Support Email Intent Dataset

A clean, ready-to-use synthetic dataset of 420 client and customer-support emails, each hand-modeled on the messages agencies and consultancies actually receive - billing questions, refund and cancellation requests, outages, security concerns, scope changes, deliverable status checks, and more. Every row is labeled with intent (20 categories), urgency (low/medium/high/critical), sentiment, the affected product area, a recommended routing action, and an SLA target in hours. Built for teams standing up an AI support triage or auto-responder: use it as few-shot examples, an evaluation set, or a fine-tuning / classification starter. The label schema is designed around real routing decisions (route_to_billing, escalate_refund_review, escalate_incident, route_to_account_manager, and so on), so a model trained on it produces outputs you can actually act on. 100% synthetic. No real people, companies, domains, or PII - every name and company is invented and generic, safe to load into any environment or share with a client. Balanced across all 20 intents (21 rows each) so no class is starved. UTF-8 CSV with a header row and RFC-4180 quoting (bodies contain commas and are properly escaped). What's inside: - 420 rows, 20 intent classes (balanced): billing_question, refund_request, cancellation, technical_issue, bug_report, feature_request, how_to, account_access, onboarding, complaint, praise, sales_inquiry, upgrade_request, integration_help, outage_report, data_export, security_concern, contract_renewal, scope_change, deliverable_status. - 14 columns: id, date, channel, customer_name, company, plan, subject, body, intent, urgency, sentiment, product_area, suggested_action, sla_hours. - Coherent, natural-language subjects and bodies (not lorem) that read like real inbound email across four channels (email, web form, chat transcript, forwarded email). - Urgency and sentiment distributions tuned per intent (e.g. outages skew critical, praise skews positive) so the data reflects real support reality. - A recommended routing action and SLA target per row, so you can train triage + prioritization, not just classification. License: standard Toolbay license - use it in your own products, evals, and client work. First-party Toolbay dataset.