How AI handles my email before I read it
Email is the tax you pay for running a company. Not the writing, the writing is the easy part. It is the standing start. Open the inbox, reconstruct who this person is, remember what I promised them three weeks ago, find the last thing I sent them, decide whether this even needs me, and only then write four sentences. Forty times a day. The reply is nothing. The runway to the reply is everything.
So I did not build a smarter spam filter. I built a colleague.
Meet Carla
FlatNine runs on an ensemble of agents, each one a named character with a job. Toshiko runs the newsletter. Bill writes the changelog. Nat audits customer notifications. Carla owns communication: the sales pipeline, and my personal inbox.
Once an hour, Carla wakes up as what we call a deep task. Not a single prompt and a single answer, but an agent with a goal, a set of tools, and permission to take a few steps on its own before it stops. Her goal this hour is narrow and clear: read my inbox, and leave me ready to act on it.
What happens every hour
She pulls my twenty most recent emails and sorts each one into exactly one of three piles.
Draft it. A real person or business that genuinely needs a reply from me, where she is confident she can write a correct, useful one. A customer with a question. A vendor confirming a date. A professor following up. She writes the reply and leaves it in my Gmail as a draft, threaded under the original, waiting for me.
Skip it. Newsletters, promotions, social notifications, no-reply senders, anything automated or personal in a way that is not hers to touch. She reads the Gmail category labels and the sender and moves on. No draft, no noise.
Flag it. This is the interesting one. If an email needs a reply but answering it would require a fact she does not have (a price, a decision, a credential, a commitment only I can make), she does not guess and she does not write a placeholder. She flags it for me with one line on why. Same if something is just important: a declined payment, an angry customer, a real deadline. She would rather hand me a clean decision than a confident wrong answer.
Drafting in my voice
A draft is only useful if it sounds like me, otherwise editing it costs as much as writing it. So Carla grounds every reply in two things.
The first is a voice skill: a written description of how I actually email. Short. Action first. Confident. Greeting and sign-off matched to who I am talking to. And one rule she is never allowed to break, which is no em dashes, because I do not use them and nothing gives away a machine faster.
The second is stronger than any rule: my own sent mail. Before she writes to someone, she pulls a handful of my real past replies to that exact person and mirrors how I talk to them. The salutation I use with them. The length. The running context. Whether we are formal or three years into an in-joke. She is not imitating a generic me, she is imitating me talking to you.
And she never invents. No made-up prices, dates, or promises. If writing the reply would force her to fabricate a fact, that email becomes a flag, not a draft. That single rule is what makes the whole thing trustworthy.
The threads I am waiting on
Drafting replies is the obvious half. The half that actually changes how it feels to run a company is the chasing.
Carla runs a Reply Zero pass: she looks for threads where I sent the last message and the other side has gone quiet, and where a nudge is genuinely warranted. Then she drafts the gentle follow-up for me. "Just circling back on this, did the contract work for you?" The things that fall through the cracks not because they are hard but because remembering to chase them is a full time job nobody wants.
She is deliberately conservative here. She would rather miss a borderline nudge than pester someone who is mid-conversation. A handful per run, no more.
The promises I made
The mirror of that is the one I find quietly remarkable. Carla reads my own sent mail for commitments I made. "I will send you the deck." "Let me check and get back to you." "I will have it by Friday." Then she works out which of those are still outstanding and now overdue, because nothing later in the thread shows I actually did them.
If the follow-through is something she can write without inventing anything, she drafts it. If it needs the actual deck, or a price, or a decision only I can make, she does not fake it, she just tells me: you promised this to this person on this date, still owed. My own past self, audited by a colleague who never forgets.
She drafts, I send
Here is the line that does not move. Carla creates drafts. She never sends, never archives, never deletes, never labels. Every reply she writes sits in my inbox until I open it, read it, maybe change a word, and send it myself.
That is not a limitation I am working around, it is the entire design. I stay the person who decides and the person who is accountable for what goes out under my name. What disappears is the standing start: the blank page, the context rebuild, the "who is this and what did I promise them." By the time I get to an email, the hard part is already done and the only thing left is the part that should be mine, the judgment and the send.
The other half: Swing
The same instinct runs my sales pipeline. Swing is our CRM, and Carla works it the way she works my inbox, except here the writing is only half the job. The other half is keeping the board honest.
Start with the writing, because it is familiar. When a new lead lands, Carla drafts the first outreach. She drafts the follow-ups on a sensible cadence, a few days out, never landing on a weekend, mirroring how a real person actually writes. She chases delinquent customers about a failed payment with a reminder that is firm but human. And when a deal has gone unanswered too many times, she stops flogging it and moves it to Lost, because knowing when to give up is part of the job too.
Now the part I find more interesting: she moves deals across the pipeline.
A sales board is a lie the moment you stop tending it. Cards sit in "Outreaching" weeks after the meeting already happened. A deal is parked in "Pending Meeting" for someone who no-showed and never rebooked. The board says one thing and reality says another, and every stale card is a small decision you forgot to make. Keeping it accurate is pure tax, and it is exactly the kind of tax Carla is good at.
So once an hour she reads every in-flight deal and asks one question: what stage should this actually be in? She does not guess from vibes. She reads the real signals, the meetings booked on my calendar and the full email and notes history on the card, and works out where the deal truly stands.
She moves along a fixed map I gave her, the same path a deal really travels:
- A deal in "To Process" or "Outreaching" with a meeting now on my calendar moves to "Pending Meeting."
- A deal in "Pending Meeting" whose meeting time has passed gets the thread read. If they showed and it went well, it moves to "Closing." If they no-showed or went quiet, it drops back to "Outreaching" to be re-engaged.
- A deal in "Closing" only leaves when the thread proves it is done: paying, dead, or paused.
That map matters as much as the judgment. "Closing" only ever moves forward, so a deal that has merely gone quiet stays in Closing instead of getting silently demoted. And the two moves that actually touch money or a relationship, marking a deal Won or Lost, she is never allowed to make on her own. Those she only suggests, and I make the call.
Everything else runs on confidence. When the signal is unambiguous, a clear future meeting, an explicit no-show, she makes the move and writes a one-line note on the card explaining why, so the history always shows who moved it and on what evidence. When she is less sure, she leaves the stage alone and just drops the suggestion on the card for me to accept or ignore.
I did not switch this on and walk away. It ran in a read-only mode first, proposing every move and changing nothing, until I had watched it be right enough times to trust it with the live board. Now it quietly keeps the pipeline matching reality, so when I open Swing I am looking at the truth instead of a week-old snapshot of it.
Inbox and pipeline turn out to be the same problem wearing two outfits. A stream of people who need a timely, in-character, factually grounded message from me, and a stream of small bookkeeping decisions that pile up the instant I look away. One agent, pointed at both, doing the standing-start work so the only thing left for me is the judgment.
Why this shape works
I did not want software that answers my email. I wanted to keep answering my own email, minus the part that was never really me to begin with. The reconstruction. The remembering. The cold start.
Carla turns my inbox from a pile of standing starts into a stack of nearly-finished decisions. That is the whole trick, and it is the same trick behind everything I build: find the expensive, unglamorous middle of a task, automate exactly that, and leave the human the part that was worth their attention in the first place.
This is the kind of thing I build at FlatNine. I post the builds @mikerubini.