Your Inbox Is a Private Intelligence Goldmine

Most people treat their inbox like a to-do list with bad UX. It is actually something stranger. A multi-year behavioural dataset of every person you work with, sitting on your hard drive. Every message is a timestamp, a relationship, and a small signal about how the sender works.

By the time you have been working for a decade, you are sitting on something like forty to sixty thousand archived emails. That is enough rows of data to build a real picture of any contact: when they are online, when they trust you with attachments, when they go quiet, when they show up for you.

If you are a freelancer juggling fifteen client threads, a founder running on three messaging apps, or anyone whose calendar is mostly a record of conversations, you already own the dataset that explains why some weeks work and others do not. You have probably never read it.

You already own a private OSINT corpus

Open-source intelligence (OSINT) is the discipline of building a picture of someone or something out of public, observable signals. Analysts use it for investigations. Recruiters use it for sourcing. Journalists use it for verification.

You can do something analogous with a corpus that is already yours: your sent and received mail. We call that personal email intelligence, because OSINT technically only covers open sources, and your inbox is a private one. The techniques rhyme. You are inferring patterns from artefacts most people throw away.

Personal email intelligence

The patterns and inferences you can extract from your own email history, including sender rhythms, response ratios, dormant relationships, and attachment archives, without sending a single byte to a third party.

The catch is that "your inbox" usually means a Gmail or Exchange tab. Search is mediocre, exports are clunky, and any "AI feature" that promises insight tends to deliver it by sending the data somewhere it should not go.

Franz's email sub-app does it differently. The signals are computed on-device, against a per-domain SQLite cache. The intelligence panel never makes a network call. Your inbox stays your inbox.

A two-second exercise before we start

Pick one contact. Someone you used to work with closely and have not heard from in a while. You probably cannot say off the top of your head when they used to email you, how often they replied, or how the rhythm of your conversation has changed. The four signals below are what the panel can recover for that person, and for everyone else, in about thirty seconds. Hold the name in your head while you read.

Signal #1: Sender rhythm

The single most actionable pattern is also the simplest: when does a given person actually email you?

Aggregate every message a sender has ever sent into a 7×24 grid, weekday by hour, in their local time. A working rhythm appears. Most people are extremely consistent. They have two or three peak hours, a coffee dip, and a hard cliff at the end of their working day.

Read the demo above like a clock. Maria sends almost everything Tuesday through Thursday between 9 and 11 her time. If you fire off a proposal at 6pm Vienna, it lands at the bottom of tomorrow's queue under twelve newer threads. If you send it at 22:00 your time the night before, just before she opens her laptop with coffee, you are at the top.

The same matrix tells you something else: when not to disturb her. Friday afternoons are cold. Weekends are dead. A "quick question" Saturday morning is not quick. It is a Monday-at-9am question with a 48-hour latency penalty.

The heatmap above runs against demo data. The same panel against your own contacts is a different experience entirely; you suddenly remember why Tuesdays at 11 have always felt productive. (If you want to skip ahead and try it, the trial details are at the bottom of the post.)

Most 'best time to send' studies average across millions of strangers. The only sender that matters for your next email is the one person who's going to read it.

Signal #2: Reciprocity

The second pattern is about the relationship, not the timing.

For every contact, the email sub-app tracks two numbers: how often they write to you, and how often you write back. The ratio is the entire relationship in one chart.

People you reply to faster than they reply to you are the relationships you are carrying. People who write three times for every one of yours are running out of patience. People who used to be 1:1 and are drifting toward 5:1 are about to disappear from your life.

Signal #3: Ghosts

The third pattern is the one nobody wants to admit they need: who you have gone silent with.

Most people have a long tail of relationships they meant to keep up with: old colleagues, former clients, that founder you promised to introduce someone to. The dropoff is invisible by default, because email clients sort by recency. The threads that need attention are the ones that do not show up.

Run against my own inbox the first time, the ghost detector surfaced between twenty and fifty contacts who had crossed the sixty-day threshold without me noticing. The honest number is that most of those relationships will not come back. A small fraction will, and that fraction is enormously valuable.

The ghost detector inverts the recency sort. It surfaces contacts you used to email regularly and have not in N days, sorted by how warm the relationship used to be. It is the most uncomfortable feature in the app, and the one most people thank us for.

Signal #4: Sender enrichment

The fourth pattern is not about you. It is about them.

Before you reply to a stranger, three things matter: who they are, what their company does, and whether anyone in your network knows them. The intelligence panel pulls a profile card together from the public signals attached to their domain and their address: favicon, public profile photo if there is one, social handles parsed from their signature, their company's About copy.

This is the part that comes closest to literal OSINT, and it is the part where we are most careful. We never scrape. We never call back to a third-party enrichment vendor with the email address. Everything you see was either in the email itself or fetched from a public endpoint that the sender's own infrastructure publishes (an /.well-known/security.txt, a public HKP key, a robots-allowed homepage). If your contact opted out of being on the public web, they will not be in the panel. The card will be mostly empty, and that is the right outcome.

Putting it together: finding their loop sweet spot

Each signal on its own is a hint. The compounding move is to look at all four at once.

That sequence (rhythm, ratio, recency, relevance) is the whole loop. Done by hand it is exhausting. Done in the panel it is a thirty-second check before you hit send.

Or skip the panel entirely

There is a quieter version of this that you do not have to think about at all. The intelligence comes to you when it actually matters: at the moment you are about to reply.

The composer reads the same per-domain database the panel reads, and it surfaces the two signals that change the decision in front of you, right under the recipient pill. Their typical reply window. Whether you are about to send into one of their attention windows or one of their dead ones. No clicks, no panel, no detour.

The point is not to nag. It is to make the cost visible at the only moment it can still be acted on. If the chip says "outside their typical hours, expect a delay", you can decide that's fine and send anyway, or you can save it as a draft and send it tomorrow morning in their time. Either is a better outcome than sending into a dead window without noticing.

What you are quietly losing without it

A small loss-frame paragraph, because it would be dishonest to skip it. Most senders in our test data were two or three times more likely to reply at their peak hour than at a random one. If you send fifty meaningful emails a year and half of them land outside the window, that is roughly a week of missed replies. Not catastrophically; quietly, and continuously, in a way that never shows up in any one inbox. Across a network of fifty contacts, the same pattern is responsible for the slow fade you can never quite explain.

The point of the panel is not to optimise every send. It is to make the cost of sending into the wrong window visible enough that you stop doing it by accident.

Local by default. Cloud by choice.

The reason most "email AI" products are bad is structural. To give you insight, they need to read your mail, so they ship it to a server, so the server becomes a target, so the privacy story dies in the second paragraph of the docs.

Franz starts somewhere different. The intelligence panel runs against a per-domain SQLite database on your own disk. Every chart in this post is built from that local cache. No network call leaves the panel. The model that classifies an email as a newsletter or a personal note is on-device. Zone parsing and attachment extraction run inside a worker thread in your own Electron process. Sender rhythm, reciprocity, ghost detection, sender enrichment from public signals: all local, always. The privacy architecture is identical on Pro whether you use your own provider key or Franz Cloud.

The intelligence panel itself is part of Franz Pro. It is available on Pro ($5.99/mo with your own key), Pro Cloud ($9.99/mo), and during the 14-day free trial. The free plan covers connecting services, workspaces, and on-device AI; the intelligence layer unlocks when you start the trial or upgrade.

The heavier AI features are where compute actually costs money: the writing assistant, deep semantic search across years of mail, smart categorisation at scale. There you get a choice, not a forced trip to a SaaS database:

• Local inference (free). A bundled model runs entirely on-device. Honestly slower than the cloud one, especially on older hardware. Nothing leaves your machine.
• Pro with your own key. Bring your own provider key. Franz acts as a thin client. The request goes from your machine to your account at OpenAI, Anthropic, or Google.
• Pro Cloud. Managed cloud inference for users who want speed without setting up keys. Privacy-first by design: we do not train on your inbox, we do not keep your messages, and the boundary you would expect from an EU-headquartered tool holds end-to-end. You are paying for compute, not for a copy of your data.

A few honest limitations, because they will come up. The panel only knows what your machine knows: there is no cross-device profile, and the desktop app is where the data lives. Franz does not have a mobile app today. The panel needs roughly a day to backfill before patterns are useful. And the heatmap is only as good as your archive; if you only just connected an account, give it a week.

The full breakdown of what stays local, what Pro Cloud is used for, and what Franz never does with your content lives in How Pro Cloud handles your data.

We did not get to the fifth signal

There is a fifth signal that did not fit in this post: the attachment archive. Every file you have ever exchanged with this person, sorted by quarter, surfaced inline. It is surprisingly emotional to scroll through. Another time.

Try it on your own inbox

Think back to the contact you held in your head at the start. The trial puts the panel against that person's history, in your own data, in about a day.