Customer service20:08
AI answers 8 types of emails from contact@ on its own. A human approves the rest
On the contact@ inbox, AI can only send repeatable confirmations on its own. We present the matrix, confidence thresholds, and a test on 300 emails.

Listen to the article summary
Synthetic voice.This guide comes from our direct practice, not a tailored legal opinion. The higher the stakes, the more you need to review it with counsel who knows your contracts. A 40-person e-commerce company, one person managing the contact@ inbox. 50–70 emails arrive daily: order status, invoice requests, delivery questions, occasional complaints. Routine messages take 3 hours a day, yet by the fifth complaint, the owner refuses to let anyone reply automatically without oversight. This provides a solid starting point: automate only a narrow scope with strict boundaries.

The answer to the title question is: yes, but only for 8 low-risk case types, not for the entire inbox. The system relies on a decision matrix: AI autonomously handles status confirmations, complaint acknowledgments, delivery updates, business hours, invoice requests, password resets, and basic FAQs. Complex complaints requiring decisions, contract withdrawals, warranties, statutory rights, disputes, official letters, sensitive data, escalations, and key accounts escalate to human agents. Automated decisions on refunds or rejecting complaints carry a different level of risk: Article 22 GDPR grants individuals the right to human intervention. Transparency requirements also apply: users must be notified that they are interacting with AI at the latest upon first contact, mandatory from August 2, 2026 under the AI Act.
What you can do this week
Monday: an operator extracts 300 real, anonymized emails from the past 6 months and tags 10 categories. Time: 3 hours. Tuesday: the business owner defines the matrix of 8 automated categories and a blocklist of trigger phrases: complaint, withdrawal, refund, ombudsman, court, penalty. These topics and strong negative sentiment block automated dispatch regardless of confidence scores. Time: 2 hours. Wednesday: we implement two thresholds. Intent classification with at least 0.90–0.95 confidence routes to auto-send, the 0.75–0.90 range generates drafts for human approval, and scores below 0.75 route directly to an agent. Thursday: shadow mode activates on incoming emails without sending. Friday: the operator reviews the 5 worst misclassifications and refines the rules. Total cost: 9–10 hours of internal work, zero new software licenses.
What to avoid doing right away
Do not activate automated dispatch for complaints and contract cancellations, even at 0.98 system confidence. This category produces legal effects, and consumer complaints require resolution within 14 calendar days under the Consumer Rights Act. Do not buy specialized software for 10,000–20,000 PLN before completing a test run on 300 emails. Do not integrate ERP systems, connect customer databases, or train models on the entire raw inbox. Never deploy automation without a kill switch that cuts off automated sending in under 15 minutes in case of campaign errors, data leaks, model failures, customer disputes, or regulatory audits.

How to verify it works after two weeks
After two weeks in shadow mode, count incorrectly resolved complaints per 100 messages and automated drafts rolled back by a human before sending. In practice, the acceptable threshold is 1–2% critical errors, with zero tolerance for the complaint or withdrawal category. If you encounter more than 2 falsely sent decisions per 100 emails, or if any complaint was rejected without human review, roll the automation back to draft mode, fix the error, test the next 200 emails, and precede the restart with a test message. Only a score below the threshold qualifies for automated sending across 8 case types.
HEXART implements this directly in your inbox, complete with correspondence logging, a category matrix, and a STOP switch. If a week of testing proves that narrow automation works, you decide whether to launch; if not, you have an established human oversight procedure and zero money spent.
Source materials
Sources consulted during research. The text above is our own; these third parties are not responsible for its content and have not authorised it.
- Aktualności - UODO
- Frequently Asked Questions | AI Act Service Desk
- Ustawa z dnia 3 lipca 2026 r. o systemach sztucznej inteligencji
- Ustawa z dnia 3 lipca 2026 r. o systemach sztucznej inteligencji
- Ustawa o systemach AI - bezpieczny rozwój sztucznej ...
- AI Act już obowiązuje. Jak korzystać z AI w świetle nowych ...
