AI email automation: triage and drafted replies

Which inboxes to automate first

The best candidates are shared inboxes with high volume and repetitive questions, where the answer lives in a system you already have:

• Orders and sales: trade customers asking for stock, prices or delivery dates. The AI recognises the request, looks up the data and drafts the reply.
• Customer service: ‘where is my order?’, address changes, returns. Sorted by type and urgency, with the order already identified.
• Accounts payable: supplier invoices, remittance queries and statement requests. Attachments go to AI document extraction and queries go to the right person.
• Purchasing: supplier confirmations and revised delivery dates, linked to the matching purchase order.
• Service desks and job enquiries: requests turned into a ticket or job with the key fields filled in.

Three levels of autonomy for email triage and replies

Not every email deserves the same treatment. We design the system so you choose the level for each type of email, and you can raise it later once you have seen how it performs. Most projects start at levels 1 and 2. Microsoft Graph and the Gmail API both let an application create a draft without sending it, so nothing leaves your mailbox unless a person, or a rule you approved, sends it.

What happens when the AI is not sure

Each classification carries a confidence score. If it falls below the threshold you set, the email goes to a manual review queue, exactly as it would have done before. Emails that mention cancellation, legal action or a complaint can be forced into review whatever the score.

Every decision is logged: how the email was classified, what was drafted, who approved it and what was changed. Those corrections are the data we use to adjust the rules. You also get a simple view of volume, response times and the most frequent topics, which often shows where your own processes or website could answer the question before anyone emails.

Copilot in Outlook, Gemini in Gmail or a bespoke assistant

If what you need is help writing emails, start with what your suite already includes. Copilot in Outlook drafts a message from a prompt and adjusts tone and length; Gemini's ‘Help me write’ in Gmail drafts and refines messages and replies. Both help the person who is writing. A shared inbox that is sorted before anyone opens it, with replies built from your ERP or CRM, is a different job:

Confidentiality and telling people when AI is involved

Emails are full of personal data and commercially sensitive information, so before starting we agree in writing which inboxes are read, what is stored, for how long and which AI providers are involved. Whoever processes emails on your behalf acts as a processor under UK GDPR, and the ICO explains what the contract between you must cover. Our approach to training, data location and AI providers is set out in full on the private ChatGPT for business page.

If you reply to customers in the EU with level 3 automatic replies, bear in mind the EU AI Act: from 2 August 2026, Article 50 requires that people are told they are dealing with an AI system unless it is obvious. We leave this ready in the design; how it applies to your business is a question for your adviser. This is not legal advice.

How we would run it and what drives cost

We start with one inbox and two or three email types, at level 1 or 2, using your real emails. You see a working demo every two weeks, you get one fixed price for the project and the code is yours, in a repository in your company's name. The rest is on how we work. What moves the cost most:

• Number of inboxes and email types to classify.
• Systems the replies draw on: answering from the ERP or CRM means building that connection.
• Email platform: Microsoft 365, Google Workspace or something else.
• Rules and exceptions per customer, product or department.
• Volume: AI model usage is paid per email processed; the market references are in the guide on what AI costs in a business.