How it works
How automatic expense tracking actually works.
AI reads each message first, but it never gets the final word. A deterministic rules layer independently checks every fact before anything is stored — so a bad AI day never means bad data. This is what makes Numero's tracking totally automatic without being blindly automatic.
Quick answer
Numero's AI reads a message the moment it arrives, but nothing it produces is trusted outright: a deterministic rules layer independently re-checks every field, drops anything implausible, and takes over entirely if the AI step is unavailable. Only evidence that survives that check ever becomes a number you see — which is why automatic tracking here means checked automatically, not guessed automatically.
The pipeline
How does automatic expense tracking actually work?
Six steps between a message and a fact. Every stage is independently checkable — nothing skips straight from an AI guess to a number on your dashboard.
01
Discovery
Before anything else, a fully deterministic, AI-free pass looks at sender, subject, and attachment patterns to decide whether a message is likely to contain financial evidence at all.
02
Extraction
AI does the first read of the message, pulling out merchant, amount, date, and category. A deterministic rules layer then independently checks every field before anything moves forward.
03
Observation
What survives validation becomes an immutable record of evidence — never edited later to match a new interpretation, only ever added to.
04
Money Movement
A purchase only ever counts toward a spending total once amount, currency, direction, and status are all explicit — this is the one and only source of any number Numero shows you.
05
Activity
The evidence resolves into the object you actually see — one purchase, one subscription, one bill — kept up to date as new evidence arrives.
06
Watchers → Moments
Focused, read-only detectors evaluate confirmed Activities for one condition each, and the results reconcile into what actually deserves your attention.
A worked example
What happens when a $14.99 renewal email arrives?
Walking one real message through all six stages makes the pipeline concrete instead of abstract.
01 · Discovery
A message arrives from a known streaming-service sender with “Your subscription has renewed” in the subject line. Sender and subject pattern alone are enough to mark it Confident, before any AI involvement.
02 · Extraction
AI reads the message and returns merchant, amount ($14.99), currency (USD), and category (subscription). A deterministic check re-verifies each field independently — amount is a valid positive number, currency is supported, merchant resolves to a known profile.
03 · Observation
The validated fields are recorded as an immutable piece of evidence — this exact snapshot, never edited later even if a follow-up message changes the picture.
04 · Money Movement
Amount, currency, direction, and settlement status are all explicit, so this renewal qualifies as a real, countable Money Movement — it will actually appear in a spending total.
05 · Activity
The evidence resolves into (or updates) one Activity: this streaming subscription, now showing its latest renewal date and amount.
06 · Watcher → Moment
The recurring-cost Watcher picks up the renewal and folds it into the current monthly total — no alert fires unless the amount changed from last cycle.
A harder case:if that same email had shown an amount like “1.234,56” — ambiguous between European and US formatting — the deterministic check can't confirm a single correct value, so the amount field is dropped rather than guessed, and the Activity is held for review instead of silently posting the wrong number.
Why a bad AI day doesn't mean bad data
What happens when the AI gets something wrong?
The same validation, regardless of who read it
Whether a field came from the AI step or the deterministic fallback, it passes through the identical validation check before it can be stored. An implausible field — a malformed amount, a currency that doesn't exist — is dropped individually rather than accepted or rejecting the whole message.
Unknown Merchant beats a wrong guess
When the evidence resolves to nothing but generic words, Numero labels the merchant Unknown rather than presenting a confident-sounding guess. An incorrect merchant name is worse than a missing one — that discipline runs through every stage, not just this one.
No financial shape, no AI call
A cheap, local check runs before any AI call at all: a message with nothing that looks financially shaped is routed straight to the free deterministic path instead of spending an AI call — or a guess — on something that was never going to be evidence.
Evidence is never rewritten
Once a piece of evidence is recorded, it stays as it was recorded. A later, different interpretation adds a new record rather than editing history to match it — so what Numero believed at each point in time stays inspectable.
What's automatic today
What's actually automatic today?
Each Watcher is a focused, read-only detector for one condition — not a general-purpose alert system.
Recurring cost summary
Tracks subscription and recurring charges by merchant, so a monthly total is always current.
Return window closing
A heads-up before a return deadline passes, computed from explicit order and policy evidence.
Invoice due
Surfaces an unresolved bill within seven days of its explicit due date.
Invoice overdue
Flags a bill that's gone unpaid past its due date.
Evidence for warranties, insurance, travel bookings, and duplicate charges is organized and searchable today, but automatic proactive alerting for those is still being built, not a current guarantee — worth saying plainly rather than leaving it implied.
No bank login, either
The evidence comes from your inbox, not a bank connection.
Everything on this page runs on read-only Gmail and Outlook access — never a bank or card login. See exactly what permission that requires and how it compares to a bank-linked tracker.
Further reading
Frequently asked
The pipeline, in plain terms.
Is this OCR?
Not for email. Attached PDFs get their text pulled out with a plain, dependency-free text-layer reader — a scanned or image-only PDF simply yields no text, the same honest outcome as any unreadable attachment. Receipt photos uploaded in the app are read directly by an AI vision model instead, since there's no text layer to extract from a photo.
What happens if the AI misreads something?
Every field — no matter which step produced it — passes through the same deterministic validation before it's stored. An implausible or malformed field is dropped rather than saved, and if the AI step is unavailable entirely, a deterministic parser takes over automatically.
Does everything get turned into an automatic alert?
No — only four things today: recurring-cost summaries, return-window-closing reminders, and invoice due/overdue reminders. Evidence for warranties, insurance, travel bookings, and duplicate charges is organized and searchable, but automatic proactive alerting for those is still being built, not a current guarantee.
Why doesn't a rough guess ever get shown as a fact?
Because an incorrect answer is worse than a missing one. When evidence doesn't clearly support a specific merchant, category, or amount, Numero labels it Unknown rather than presenting a confident-sounding guess — the same discipline applies at every stage of the pipeline, not just merchant names.
How is this different from a simple email parser or a Gmail filter?
A filter or basic parser matches keywords and stops there — it can't tell a real renewal from a marketing email using the same words, and it has no validation step. This pipeline reads for meaning, independently re-checks every extracted field before storing it, and connects related evidence (a renewal, then a price change, then a cancellation) into one ongoing Activity instead of treating each message in isolation.
Is "totally automatic" the same as "fully trusted without checking"?
No — that's the specific distinction this page exists to make. Automatic means you don't have to enter anything by hand. It doesn't mean AI output is accepted as-is: every field is independently validated, implausible values are dropped, and Unknown is preferred over a confident wrong answer.
See it work on your own inbox.
Connect Gmail or Outlook in under a minute. Free to start, no bank login.