AI Transaction Categorizer
How your agent works
A short guide you can keep, and hand to anyone who takes over your books.
The one thing to know first: the agent has never written anything to your
books, and it will not until you and I turn that on deliberately, together.
Everything you see on your dashboard today is a proposal, waiting for
your yes.
And when we do turn it on, nothing moves without you. Writing is off right
now. When you decide you want it, it works like this: each proposed transaction
gets a checkbox, you tick the ones you agree with, and the dashboard shows you a
summary and asks for your password before anything is sent. Anything that does get
written keeps an Undo next to it. So the agent never decides on its own —
it asks, one transaction at a time, and you can always put it back.
What it does, once a day
Every morning at 8:00 your time, on its own, without anyone
starting it and whether or not any computer is on:
- It reads your chart of accounts, so it can only ever suggest accounts you
actually use. It cannot invent a category.
- It reads the transactions sitting in an uncategorized account.
- For each one it decides an account and gives itself a confidence
score between 0 and 1.
- Anything scoring 0.75 or higher is proposed with that account.
Anything below goes to your review queue instead, with the reason it hesitated.
- It publishes the result to your private dashboard.
It takes about 1.4 seconds per transaction. A month of roughly
400 takes about ten minutes.
Reading your dashboard
agc-quickbooks-ai.pages.dev/report — you sign in once and
it remembers you for 90 days. It works on your phone.
| What you see | What it means |
| Human Review Queue | The ones it refused to guess, each with its reason. This is the part worth your attention — it is short by design. |
| Confidence | How sure it was. Below 0.75 it does not propose an account at all. |
| By Rule | Decided by a saved rule rather than the AI. Instant, free, and perfectly consistent. |
| Dry run | Confirmation that nothing was written to your books. |
| Manual Work Saved | An estimate, at 1.5 minutes per transaction categorized by hand. |
If the dashboard looks empty or old, it is telling you the truth, not breaking.
It reports what was in your books at the moment it ran. If a run cannot finish, the
page says so outright rather than quietly showing you yesterday's figures as if they
were today's.
Correcting a categorization
When the agent proposes an account you disagree with, tell me the transaction and
the account it should have used. That correction is saved as a permanent rule — this
payee always goes to this account — and from then on it is applied instantly, with no
AI involved and no chance of a different answer next month.
This is worth doing early rather than late. A correction you make in January is
applied to every matching transaction after it; the same correction made in December
leaves eleven months of inconsistency behind it.
Nothing leaves that page without your password. Even once approvals are
switched on, the dashboard will not change an account just because a button was
tapped — it asks you to confirm with your password first. That is deliberate: a page
that can edit your books is a page that can be left open on a phone.
What it does not do
- It does not reconcile. Categorizing decides which account an
expense belongs to. Reconciling matches your bank statement against your books and
marks things cleared. They are separate jobs, and this agent does the first one.
- It cannot see the Banking → For Review queue. Intuit does not
expose the bank feed queue through its API, so no outside app can read from it or
accept items out of it. What the agent works on is transactions already recorded in
your books.
- It does not guess. When it cannot tell, it says so and stops.
That is the whole design — a wrong category filed with confidence is worse than an
honest question.
Stopping it
You are never locked in, and you do not need me to stop it.
- To revoke its access entirely: in QuickBooks, gear icon →
Settings → Apps, find AI Transaction
Categorizer, and disconnect. One click. It loses all access immediately.
- To sign out of the dashboard on a device: sign out.
- To put it back to proposals only (once approvals are on): tell
me, and I switch writing off for your company. It takes seconds, it does not need a
new version of anything, and the agent goes back to only suggesting. Your existing
reports are unaffected.
Writing is off today. If you ever want it on, that is a change we make together —
and it still asks for your password on every batch, so "on" never means "on its
own".
What it costs to run
The dashboard and the scheduled run sit on free tiers — not trials, permanent free
tiers. Today they run under my accounts, which is what let me set this up without
asking you for logins. Your data is yours: exportable at any time, and if
you ever want the whole setup moved to accounts in your own name, say so and I will
hand it over.
The only real cost is the AI processing itself, billed per use, and it falls as more
of your vendors become saved rules that never touch the AI again. At your volume that is
a few dollars a month at most — and for now it comes out of what you have already paid,
not out of your pocket.