AI assistant · our own back office

One person is the bottleneck.
This is the hour we take off them.

An owner-operator spends the first hour of every day harvesting, sorting, drafting and chasing. None of it is the work. All of it has to happen before the work starts.

Adapton runs this on its own back office, every day. What follows is the architecture and the constraints it was built under.

How it works Book a 15-minute call Your numbers, on the call
Three jobs run themselves; the last step is a person Harvest at 06:30, then triage against written rules, then drafting — each feeding the next. The chain stops at a refusal marked approval gate: nothing sends on your behalf. Below it, a person clicks send. 06:30 Harvestthe places the work already livesTriagewritten rules, not vibesDrafttext produced, nothing sentAPPROVAL GATE — nothing sends on your behalfYou click send
Three jobs run themselves. The last step is a person, deliberately.

Every figure here is a build fact — counted, not claimed. No outcome numbers appear on this site until a client’s system produces one.

  • dailyrunning on our own operation
  • 0messages sent without a person clicking send
  • 0send tools the assistant can reach
  • 1account — the AI key stays yours
  • 3–4named jobs, each with a written shape
  • 06:30it has already run by then
  • 6steps, and the last one is you
  • append-onlyhow every run is recorded

Three or four named jobs, not a personality.

Assistant products fail in one of two ways. Either they are so general that nobody can name what they do, or they are so eager that they send something embarrassing on your behalf. Both come from the same mistake: treating “assistant” as a character rather than as three or four named jobs with a written shape.

So this one is named jobs. Harvesting pulls from the places the work already lives. Triage sorts against written rules rather than vibes. Drafting produces text a person reads before it moves. One page shows what ran, what it produced and what is waiting on you. Nothing sends on your behalf — that is not a limitation we are working around, it is the reason this can touch a mailbox at all.

How it works

Six steps. The last one is a person.

Everything above the approval gate runs unattended. Nothing crosses it without a click.

  1. 01

    Scheduled trigger 06:30

    It runs before the day starts, on a timetable, whether or not anyone remembers it exists. A routine that has to be triggered is a task with extra steps.

  2. 02

    Harvest unattended

    Pulls from the places the work already lives — mail, calendar, sheets, the files you already keep. We read what you already record rather than asking you to record it differently.

  3. 03

    Triage written rules

    Sorts against criteria written down in advance, so the same input produces the same answer next week. A rule you can read is a rule you can argue with; a judgement you cannot see is one you have to trust.

  4. 04

    Draft never sends

    Produces the text and stops. Replies, summaries, the message you were going to write anyway — sitting in a queue with your name still on the send button.

  5. 05

    Log append-only

    Every run recorded, including the ones that found nothing. A run that found nothing and a run that died must never look identical — that is the failure mode that lets a broken routine sit unnoticed for a month.

  6. 06

    One page ran · produced · waiting

    Not a terminal. One page that says what ran, what it produced and what is waiting for approval, because a system nobody can see the state of is a system nobody trusts.

The rule that makes it safe

It drafts, queues and shows. A person clicks send. Every eager-assistant story ends the same way, and the fix is not a better model — it is a gate the software cannot open by itself.

A morning, end to end Six stages top to bottom: a scheduled 06:30 trigger, harvesting from mail, calendar, sheets and files, triage by written rule, drafting that never sends, an append-only log, and one page showing what ran, what it produced and what is waiting. Between the log and the page sits a refusal — a person clicks send. RUNS WHILE NOBODY IS WATCHING01 Scheduled trigger — 06:3002 Harvestmail, calendar, sheets, your own files03 Triagethe few that matter, by written rule04 Draftreply text produced, never sent05 Logappend-only — found nothing is not diedA PERSON CLICKS SEND06 One page: ran · produced · waitingA run that found nothing and a run that diedmust never look the same in the log.
Everything above the gate runs unattended. Nothing crosses it by itself.

What you get

How it actually runs, and on whose account.

Every morning

On a timetable

It runs unattended at a fixed hour. Nobody starts it, and nobody has to remember it exists.

Your key

Your own AI account

The client holds their own account and their own key, so the usage is visible to them and the cost is theirs to see. Nothing is locked to us, and we are not reselling tokens with a markup nobody can audit.

Week one

It will be wrong

A first version is tuned against real work for about a week. That is in the scope and in the price, because pretending otherwise only moves the disappointment later.

Anyone unwilling to hold their own AI account is a prospect we decline. There is no version of this we can sell them.

Want this one walked through against your own numbers? Email info@adapton.io and we will show you the parts that transfer and the parts that do not.

The limits

What it will not do.

The boundaries are structural rather than promised, which is the only kind worth writing down.

  • It removes tasks, not headcount. Anyone buying it to replace a person is buying the wrong thing.
  • It does not touch medical, legal or financial records. Regulated work is out of scope, not discounted.
  • It has no send tool. Not a disabled one — there is no reachable path from the assistant to an outbound message.
  • It will be wrong in week one, and the week of tuning against real work is the fix rather than the apology.
  • It runs on our own operation, daily. It is not yet running for a client, and we will say so until it is.
Whose account it runs on, and what is unreachable The client’s own AI account feeds a spend cap enforced in code, which sums the month and exits before the call. Below a refusal line sit two things the assistant cannot reach at all: a send button, and medical, legal or financial records. WHOSE ACCOUNT IT RUNS ONThe client’s own AI accounttheir key, their bill, their usage pageSpend cap in codesums the month, exits before the callNOT REACHABLE, BY CONSTRUCTIONA send button the assistant can pressMedical, legal or financial records
Below the dashed line is not “switched off”. It is not reachable.

Next step

Show us the thing you do every morning.

Fifteen minutes. We ask three questions and do the arithmetic with your real numbers. If it is not worth automating we will say so on the call.

Send us the details

All case studies · info@adapton.io