Envelope · coming soon
It’s us against the machines. Gmail’s tabs and Outlook’s Focused Inbox already sort out the newsletters and receipts. What they miss is software built to look like a person: cold email, automated sequences, fake “Re:” follow-ups. In my inbox, Gmail put 97% of cold outreach in the Primary tab and marked 95% of it Important, the same as mail from real people (the numbers).
Envelope tells the two apart and files the machines out of your way. It never reads a word anyone wrote.
Why this exists
It started with one inbox. I counted every email that reached me over three years: 138,247 of them. 10% were sent by a person, and this year it is down to 8%. Machine-sent mail is up 6× since 2017 and hasn’t let up. By the published research it costs about 2.1 hours a day just to decide what not to read. The whole study follows, exactly as I ran it.
Who is on the other side
A whole industry exists to get a machine’s email into your inbox looking like a person’s. Clay and ZoomInfo find and enrich the names: Clay is valued at $7.1 billion, and ZoomInfo had $1.2 billion in revenue last year. Outreach ($4.4 billion) and Apollo ($1.6 billion) run the sequences, and HubSpot, which took in $3.1 billion last year, sells them in its Sales Hub. AI “SDR” startups like 11x and Artisan, which write and send the emails for you, have raised about $100 million between them. And two of the tools that send the most cold email I receive, Instantly and Smartlead, never needed investors at all: sending it pays for itself.
Billions of dollars, all pointed at your inbox. Envelope reads what they can’t disguise, the envelope every email arrives in, and moves the machine-sent mail out of your way.
Valuations at each private company’s latest priced round; HubSpot’s reported revenue (it does not break out Sales Hub) and ZoomInfo’s; amounts raised for the AI SDRs: Clay, $7.1B, Sep 2026, ZoomInfo, $1.25B revenue, 2025, Outreach, $4.4B, 2021, Apollo, $1.6B, 2023, HubSpot, $3.13B revenue, 2025, 11x, $74M, Artisan, $25M Series A.
You guessed 50%. In my inbox it was 10% over three years, and 8% this year. Two friends join and we count yours.
What joining gets you
1A place on the waitlist
Join and you’re in line for Envelope itself. When it’s ready, we’ll let people in by their place in line. You’ll get one email, and that’s all joining asks of you.
2Five envelope invites
Each one is an email you send from your own inbox to someone you actually write to. A friend who joins with yours moves 5 places up the line, and you move up 10 places with each one. You also get a share link to post anywhere: each person who joins from it earns you another envelope invite, up to 20 more.
3A study of your own inbox
When friends join with your envelopes, you earn the same study as below, run on your own mail:
The study runs on your own computer, from a Google Takeout of your mail. Only the counts leave it.
We will not email you updates. One email when Envelope is ready.
Thank you. One email when it is ready.
The study · my mailbox ยท Sep '23 to Aug '26
138,247 emails that reached me from outside my company, grouped by what sent them and when. Spam folder excluded.
Spam is excluded: Google Workspace keeps only about thirty days of it, so it cannot be compared across years.
The share of each year’s mail sent by a person, against everything sent by a machine.
In 2017, 38% of the mail that reached me was sent by a person. In 2026 it is 8%. People did not stop writing to me; they send about as much as they ever did. Machine mail went from 18 a day to 117, 6 times as much, while mail from people stayed between 10 and 15 a day. The share fell because the machines multiplied.
| Year | From a person, a day | From a machine, a day | Sent by a person |
|---|---|---|---|
| 2017 | 11 | 18 | 38% |
| 2018 | 12 | 24 | 33% |
| 2019 | 14 | 60 | 19% |
| 2020 | 12 | 82 | 12% |
| 2021 | 15 | 88 | 14% |
| 2022 | 13 | 113 | 11% |
| 2023 | 14 | 109 | 12% |
| 2024 | 12 | 102 | 11% |
| 2025 | 13 | 125 | 10% |
| 2026 (to Aug) | 10 | 117 | 8% |
Read this as a floor. “Not sent by a person” counts only mail where something identified a machine. Cold email sent by hand from an ordinary mailbox lands on the person side, because from the outside it is indistinguishable from a stranger writing to me. So at least 92% of last year’s mail was not sent by a person; the real figure is higher and cannot be measured this way.
Not for this study: for looking at it as it came in.
Every message counted in this report was looked at by me, if only long enough to decide not to read it. The research on what that looking costs is consistent, and it is not mainly about reading time. It is about the interruption, the switch, and the attention that does not come back when a person returns.
More email, measurably worse. In the CHI 2016 study, the more of the day a person spent in email, the lower they rated their own productivity and the higher their stress read on a heart-rate monitor (source 4). Cutting checks to three a day lowered daily stress in a randomised trial (source 7); a week without email at all lowered heart-rate-variability stress and lengthened the time people stayed in one window (source 8). Part of the cost is what Leroy calls attention residue: when a person switches before a task is finished, some of their attention stays behind on it and the next task suffers (source 9).
My mailbox, through those numbers. This is my arithmetic, not a study’s. 126 messages a day arrived over the window. At the Microsoft skim ceiling of one minute each, that is 2.1 hours a day spent only deciding, about 1.3 working days out of every five, before a single reply is written and before any of the switching costs above are counted. Envelope’s claim is not that it helps me read faster. It is that most of these messages never needed the look.
Every message from outside my company, stacked by class: a sales-sequencing tool, a marketing platform, a mailing list, a transactional sender, a machine I could not name, or no sign of automation at all.
| Month | Cold | Marketing | Notifications | Lists | Person | Unnamed |
|---|---|---|---|---|---|---|
| Sep '23 | 3.1% | 51.2% | 28.1% | 1.7% | 11.1% | 4.8% |
| Oct '23 | 2.4% | 49.0% | 31.3% | 1.8% | 10.7% | 4.8% |
| Nov '23 | 2.3% | 46.9% | 37.5% | 2.2% | 8.7% | 2.5% |
| Dec '23 | 2.4% | 51.3% | 31.1% | 1.8% | 10.4% | 3.0% |
| Jan '24 | 2.5% | 38.7% | 40.4% | 3.7% | 11.2% | 3.4% |
| Feb '24 | 1.6% | 38.8% | 40.7% | 3.8% | 12.3% | 2.7% |
| Mar '24 | 1.1% | 33.4% | 51.3% | 2.5% | 9.2% | 2.4% |
| Apr '24 | 1.5% | 34.8% | 47.5% | 1.9% | 12.1% | 2.2% |
| May '24 | 1.8% | 36.8% | 45.2% | 2.0% | 11.2% | 3.1% |
| Jun '24 | 2.1% | 33.2% | 53.1% | 0.2% | 9.2% | 2.2% |
| Jul '24 | 3.5% | 33.9% | 45.2% | 1.9% | 12.2% | 3.3% |
| Aug '24 | 3.1% | 34.8% | 46.1% | 2.1% | 11.0% | 2.9% |
| Sep '24 | 2.3% | 34.6% | 46.3% | 1.3% | 11.2% | 4.3% |
| Oct '24 | 2.4% | 36.2% | 46.7% | 2.2% | 9.6% | 2.9% |
| Nov '24 | 3.3% | 41.1% | 41.6% | 0.8% | 9.0% | 4.2% |
| Dec '24 | 3.1% | 38.5% | 44.0% | 3.2% | 8.7% | 2.6% |
| Jan '25 | 3.1% | 37.2% | 44.9% | 1.0% | 10.8% | 3.1% |
| Feb '25 | 3.7% | 35.5% | 44.2% | 0.5% | 13.0% | 3.1% |
| Mar '25 | 4.9% | 33.5% | 45.3% | 0.1% | 12.3% | 3.8% |
| Apr '25 | 5.7% | 29.9% | 51.6% | 0.2% | 9.8% | 2.8% |
| May '25 | 5.9% | 26.0% | 53.9% | 0.4% | 11.3% | 2.5% |
| Jun '25 | 4.2% | 28.6% | 51.9% | 0.7% | 10.3% | 4.3% |
| Jul '25 | 3.9% | 25.9% | 56.0% | 0.3% | 10.7% | 3.1% |
| Aug '25 | 4.5% | 28.5% | 53.2% | 1.6% | 8.5% | 3.7% |
| Sep '25 | 4.5% | 27.8% | 56.6% | 0.7% | 7.3% | 3.1% |
| Oct '25 | 5.1% | 30.8% | 53.1% | 0.5% | 8.2% | 2.4% |
| Nov '25 | 6.4% | 37.4% | 46.2% | 0.4% | 6.5% | 3.1% |
| Dec '25 | 4.8% | 42.6% | 42.4% | 0.2% | 7.4% | 2.6% |
| Jan '26 | 7.0% | 35.5% | 46.5% | 0.3% | 7.7% | 2.9% |
| Feb '26 | 6.0% | 30.9% | 49.1% | 0.3% | 11.4% | 2.3% |
| Mar '26 | 6.4% | 36.2% | 43.0% | 0.3% | 11.0% | 3.0% |
| Apr '26 | 7.6% | 45.1% | 37.7% | 0.1% | 6.0% | 3.5% |
| May '26 | 7.0% | 47.1% | 34.0% | 0.2% | 7.8% | 4.0% |
| Jun '26 | 5.8% | 45.2% | 38.5% | 0.1% | 6.7% | 3.7% |
| Jul '26 | 5.9% | 46.4% | 38.6% | 0.2% | 5.7% | 3.2% |
| Aug '26 | 7.1% | 48.9% | 30.6% | 0.4% | 9.1% | 3.9% |
Monthly columns across 10 years are a picket fence, so the long arc gets one column per year: 2017 to 2026, 336,939 messages. Same six classes, same colours as above. The last column covers only January to August 2026. The archive itself reaches back to 2013; everything before 2017 is left out on purpose. The figures and charts above are the last 36 months of this.
A ranked list says who is loudest now. This says when my mailbox acquired them: the dot is the first message in this window I can attribute to that tool, the bar is how long it has kept going, and the figure on the right is its share of the last twelve months’ cold outreach, so a tool that has gone quiet reads as 0%.
The six marketing platforms that delivered the most mail in this window, one panel each, all on the same vertical scale so the panels can be compared. The figure at the top right of each panel is its share of all marketing mail in the window.
The same view for the six sales-sequencing tools with the most mail; the figure at the top right is its share of all cold outreach in the window. These are the systems reps use to send a scheduled series of one-to-one-looking messages to people who have not asked for them.
| Platform | Share | 12-mo |
|---|---|---|
| INInstantly | 28.7% | ▼ |
| SESendGrid | 18.3% | ▲ |
| ASamazon-ses | 11.4% | ▲ |
| MRunbranded, via Microsoft 365 | 9.7% | ▲ |
| SMSmartlead | 9.6% | ▼ |
| MAmailgun | 6.9% | ▲ |
| HSHubSpot Sales | 3.6% | ▼ |
| OUOutreach | 3.5% | ▼ |
| MFunbranded mailbox farm | 1.7% | ▲ |
| SASalesloft | 0.8% | → |
| GAgmail-api-client | 0.5% | ▼ |
| LElemlist | 0.5% | ▼ |
| BUbuzzstream | 0.4% | → |
| APApollo | 0.4% | ▲ |
| LIlinkdoctor | 0.4% | ▲ |
| RIReply.io | 0.3% | → |
| UTunidentified tracker | 0.3% | ▼ |
| GRgrowthjack | 0.3% | → |
| FMfmailroute (sending infra) | 0.2% | → |
| RErespona | 0.2% | → |
| Platform | Share | 12-mo |
|---|---|---|
| HUhubspot | 25.4% | → |
| SEsfmc-exacttarget | 15.6% | ▼ |
| MAmarketo | 8.8% | ▼ |
| MAmailchimp | 8.0% | ▼ |
| BEbeehiiv | 6.0% | → |
| unidentified vendor | 4.3% | → |
| KLklaviyo | 3.9% | ▲ |
| SAsailthru | 3.9% | ▼ |
| CCconstant-contact | 3.7% | → |
| SUsubstack | 2.7% | ▲ |
| SESendGrid | 2.5% | ▲ |
| ITiterable | 2.1% | ▲ |
| CIcustomer.io | 1.9% | → |
| SEsendinblue | 1.7% | → |
| ASamazon-ses | 1.6% | ▲ |
| MAmailgun | 1.0% | ▲ |
| CMcampaign-monitor | 0.8% | ▼ |
| KCkit-convertkit | 0.8% | ▲ |
| ACactivecampaign | 0.7% | → |
| MAmailjet | 0.6% | ▲ |
A cold email is rarely alone. The tool that sends it schedules the next one before the first is opened, and the whole run is one sequence: same sender, same thread, a few days apart, until someone answers or the list runs out. Grouping the 2,947 cold messages of Sep '25 to Aug '26 that way gives the shape below.
Most sequences show a single message here, 1.27 on average. Some were one and done. Some followed up under a fresh subject line, which this count treats as a new sequence. And some follow-ups went to spam, which Google Workspace discards after thirty days, so they are not in the archive at all. The follow-up share is a floor.
Where a sequence did follow up, this is how long it waited. 86% of follow-ups came within a week; the most common gap is three to four days, which is the default in most sequencing tools.
One more tell. 99% of the marketing mail in this window carried an unsubscribe header, because the law and the mailbox providers require one for bulk mail. 5% of the cold outreach did. The sequencers send one message at a time to one address at a time, so nothing about the envelope says bulk, and no unsubscribe link is offered, because that would give the game away.
An open-tracking pixel is an invisible image in the HTML of a message that reports back when it is loaded. This study never reads a body, so it cannot count pixels. It can count the messages where a pixel is impossible: a message sent as plain text has no HTML to hide one in. 65% of the cold outreach in this window was plain text only, against 0% of the marketing and 2% of the mail from people. That is a choice, not a limitation: the sequencers default to plain text because a pixel is the fastest way to a spam folder, and because a person writing one email rarely sends HTML with no text at all. The rest could carry one; how many do, the envelope cannot say.
| Class | Plain text only (no pixel possible) | Text and HTML | HTML only |
|---|---|---|---|
| Cold outreach | 65% | 26% | 9% |
| Marketing | 0% | 96% | 4% |
| Sent by a person | 2% | 97% | 1% |
| Automated, sender I know | 0% | 87% | 13% |
| Notifications, kept | 1% | 94% | 5% |
| Notifications, filed | 0% | 96% | 4% |
Every sequencing tool leaves its own habits in the envelope, and they differ more than the copy does. HubSpot Sales runs the longest sequences (2.1 touches on average) and unbranded mailbox farm the shortest (1.0). Unbranded mailbox farm fakes the thread most often, on 100% of its messages, by referencing an id that never arrived or putting “Re:” on a message with no parent. Unbranded, via Microsoft 365 sends 88% of its mail as plain text, where no tracking pixel can live; Apollo sends 0%. Outreach offers a way out most often, an unsubscribe header on 43% of its mail; no other tool passes 8%. HubSpot Sales sends 33% of its mail outside 8am to 6pm my time. And unbranded, via Microsoft 365 burns through sending domains fastest: 502 messages from 314 domains, fewer than 1.6 each.
| Tool | Msgs | Sequences | Touches | Days apart | Plain text | Unsubscribe | Fake thread | Off hours | Peak |
|---|---|---|---|---|---|---|---|---|---|
| Apollo | 21 | 17 | 1.2 | 5 | 0% | 5% | 5% | 14% | 10am |
| HubSpot Sales | 98 | 46 | 2.1 | 4 | 0% | 8% | 1% | 33% | 1pm |
| Instantly | 930 | 723 | 1.3 | 4 | 75% | 6% | 1% | 9% | 9am |
| unbranded, via Microsoft 365 | 502 | 446 | 1.1 | 3 | 88% | 2% | 38% | 8% | 10am |
| unbranded mailbox farm | 276 | 276 | 1.0 | – | 83% | 0% | 100% | 10% | 12pm |
| Outreach | 56 | 45 | 1.2 | 5 | 0% | 43% | 0% | 11% | 9am |
| Smartlead | 300 | 192 | 1.6 | 3 | 59% | 1% | 0% | 11% | 10am |
| no platform named | 582 | 451 | 1.3 | 4 | 58% | 3% | 14% | 15% | 11am |
Sep '25 to Aug '26, tools with at least 20 cold messages. Touches: mean messages per sequence. Days apart: median gap between touches. Plain text: share with no HTML part, so no pixel is possible. Fake thread: share whose References or “Re:” point at nothing that arrived. Off hours: share sent outside 8am–6pm in my time zone. Peak: the commonest send hour. All from headers; counts only.
Where in the week cold outreach lands, by the day it reached my inbox, local time, Sep '25 to Aug '26. Each bar is that day’s share of a typical week, with every weekday divided by how many times it actually fell in the window, so Saturday is not divided by the same number as Tuesday.
The mail is compressed into five days, and Monday carries the most of it.
Share of cold outreach by the hour it arrived, my local time, Sep '25 to Aug '26.
82% arrives between 9am and 6pm, and the busiest hour is 9am. The tools that send this mail are scheduled to land in my working day.
Every day from Sep '25 to Aug '26, one square each, Monday at the top of each column. Darker is more mail; the shading is relative within each grid, so the two grids compare shape, not size. Hover a square for the day and its count.
The machines never took a day off: 0 days in the year with nothing from them. People had 4 weekdays with nothing at all, and the weekends are the two faint rows at the bottom of the second grid.
The holidays are the tell. On the days offices actually close (Christmas Day, Christmas Eve, New Year's Day, Thanksgiving) the people writing to me drop to 2 a day against 13 on an ordinary weekday, and the cold outreach drops with them, 3 against 11, because a sequencer is set to send on business days and knows the calendar. The rest of the machines do not: 64 a day against 145. Christmas Day: 0 from people, 70 from machines. The Monday holidays (Labor Day, MLK Day, Presidents' Day, Memorial Day) split the room: people mostly take them, 3 a day, while the sequencers treat them as a working Monday, 12 cold a day, and the rest of the machines send 154.
| Weekday holiday | Machines | of which cold outreach | People | |
|---|---|---|---|---|
| Labor Day | Mon 01 Sep 2025 | 194 | 7 | 1 |
| Thanksgiving | Thu 27 Nov 2025 | 78 | 3 | 0 |
| Day after Thanksgiving | Fri 28 Nov 2025 | 95 | 10 | 2 |
| Christmas Eve | Wed 24 Dec 2025 | 57 | 2 | 5 |
| Christmas Day | Thu 25 Dec 2025 | 70 | 2 | 0 |
| New Year's Eve | Wed 31 Dec 2025 | 74 | 7 | 1 |
| New Year's Day | Thu 01 Jan 2026 | 50 | 5 | 1 |
| MLK Day | Mon 19 Jan 2026 | 159 | 18 | 4 |
| Presidents' Day | Mon 16 Feb 2026 | 172 | 9 | 7 |
| Memorial Day | Mon 25 May 2026 | 93 | 12 | 1 |
| Ordinary weekday (median) | 145 | 11 | 13 |
The other half of the mailbox is mine: what I sent, and how fast, Sep '25 to Aug '26. The same pass that counted what reached me matched my replies to the messages they answered, and kept only the tallies.
Counted by conversation, not by message: a thread with six messages from one person and two replies from me is one conversation, answered. Only mail addressed to me directly counts; mail I was copied on does not. Message by message, 34% of what people sent me directly got its own reply. Read it as a floor: "sent by a person" is whatever the sorter could not prove was a machine, so every well-disguised sequence, every FYI that needed no answer, and every question I answered by phone or in person counts here as a conversation I ignored.
My replies to people, by how long after their message mine went out. A quarter went within 21 min; three quarters within 7 h 16 min; 50% within two hours.
Share of the email I sent, by the hour it left, my local time. 70% between 9am and 6pm, 14% after 8pm, and the busiest hour is 12pm. Tuesday carries the most of the week; 7% goes out at the weekend.
Put the two halves together and the shape of the day is clear. The machines send at 9am. I answer people around 12pm. Everything between those two is the sorting.
Messages a day, by year, three lines. Mine has run between 11 to 24 a day since 2017 and the people writing to me between 10 to 15. The third line is the machines, 18 to 117, 6 times over. Nothing about the humans in this picture changed. The volume did.
Every one of these messages existed to get a reply. Here is how often it worked on me, Sep '25 to Aug '26.
22 replies to 11 sequences: 9 of them on the first touch, the rest after being followed up. When I did answer, I took a median of 4 h 3 min. After my reply, a person answered in 6 of the 11 sequences. In 1 the tool never noticed and fired its next scheduled message anyway. 4 went silent. Since 2017 the shape is the same: 231 replies to 8,895 cold emails (2.6%), in 159 sequences; a person wrote back in 56% of those sequences, 10 rolled on as if I had said nothing, and 60 never wrote again.
Every message in a Takeout archive carries the labels Gmail gave it: which tab it went to, whether it was marked Important, whether it was ever opened. So the mailbox provider's verdict on each class can be tallied against mine, Sep '25 to Aug '26.
Gmail is good at marketing: 71% of it went to Promotions and only 17% was marked Important. It cannot see cold outreach at all. 95% was marked Important and 97% went to Primary, the same treatment as mail from a person (97% and 93%). The difference shows up in what I did next: I left 44% of the cold outreach unread against 11% of what people sent me. Gmail filed it as a person; I could tell it was not; the sorting was left to me.
| Class | Marked Important | Primary tab | Never opened |
|---|---|---|---|
| Sent by a person | 97% | 93% | 11% |
| Cold outreach | 95% | 97% | 44% |
| Marketing | 17% | 1% | 87% |
| Automated, sender I know | 28% | 10% | 75% |
| Notifications, kept | 55% | 36% | 58% |
| Notifications, filed | 33% | 4% | 72% |
| Lists | 12% | 1% | 88% |
Why the machines win here: Gmail's Important and Primary decisions lean on whether the sender looks like a person writing one message to one address, and a sequencer is built to look exactly like that. Marketing platforms declare themselves in the headers; sequencers do the opposite.
Everything above is mail that reached my inbox. Google Workspace keeps only about thirty days of spam, so this is a snapshot, not a trend: Sep '26. It is here because it answers the obvious objection to the cold-outreach number above: most cold email never reaches the inbox at all.
| Class | Messages | Share |
|---|---|---|
| Cold outreach | 285 | 67% |
| Marketing | 71 | 17% |
| Sent by a person | 35 | 8% |
| Notifications | 22 | 5% |
| Machine, transport only | 12 | 3% |
| Automated, sender you know | 2 | 0% |
| Machine, platform unknown | 1 | 0% |
The same tools that reach my inbox are behind the mail that did not. Of the cold outreach in spam, 32% came through Microsoft 365 with no vendor name attached, and 37% carried the name of a sending tool.
| Platform | Share |
|---|---|
| MRunbranded, via Microsoft 365 | 32% |
| unidentified vendor | 31% |
| INInstantly | 29% |
| SESendGrid | 2% |
| FMfmailroute (sending infra) | 2% |
| ACazure-comm-services | 1% |
| GAgmail-api-client | 1% |
| HSHubSpot Sales | 1% |
| APApollo | <1%% |
| ARartisan (AI SDR) | <1%% |
| GRgrowthjack | <1%% |
| KAkarmalinks (tracker) | <1%% |
| Platform | Share |
|---|---|
| SAsalesforce | 31% |
| unidentified platform | 23% |
| SEsendinblue | 11% |
| SEsfmc-exacttarget | 11% |
| MAmailgun | 10% |
| MAmarketo | 10% |
| HUhubspot | 1% |
| IKinfusionsoft-keap | 1% |
| SESendGrid | 1% |
This is the one place the method can be graded against a known answer. Every message here is junk; Google Workspace already decided that. If the method were perfect, all 428 would show a machine behind them. In fact 83% do: 285 came from a sales-sequencing tool and 71 from a marketing platform. Another 9% show automation of some kind (a relay, a notification template, a mailing list) without a nameable sender. The remaining 8%, 35 messages, show no sign of automation at all. These are not people writing to me. They are spammers sending from an ordinary Gmail or Outlook mailbox, or from a domain registered last week, with no tooling attached, and from the outside that is indistinguishable from a stranger typing me a note by hand. That is the ceiling on this kind of filing: it catches the automation, and here 8% of the junk had none. It is why Envelope sits on top of the spam filter rather than replacing it.
This is my mailbox. Everything here is an exact count of what arrived in one inbox, mine, over 36 months. It is not a survey, not a sample of the market, and not a claim about anyone else's mail. It is one mailbox measured properly, which is a different thing from a small mailbox measured loosely.
Marketing and cold outreach are split by what sent the mail, not by what it says. Mail that leaves through a marketing platform (Mailchimp, Marketo, HubSpot's marketing hub) counts as marketing, even when the campaign is aimed at strangers. Mail that leaves through a sales-sequencing tool (Outreach, Salesloft, Apollo, HubSpot Sales) counts as cold outreach, even when it looks like a note from one person. Where one vendor offers both, as HubSpot does, the two products are counted separately. Anything with no identifiable sender is counted as machine, unnamed rather than guessed into one of the two; that bar is the honest measure of what I could not attribute.
“Notifications” means system mail, and I do not try to tell a receipt from an alert. That would take reading the message, and nothing here reads the message. What the headers can tell is whether anybody is home: a ticket or case system, an e-signature request, a message in a running thread or from a shared inbox is kept; a message from a no-reply address, or one the sending platform itself labelled transactional, is filed. The class used to be called Receipts; the hand check above is why it is not.
Sent mail, spam, trash and drafts are excluded, as is anything from my own domains.
Join, send your five envelopes, and when five friends join we run it on your mail.
Thank you. One email when it is ready.