Make your agents Claude ChatGPT do more with Business brain.

Nothing has to be prepared first. Old exports, several people's mailboxes, a folder nobody ever named — that is the normal starting condition, and it is what we are for.

It reads your mail the way a backup does. A message is never marked read, never moved, never deleted, and nothing on your mail server changes. Documents come from the folder you already keep: offers, invoices, drawings, spreadsheets, read down to the page.

On the one system running today this happens every ten minutes, from a small always-on machine. On a single-computer build, the same job runs on the one computer and nothing ever leaves the building except the step below.

Every number and every date has to be literally present in the message before it is stored, and it is stored with a verbatim quote of the line it came from. Anything not stated is left empty. Nothing is inferred, rounded, or converted between currencies.

A fact that does not fit the contract is quarantined rather than guessed at. That is the whole reason an answer can be checked in one click instead of taken on trust.

This is the step in the pipeline that sends text out for processing. It runs on your own key, as a bounded job you authorise — not as something ticking away in the background — and it is the most heavily gated part of the system.

Keyword first. Never stale, and it reports the true total rather than the first page of it.

Then meaning, so a thread that never used your words still surfaces. That index is built on your own machine. Nothing leaves it to make it work.

Then the same question in both languages. A business that trades across a border writes its archive in more than one, and a question asked in only one of them misses the thread that actually answers it. When both phrasings surface the same thread, that agreement is itself a signal.

Then the text itself, in true date order, forwarded originals included. Everything is read-only and capped, and it flags itself when it has been truncated rather than quietly returning less. Transcripts are searched word for word rather than by meaning — deliberately, so a search for what you actually typed finds it.

A decision settled in one conversation is filed the same turn and quoted back in the next, instead of being re-decided from scratch a year later. A fact that comes up twice gets promoted out of the daily log into a note of its own.

Names never move once they exist. What was filed last year is still where you left it this year, which is the whole difference between a memory and a pile.

The conversation itself is kept as well, word for word. You can go back to what was decided in March and pick it up where it stopped — including with a different person at the keyboard.

Not our rules. Yours — the part a general AI tool cannot give you, because nobody has written them for your business. Every rule on the live system was added after a real mistake, which is why the file is worth more than the code around it.

Nothing is learned inside the AI. The model does not change. The rules file, the approved notes and the transcripts are files on your disk — which is exactly why pointing a different assistant at the same folder loses nothing.

And nothing the system writes counts as fact until a person has read it and said so. Generated and approved are two different states, and the file records which one a note is in.

Ask a general assistant for a price and it will produce a plausible number. Ask this one, and if the price is not in your records the answer is "to be confirmed". We would rather it be useless on a question than confident and wrong.

Dates and durations are computed rather than recalled. How long a client usually takes to reply is arithmetic over real timestamps in the archive, not an impression of them.

It also knows the difference between "no" and "I have not read that far back yet", and it is required to tell you which of the two you are looking at. An unread archive must never read as an empty one.

There is no mail-sending code anywhere in it. It cannot email a customer, because the capability does not exist rather than because it has been switched off.

The connection to your mail is read-only in the strict sense: nothing is marked read, moved or deleted, and the mail server itself is never altered.

The database is opened read-only and there is no tool to write to it. It could not change your records if you asked it to.

The owner. The person who has to remember what was promised in 2018, and currently does it at ten at night.

The staff. Sales, procurement, whoever answers the mail. Same archive, same cited answers, same rules — and one provider seat per named person, never a shared login.

And not only the running work. Throw a half-formed idea at it. It gets filed where it belongs without disturbing anything else, and it is still there next month.

Copy the folder and you have copied the system. Any software engineer can pick it up — plain files and a database with a written schema, not a proprietary format and not an account on a platform of ours.

It is not locked to one AI either. The brain is your working folder, so pointing a different assistant at it is a configuration change rather than a migration. Nothing is stored inside the assistant, so nothing has to come out of it.

Vendor marks say what the system connects to. Not a partnership, not an endorsement.

Claude is what the one live deployment runs on. The others are compatible with the same folder — which is a property of plain files, not something we have demonstrated for you. We will say so until we have.

It ages well, too. A folder-and-database build moves onto whatever gets better next. That is what a plain file is for. A habit of pasting things into a chat window every day does not migrate anywhere — it stays a habit, and it takes nothing with it.

It works on the live mailbox. A decade of exported mail, and the documents sitting beside it, are not in the mailbox any more — and that is where the expensive questions live.

And it does not carry your rules. No approved price table, no "never invent a number", no file you can edit the day it gets something wrong.

If the question is about last week, use the tool you already have. This is for the other questions.

Small. A few thousand messages is a weekend of tidying and the same steps as everything else.

Large. Past a few hundred thousand it stops being a search problem and becomes an indexing one, and that is the part the quote turns on.

Either way. Same rules file, same read-only access, same refusal to send. Scale changes the price, not the behaviour.

A paid plan gives each named person an allowance that resets on the provider's schedule. Providers can apply session, weekly or other limits, so we choose the account and tier against real use rather than promise an unlimited seat.

Metered API usage is separate. The pipeline uses it to collect, extract and index the archive, then to process new material after syncs. Every call and charge appears on the account you own.

Prices and limits can change. The durable part is the split: your files and database stay yours, everyday reasoning runs on your provider account, and pipeline usage has a meter you can see.

You buy it, you hold it, you cancel it. Business deployments default to the provider's business contract. Consumer tiers are written for individuals, and we do not build around a shared login.

Which tier suits you depends on how hard the business leans on it, and that is a conversation with your own usage in front of us rather than a number we put on a web page.

It is also why swapping assistants later is a configuration change rather than a migration. The brain is a folder on your disk. The plan is just what reads it.

The initial backfill is a bounded job you authorise. After that, the same pipeline keeps collecting and processing new material. Its ongoing cost moves with volume, model and freshness, and the written scope includes an estimate.

Questions, reasoning and long working sessions can run on each named person's plan. Pipeline calls remain visible on your own API meter rather than hidden inside a KUUP licence.

If you would rather run the whole thing on an API key instead, it does that too, on your own account, with no rebuild. It is the same code and the same data either way.