AI email triage for a small professional services firm

AI email triage reads each incoming message, decides what it is and who should handle it, and either routes it or drafts a reply for a person to check. For a firm of 5 to 100 people, the shared inbox (info@, office@, claims@) is usually the right place to start, because one workflow serves the whole team and the rules are easy to agree.

In most firms I talk to, nobody owns the shared inbox. Everyone glances at it, the person who sits nearest answers the easy ones, and the hard ones wait. Triage does not replace the people who answer; it makes sure the right person sees the right message the same morning.

What triage actually does

A triage workflow has four steps, and each can be switched on separately:

1. Classify. Is this a new enquiry, a client question on an existing matter, a document delivery, an invoice, a court or authority letter, a claim, or noise?
2. Extract. Pull out the client or matter name, any deadline, and what is being asked.
3. Route. Assign it to a person or a queue, or create a task in the tool the team already uses.
4. Draft. For routine questions, prepare a reply from your templates or the client file for someone to check and send.

Most firms get the largest gain from the first three. Drafting is useful, but it is the step that needs the most care.

Two ways to set it up

A UK guide published this year for small professional services firms on Microsoft 365 reaches the same conclusion: per-user Copilot for personal inboxes, a single classification workflow for shared ones, with a default route for anything unclassified so nothing is lost. [3] That matches what I would recommend for a Swiss firm, with the data protection points below added.

What to leave with people

• Sending. A person sends every reply that goes to a client, court, authority or insurer. AI drafts; people send.
• Deadlines. AI can spot a likely deadline and flag it. The responsible person confirms it and puts it in the diary.
• Anything ambiguous. If the workflow is not confident, the message goes to a named person, not to a guess.
• Complaints and sensitive matters. Route them straight to a partner or owner, untouched.

The data protection points for a Swiss firm

An email triage workflow reads client correspondence, so the Swiss data protection act applies to it in full. The FDPIC has confirmed the Act applies directly to AI-supported processing. [5] In practice:

• The AI provider is a processor. You need a contract with it, it must not breach any duty of confidentiality you owe, you must be satisfied it can keep the data secure, and it may only use sub-processors with your approval. [4]
• Use a business service, not a personal account. Microsoft's work-account Copilot Chat, for example, applies its enterprise data protection terms with Microsoft acting as processor, and prompts and responses are logged in your own tenant for audit. [1]
• Read the minimum. Classify on subject and body text; only open attachments where the workflow needs them.
• Tell clients. A line in your privacy notice about AI-assisted handling of correspondence covers the transparency point.
• For lawyers, professional secrecy is the higher bar. See can a Swiss law firm use ChatGPT or Copilot with client data?

How to start without a big project

1. Take a week of the shared inbox and sort it by hand into six to ten categories. If people disagree about the categories, settle that first; AI will not.
2. Name an owner for each category and a fallback person for everything else.
3. Run the workflow alongside people for two weeks, labelling but not routing, and check where it was wrong.
4. Switch on routing, then drafting for the two or three most routine question types.

That is a typical first project under my AI Workflow Automation service, from CHF 1,500. If you are not sure the inbox is the right first step, the AI Opportunity Audit answers that. Sector detail: law firms, fiduciaries and accountants, insurance brokers.