AI Can Sort Your Inbox. Which Replies Still Need a Human?

Monday morning, 140 unread emails. Most are receipts, newsletters and app notifications. Somewhere in the pile are a refund request from an annoyed customer, a supplier asking you to confirm a price, and a prospect who wants a quote by Wednesday. AI email tools can now find most of the noise for you. The harder question is which of the remaining messages a person still has to read before anything goes out.

The short answer: let AI sort and summarise first, let it draft second, and let it send on its own only for low-risk messages where a wrong reply costs nothing. Complaints, anything about money, and anything that commits your business to a price, date or exception should always reach a human. The tools covered below fit that pattern already. Superhuman says its Auto Drafts are "never sent automatically", and Microsoft's own Copilot in Outlook FAQ tells users to "review, edit, and verify" anything Copilot creates.

What AI email triage actually does

Triage means deciding what each message is and what should happen next: reply, wait, file or ignore. Current tools do this in three ways.

• Priority flags. Microsoft's Prioritize my inbox in Outlook marks new Inbox mail as high, normal or low priority and shows a short reason. You give it plain-language instructions about what matters, such as "It's about a customer complaint". It only judges mail that arrives after you switch it on, and it skips mail that rules have already moved to other folders.
• Fixed categories. Superhuman's Email Assistant gives every new email one of six labels: Respond, Waiting, FYI, Notifications, Promotions or News. It can archive chosen types of Notifications, News and Promotions, but not Respond, Waiting or FYI mail.
• Summaries and to-do lists. Google's AI Inbox for Gmail pulls "suggested to-dos" and "topics to catch up on" out of your Primary tab. It is in beta, and it ignores attachments, delegated inboxes and encrypted mail.

Rule-based tools sit alongside these. SaneBox plans list features such as SaneLater, a Daily Digest, Reminders, and subject and domain filters. Plain Gmail filters can also act on sender, subject or keywords without any AI.

Categories that work for a small team

Built-in labels are a start, but they describe your workload, not your risk. A useful set for a solo founder or a team of two to ten adds a risk lens:

1. Noise: newsletters, promotions, routine notifications. Safe to file automatically.
2. Information: receipts, delivery updates, FYI copies. File, but keep searchable.
3. Routine questions: opening hours, where to find an invoice, how to book. Good candidates for drafts.
4. Sales and quotes: anything that asks what something costs or when it can be delivered. Drafts allowed, sending only by a person.
5. Complaints and refunds: unhappy customers, chargebacks, cancellations. Always escalate to a named person.
6. Legal, security and money: contracts, payment-detail changes, sign-in alerts, data requests. Always a human, ideally the owner.

Turning these categories into reliable rules and instructions for your own mail is where most of the setup effort goes, and it differs by tool.

Where a human must stay in the loop

"Human in the loop" means a person approves an action before it takes effect. For email, the action that matters is sending. An AI that files a newsletter in the wrong folder wastes a minute. An AI that tells a customer "yes, we'll refund that" or quotes last year's price can create an obligation you then have to honour or argue about.

Three kinds of mail deserve a hard stop.

Complaints

Route these to a named person with a response target, for example "same working day". Use the AI summary to save reading time, but have a person write or approve the reply. A tone-matched draft can sound confident about facts it has not checked.

Pricing and commitments

Any reply that names a price, discount, delivery date, refund or exception is a commitment. Drafting tools build replies from your past emails and the thread. Superhuman says Auto Drafts use "how you have communicated with that recipient in the past". That is helpful for tone and risky for numbers: an old quote in your sent folder is not your current price list. Check figures against a single current source before sending.

Security and payment changes

Requests to change bank details or reset access can be fraud attempts. Treat them as human-only, and confirm through a channel you already trust, not by replying to the email.

A staged rollout: sort, then draft, then send

Each stage gives you evidence before you hand over more control.

Stage 1: sort only (two to four weeks)

Turn on labels or priority flags and change nothing else. Each day, check the archived or low-priority pile for anything important. Superhuman's Daily Digest lists what Auto Archive moved. Note that removing a wrong label in Email Assistant fixes that thread only. It "does not create a rule" for similar messages, so a person has to adjust the settings.

Stage 2: drafts for routine questions

Let the tool prepare replies, and have a person read every draft before it goes. Gmail's Gemini features include Suggested Replies and "Summarize and draft emails" on eligible Workspace plans. Copilot in Outlook offers Draft with Copilot, and Superhuman Email Assistant prepares response drafts in your voice.

Stage 3: automatic sending for low-risk mail only

Fully automatic replies should be narrow and fixed. Gmail, for example, can already send a saved template as an automatic reply to matching mail. That suits an acknowledgement such as "Thanks, we reply within one working day". It does not suit anything that needs judgement. Keep complaints, quotes and payments out of any auto-send rule, and review the rule's matches every week.

If you would rather not build and tune these stages yourself, Evoogen can set them up around the tools you already use.

Tools compared, with prices

Sources for the table: Google Workspace payment plans (GBP prices), Google Workspace plan features, the AI Inbox announcement of 15 October 2026, Microsoft's UK add-on pricing and Copilot Business FAQ, the Copilot in Outlook FAQ, Superhuman plans, SaneBox pricing and Help Scout pricing. The Microsoft promotional price runs from 1 July to 31 December 2026 for existing customers with an eligible Business plan. Microsoft's Outlook help says Prioritize my inbox is rolling out gradually, so check that it appears in your account before you rely on it.

The prices are not directly comparable. Google's figure buys the whole Workspace suite with email included. Microsoft's is an add-on on top of a Microsoft 365 plan you already pay for. Superhuman and SaneBox sit on top of an existing Gmail or Outlook account. Help Scout replaces a shared mailbox.

A worked example: a three-person online shop

This example is hypothetical. The volumes and timings are assumptions for illustration, not measurements from any tool.

Assumptions. A three-person UK shop already on Google Workspace Business Standard receives 400 emails a week across one shared address and three personal inboxes. The mix: 160 noise, 80 information, 90 routine questions, 40 sales and quote requests, 20 complaints or refunds, 10 legal, security or payment messages. Opening and deciding on each email takes 30 seconds; writing a routine reply from scratch takes 4 minutes.

Before AI. Opening and sorting: 400 × 30 seconds = 200 minutes a week. Routine replies: 90 × 4 minutes = 360 minutes. That is about 9.3 hours a week before quotes and complaints.

Stage 1 (sort only). If the tool files the 240 noise and information messages correctly, the team no longer opens them, saving 240 × 30 seconds = 120 minutes. Checking the filed pile and the daily digest might take 15 minutes a day, or 75 minutes a week over five working days. Net saving: about 45 minutes a week. That is modest, and it shrinks if misfiles force you to check more often.

Stage 2 (drafts for routine questions). If reviewing and editing a draft takes 1.5 minutes instead of 4 minutes of writing, the 90 routine replies drop from 360 to 135 minutes, saving 225 minutes. Quotes, complaints and payment messages are still handled by people, so their time does not change.

Cost. Gemini in Gmail comes with the Business Standard plan this shop already pays for: 3 × £11.80 = £35.40 a month on an annual plan. A shop on Microsoft 365 would add Copilot Business at 3 × £13.80 = £41.40 a month in the first year under the promotion, then 3 × £16.10 = £48.30 at the listed price, on top of its existing plan. If your real numbers show the review time eating the saving, stop at Stage 1.

What to measure

Track a few numbers from the first week, and compare them with a week before the tool was switched on.

• Misfile rate: important emails found in the archived or low-priority pile, per week. Any complaint or payment message misfiled is a reason to tighten rules before moving on.
• Time to first human reply for complaints and quotes. Triage should make this faster, not slower.
• Draft edit level: how many drafts are sent unchanged, lightly edited or rewritten. Many rewrites mean drafting is not saving time yet.
• Wrong commitments: replies that stated an out-of-date price, date or policy. The target is zero.
• Auto-send matches: for any automatic reply rule, review which emails triggered it each week.

How to decide

If you already pay for Google Workspace or Microsoft 365, check which built-in features your plan includes before buying a separate tool. Decide which categories are human-only before anything is automated. Move from sorting to drafts only when complaints and payment mail are no longer being misfiled, and keep automatic sending limited to fixed acknowledgements. Recheck vendor prices before you commit, because promotions and plan names change.

If you want this set up for your business, Evoogen can map how email moves through your team, agree which messages stay with a person, and configure the sorting and drafting stages for you. Talk to Evoogen about setting up AI email triage.