Research-backed calculator guide

AI Email Triage Cost per Request: Classify, Extract, Route and Resolve

Estimate AI email triage cost per request by separating classification, extraction, routing, resolution, review, retries, and operational overhead.

What counts as an email triage request?

Define the unit before choosing a model. In this guide, one request is one inbound email plus the thread context and attachments you actually send to the workflow. The workflow may perform several actions:

1. Classify: identify intent, urgency, customer type, or policy category.
2. Extract: return structured fields such as order ID, invoice number, deadline, or requested service.
3. Route: choose a queue, owner, SLA, or next tool.
4. Resolve: draft an answer, create a ticket, retrieve a policy, or trigger a safe automated step.
5. Review or escalate: send uncertain, sensitive, high-value, or exceptional cases to a person.

A simple inbox label is not the same unit as a correctly routed request. A label can be technically correct while the message still reaches the wrong team, lacks a required field, or waits in a queue. If your business question is “what does AI triage cost per email?” use model cost. If it is “what does it cost per correctly routed request?” include errors, review, retries, and the systems that enforce the route.

The pages appearing for this query commonly describe these workflow actions, but they use different units. Tool roundups quote seats or subscriptions. Tutorials quote a monthly automation stack. Case studies emphasize turnaround time. Those are useful perspectives, but they cannot be compared until volume, scope, and outcome are aligned.

The cost formula

Track a request-level ledger with these fields:

• model_cost: input tokens and output tokens for every model call.
• tool_cost: paid search, OCR, ticketing, storage, or automation calls attributable to the request.
• review_cost: the probability of human review multiplied by review minutes and loaded hourly cost.
• retry_cost: expected duplicate or corrective calls after a timeout, malformed output, or failed tool action.
• exception_cost: expected cost of cases that need specialist handling, rework, delay, or a second system.

For a model call, the basic calculation is:

call cost = (input tokens × input price + output tokens × output price) / 1,000,000

Use the price card for the exact model and service tier. OpenAI’s API pricing page separates input, cached input, and output tokens and also documents service options such as Batch pricing; the page states that API usage is billed separately from ChatGPT subscriptions. ^[1] Anthropic likewise lists base input, cache-write, cache-hit, and output rates, and explains that prompt caching reuses repeated prompt content at a lower input cost. ^[2] Google’s Gemini pricing page separates free and paid tiers, token prices, context caching, and batch processing. ^[3]

Do not assume that a tool call is free because the language-model call is cheap. A ticket creation, OCR pass, web search, vector lookup, or paid automation operation may have its own meter. Likewise, cached context can reduce input cost, but only if the workflow really reuses the same content under the provider’s caching rules.

Worked cost example

The following is an original planning scenario, not a quote from OpenAI or any packaged email product. It uses the prices visible on OpenAI’s public pricing page on 2026-11-02: GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens, and GPT-6.1 Sol at $2.00 per million input tokens and $10.00 per million output tokens. Rates, model names, availability, taxes, regional arrangements, and service tiers can change; re-check the vendor page before budgeting.

Assume a 300-word message, a short thread context, no attachment OCR, JSON output for intermediate steps, and a routine reply for most requests. The workflow sends a request through three Luna calls, then sends the resolution step to Luna for 90% of cases and Sol for 10% of cases that meet an escalation rule.

The expected model total is therefore about 0.1364 cents per request, before review, tools, and failure handling. At 120 requests per day and 30 days, that is about $4.91 per month in raw model spend under these assumptions. At 300 requests per day, it is about $12.28 per month. Those are estimates, not promises of what any vendor plan will charge.

Now add the operational allowances. Suppose 1.5% of requests enter a human review queue. If review takes 45 seconds and the loaded labor rate is assumed to be $60 per hour, review costs 1.5% × $0.75 = $0.01125 per request. Add a 3% retry allowance to the model total: 3% × $0.001364 = $0.000041. The result is:

$0.001364 + $0.01125 + $0.000041 = $0.012655, or approximately 1.27 cents per request.

At 120 requests per day, this fully loaded planning estimate is approximately $45.56 per 30-day month, before fixed platform subscriptions and exception-specific losses. The review assumption drives the result. If review takes two minutes, or if 10% of traffic needs a specialist, the effective cost rises sharply even though token prices have not changed.

How to make the estimate fit your inbox

Start with a sample of at least 200 recent messages. Record message tokens, thread tokens, attachment text, output tokens, number of model calls, tool calls, human-review rate, retry rate, and the final outcome. Keep separate samples for routine, urgent, high-value, ambiguous, and attachment-heavy mail. A single average hides the long tail that creates cost and risk.

Next, decide which steps need a capable model. Classification and extraction often have compact outputs. Resolution can require more context and a more capable model because the answer has to be accurate, policy-compliant, and safe to send. A tiered policy is usually easier to audit than sending every step to the most expensive model. It is also easier to change when your evaluation set shows that one category is often confused.

Count context, not just the visible email body. Repeated system instructions, customer history, signatures, policy documents, and previous messages all consume input tokens unless your implementation trims or caches them. Anthropic’s pricing documentation explicitly describes cache writes and cache reads; OpenAI and Google also expose cached-input or context-caching concepts on their pricing pages. ^{[1] [2] [3]} Do not apply a caching discount unless you have verified that your request pattern qualifies.

Then measure “correctly routed.” A label accuracy score is not enough. Check whether the selected owner was right, the extracted fields were complete, the SLA was appropriate, and the request reached the intended system. For high-risk mail, let the model classify, extract, and draft, but require approval before sending or changing records. A review queue is a cost line, but it is also a control against wrong routing, unsafe automation, and prompt injection hidden in message content.

Finally, reconcile the estimate to the invoice. Export provider usage and join it to a request ID, workflow version, model, and outcome. The AI API invoice reconciliation calculator is useful when estimated tokens and billed usage diverge. For a broader monthly view, use the AI Agent Cost Calculator, and use the AI token cost calculator when you need to validate token assumptions independently.

What this estimate excludes

This example excludes email-provider licensing, help-desk seats, automation platform subscriptions, mailbox polling, storage, OCR, embeddings, retrieval, paid search, taxes, data-transfer charges, implementation, monitoring, evaluation, support, and the business cost of a wrong or late route. It also assumes text-only input and does not price attachments.

It does not claim that direct API usage is always cheaper than a packaged assistant. A packaged plan may include connectors, permissions, audit logs, retries, support, and a user interface. Compare like with like: raw model spend against raw model spend, or fully loaded cost against fully loaded cost. The public result set for this query shows why this distinction matters: subscription and seat figures are not interchangeable with token economics.

The figures are also not an ROI calculation. To estimate ROI, measure baseline review minutes, first-response time, reassignments, missed SLAs, and the value of prevented errors for your own inbox. Do not multiply a vendor’s accuracy or time-saved claim by your volume without testing it on representative messages.

Frequently asked questions

What is the cheapest way to calculate AI email triage cost per request?

Begin with the provider’s current input and output prices, then count the tokens and model calls used by one real workflow. Multiply each call, add tool charges, and include expected review and retries. A token-only number is a useful lower bound, not the cost of a correctly routed request.

Does classification cost more than drafting a reply?

Usually, drafting costs more because it produces substantially more output and often carries more thread context. The exact relationship depends on the prompt, output length, model, and whether the workflow runs additional calls. Measure both rather than assuming that a short visible email means a short billed request.

How should I price human review?

Use the review probability, average review time, and loaded hourly rate: review rate × minutes ÷ 60 × hourly rate. Include payroll burden or contractor overhead when appropriate. If reviewers handle only exceptions, measure exception mix separately instead of spreading the same average across every request.

Are packaged email-AI plans comparable with API cost per email?

Not directly. A packaged plan may charge per seat, per resolution, or through credits while bundling integrations, storage, support, and controls. API cost is only one component. Ask for the plan’s usage unit and compare it with your own observed request volume and outcome definition.

Should every triage step use the cheapest model?

No. Use a small model where a compact structured decision is sufficient, then test a stronger model for ambiguous or high-impact cases. The right policy minimizes expected total cost while meeting your accuracy and safety threshold. A model that creates more retries or reviews can be more expensive in production despite a lower list price.

What is a sensible first measurement before launch?

Label a representative set of at least 200 recent messages. For each, record the correct category, owner, required fields, urgency, expected action, and whether a human would approve automation. Use that set to measure routing accuracy, review rate, retries, and token use before you automate live mail.