Home / Case studies / AI Inbox Manager · Built by Evisent
Built by Evisent · In active use
Case study · Inbox management · Microsoft 365 + Power Automate + Anthropic Claude

Email = managed.
3.5 hours back, every week.

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 AI-powered inbox manager to automate the bulk of that triage. In active use, it's reclaiming 3.5 hours of focused time every week for its operator.

Want one for your business? → Take the readiness quiz
Focused work environment without inbox noise
M365 + POWER AUTOMATE + CLAUDE
3.5 hrs
Reclaimed per week
per operator
~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

Microsoft-aligned on infrastructure. Vendor-neutral on the AI.

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.

We chose Claude for the language layer because its reasoning quality on classification-and-summarisation workloads is consistently strong, but the architecture is intentionally swappable — the same workflow runs against Azure OpenAI or Microsoft Copilot if a client prefers a fully Microsoft-stack deployment. Vendor-neutral on the AI choice; Microsoft-aligned on everything else.

The outcome

3.5 hours a week reclaimed. 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.

  • ~3.5 hours per week 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.
If you'd like one for your business

The inbox manager is a Build, not a product.

We haven't packaged the inbox manager as a per-seat SaaS subscription — and we won't. Every operator's inbox is different, every team's email pattern is different, and the classification logic only works when it's tuned to the operator. We deliver it as a Build & Operate engagement: scoped to your team, deployed in your Microsoft 365 tenant, tuned over the first 30 days, then maintained as it learns.

Single operator

A senior executive, founder or accounts-payable lead drowning in email. Build typically delivered in 1-2 weeks, tuned over the following 30 days, Operate retainer optional.

See the Build & Operate engagement →

Team rollout (5-15 inboxes)

A management team, sales pod, or accounts function. Per-operator tuning happens in parallel. Annualised time-reclaim across the team typically pays back the Build inside the first month.

See the Build engagement →

Existing AI Governance client

The Bundle already covers governance for Copilot and any other AI in use. Adding the inbox manager as a Build sits inside the same governance framework — no separate AUP work needed.

See the AI Governance Bundle →
Three ways to get your week back

Take the quiz. Scope a Build directly. Or have a chat.

The AI Readiness Quiz routes you to the right SKU based on your AI footprint, urgency and use case. Or skip the quiz and scope a Build engagement directly. Or pick up the phone.

Take the AI Readiness Quiz → Scope a Build Call · 1300 384 736