knowledge benchmark, not an email task. For Claude Haiku 4.5, Spanish, French and German score 94–97% of its English result, Swahili 78% and Yoruba 53% (source). Anthropic also recommends stating the response language in the system prompt. Mixed-language emails and dialects need their own test cases.

Does the AI read the full email or just the subject line?

A well-built system passes the full email body to the model. Subject-only classification misses tone, urgency and content that contradicts the subject line. The full body uses more tokens, but on Anthropic’s prices the difference is small: in our worked example above, a full email with a short draft costs under a cent with Claude Haiku 5.5 or Claude Sonnet 5.5.

What happens to email content sent to the Claude API?

Anthropic says API inputs and outputs aren’t used to train its models by default, and it deletes them from its backend within 30 days, with exceptions such as the Files API, a zero data retention agreement, Usage Policy enforcement and legal requirements. Compliance is still your job: send only what the task needs, document the data flow and get legal advice if you handle health, legal or financial information.

How long does a proper AI email triage build take?

It depends on how complex your routing rules are, whether you connect to Gmail, Outlook or both, and whether classification needs CRM lookups, so we don’t quote a standard timeline. The stages are the same each time: map your real email into categories, write the escalation rules, test the prompt on emails you’ve already sorted, connect the mailbox with logging and error handling, then run it alongside your team until the results hold up.