Built by Evisent · In active use
Case study · Inbox management · Built inside Microsoft 365

Email, managed.
15 hours back, every month, per person.

Reading email is the modern professional's biggest unbilled time cost. Triaging it, deciding what matters, deciding what to ignore, drafting replies to the predictable ones, the cognitive load adds up to hours a week of unfocused work. Evisent built an inbox manager that is reclaiming 15 hours of focused time every month for each operator using it.

EVISENT // INBOX RUNNING
Live triage
7 INTENT CLASSES · DRAFT REPLIES
  • Time-sensitive · clientDRAFT REPLY
  • Needs decision · internalSUMMARISE
  • FYI / vendorARCHIVE
  • Newsletter / promoDELETE
Outcome (operator)
15 hrs/mo per person
~180 hrs
Annualised time
back per operator
~25 → 3
Inbox checks per day
before vs after
1–2 wks
Typical Build
delivery time
"The point of automation isn't to remove the work. It's to remove the decisions you don't need to make, so the work that's left is the work that actually matters."
The problem we set out to solve

Inbox triage is invisible work, and most of it doesn't need a human at all.

For most professionals, the inbox is the single largest source of cognitive load and context-switching across a working day. Industry research consistently puts unfocused email handling at 2–4 hours per knowledge worker per day when measured across triage, reading, deciding what matters, drafting replies to routine items, and the multiple times an inbox is "checked" without action being taken.

The deeper cost isn't the minutes. It's the mental-state cost, every inbox visit is a context switch that pulls attention away from focused work, and most of the decisions made during that visit are low-value triage decisions that an AI can make as well or better than a human: this is a newsletter, this is spam dressed up as urgency, this is a routine confirmation, this can wait until tomorrow, this needs your attention now.

We set out to remove the bulk of that decision layer.

The brief in one line

"When I open my inbox, I want to see only the emails that need a human decision, and I want a summary, a recommended action, and a draft reply for everything that does."

What the automation does

Six things, on a schedule, without intervention.

The inbox manager runs on a continuous schedule against the operator's Microsoft 365 mailbox. It uses Anthropic Claude for the language-judgement steps and Power Automate for the rules-based plumbing. Every incoming email passes through six processing layers.

Layer 1

Classify by intent

Every incoming email is classified into one of seven intents, action required, FYI, newsletter, marketing, vendor billing, calendar, or noise. Claude reads sender, subject, body and thread context to decide.

Layer 2

Auto-route by classification

Emails classified as newsletter, marketing or noise are moved to dedicated folders without notification. The operator never sees them in the primary inbox unless they choose to look.

Layer 3

Summarise long threads

Email threads over a certain length are summarised into a 3-sentence brief at the top of the email. The operator reads the brief, decides whether the full thread needs attention, and saves the scroll.

Layer 4

Draft routine replies

For predictable email patterns, confirmation requests, calendar coordination, routine status updates, a draft reply is generated and sat in the email itself for the operator to send, edit or discard with a single click.

Layer 5

Daily priority digest

Every morning, a single email digest summarises what needs attention today, what's been auto-handled, and what's been routed elsewhere. The operator reads one email instead of fifty.

Layer 6

Continuous learning

Operator corrections (re-classifying, restoring routed emails, rejecting draft replies) feed back into the classification logic. The system gets sharper over the first 30 days as it learns the operator's preferences.

The toolchain

Built inside Microsoft 365, on infrastructure the operator already owns.

The entire automation runs inside the operator's existing Microsoft 365 tenant. No external SaaS subscription, no email content leaving the M365 boundary except for the explicit Claude API call (with the operator's data flow agreement). No new vendor relationship for the buyer.

Microsoft 365

Exchange Online, Outlook, Graph API. The host environment.

Power Automate

The orchestration layer. Triggers, routing, folder moves, draft creation.

Anthropic Claude

The language judgement, classification, summarisation, draft generation.

Microsoft Purview

DLP and sensitivity-label aware so privileged content stays inside the boundary.

An honest note on the language layer. This is our own inbox, and it uses Anthropic Claude for the judgement steps, under the data flow agreement above. That is internal use, and we are open about it. Client automations follow a different rule: anything that handles client PII or sensitive data runs on Australian-hosted solutions only.

The outcome

15 hours a month reclaimed, per person. And the email stress is gone.

Time savings are the headline number, but they aren't the most interesting outcome. The deeper change is in mental-state: the inbox stopped being a source of anxiety, and started being a tool used on the operator's schedule rather than the other way around.

  • ~15 hours per month, per person reclaimed on triage, reading, and routine reply drafting. Annualised: ~180 hours a year, per operator, close to a full working month of focused-work capacity.
  • Inbox check frequency dropped from ~25 times a day to ~3 times a day: morning digest, post-lunch sweep, end-of-day. Context switching cost collapsed.
  • Stress reduction was the unexpected outcome. Email stopped being a background source of anxiety. The operator describes the change as "less fearful of opening the inbox", a phrase that came up unprompted in retrospective.
  • Newsletter / marketing volume from the primary inbox dropped by 94%. Auto-routed without ever being seen, accessible if needed but invisible otherwise.
  • Routine reply draft acceptance rate is ~68%, meaning two-thirds of predictable replies are sent as drafted or with a minor edit, rather than written from scratch.

Want your inbox to work like this? Ask us what it would take. There is no pitch at the end.

↳ Want less admin in your inbox?

Book a 30-minute practice review.

Practice Automation packages inbox work like this into fixed-price products for accounting, advice, architecture and engineering practices: filing client email where it belongs, finding TFNs before they leak, and flagging bank-change requests. Products from $1,450 setup, built in your own tenant.