What Training an AI on Your Own Documents Actually Involves

The demonstrations make this look like pouring a folder into a machine. The reality has more steps, most of them dull, and it is worth knowing which ones will need you.

This is what the work actually consists of.

One: deciding what is in scope

Before anything technical, somebody has to decide which material is included. Not everything should be. Material belonging to other people, material you hold under an obligation, material that is simply enormous and irrelevant — each is a reason to leave something out.

This step is entirely yours. Nobody else can decide what you are entitled to use or what matters to you. It is usually an afternoon of thinking rather than a week of work, and skipping it is the most common way this goes wrong.

Two: getting the words out

Documents have to become text. Typed files already are. Scans have to be read, which works well on clean pages and badly on photographs taken at an angle. Spreadsheets need their structure understood, because a number without its column heading means nothing.

This is where archives reveal themselves. Nearly everyone has more unreadable material than they expect — the decade of receipts photographed hastily, the contracts that exist only as faxes of faxes. Some can be salvaged. Some genuinely cannot, and being told so early is better than discovering it later.

Three: resolving the contradictions

The hardest step, and the one that cannot be automated.

Real archives contradict themselves. Four versions of an agreement. A figure recorded differently in two places. A policy superseded in an email nobody filed. A system trained on all of it learns every version equally and will later tell you whichever it reaches first, with complete confidence.

So somebody has to decide which version is authoritative. That is a judgement about your affairs. We can find the contradictions and show them to you — that is mechanical. Deciding is yours.

Four: the first training run

Only now does the part everyone pictures happen. The prepared material is used to train, on the machine in your building, and what comes out is a model shaped by your world.

It takes hours rather than minutes, and it is uneventful. Nobody needs to watch it.

Five: finding out what it got wrong

The step people forget. A newly trained assistant is confidently wrong about some things, and the only way to find out which is to ask it questions you already know the answers to.

This is why the demonstration matters. You are not being shown a trick; you are checking whether it understood your material, using questions where you can grade the answer yourself. Where it is wrong, that usually points back at step three — a contradiction nobody resolved, or a document that meant something different from what it appeared to mean.

Six: keeping up

Your material does not stop arriving. New correspondence, new documents, new decisions. An assistant trained once and never again is slowly becoming a record of the past.

So retraining is part of the service rather than an event — it happens as your material grows, and it is our job rather than something you schedule.

What this costs you in time

Being honest about the only currency that matters here:

  • Deciding scope: an afternoon, yours, unavoidable.
  • Resolving contradictions: a few hours, yours, spread over the work.
  • Checking the demonstration: an hour, yours, and worth taking seriously.
  • Everything else: ours.

Call it a day of your attention, not spent in one sitting. Anyone promising less than that is either doing less than this or intending to decide things on your behalf that they should not be deciding.

Where it starts

Not with a year and not with a machine. It starts with finding out what your material actually supports, in writing, before anyone asks you to commit — which is exactly what the Archive Assessment is for.

How an assistant uses what it has learned →

Local AI. Private data. Local training.

$8,995 the first year — everything included. $4,995 each year after. The first year costs more because it contains the Jetson Orin placed and configured, your archive loaded and the first training run; every year after is the service running.

Begin the first year — $8,995

Or add the Archive Assessment first — $1,950, credited in full

You order and pay at checkout; we write back within days — a decline returns every dollar, and until our letter confirms the year you may withdraw in writing. From that letter the year is final, and it arrives on the date the letter names. Prefer to write to us first?

One decision, and then it is handled

One price, and one way to begin.


The same assistant at either address, for the same price. Begin with the first year; if you like, add the Archive Assessment before it — delivered work you keep, credited in full.

The first year

$8,995

Everything included. $4,995 each year after.

Begin the first year — $8,995

Or add the Archive Assessment first — $1,950, credited in full

The first year costs more because it contains the Jetson Orin placed and configured, your archive loaded and the first training run; every year after is the service running. You order and pay at checkout; we write back within days — a decline returns every dollar, and until our letter confirms the year you may withdraw in writing. From that letter the year is final, and it arrives on the date the letter names.

We answer in writing, within one business day — and if we cannot serve you well we say so before our letter confirms a year. Prefer to write to us first?