What this article is about
TaxItEasy doesn't just match bank transactions to invoices. It pays attention to what you do with those matches. Confirm a suggestion, reject one, unlink a match, or choose a category for a bank transaction, and the system remembers. This article explains exactly what is learned, what the learned memory is allowed to do (and, more importantly, what it is never allowed to do), and where your data stays.
One thing up front: learning starts from your live usage. It builds as you confirm and correct, and it is not pre-trained on anything.
Trusted vendors
The core of the memory is a per-company list of trusted vendors. A vendor becomes trusted after you've confirmed 3 matches for it, as long as you haven't been rejecting its suggestions too often (a high reject rate keeps a vendor out of the trusted list, and frequent rejections scale the effect back down).
Once a vendor is trusted, its future match candidates get a confidence boost of up to 12 points on the 100-point match score. In practice: a plausible-but-not-obvious payment from a vendor you've confirmed many times ranks higher and surfaces sooner than the same payment from a vendor the system has never seen you confirm.
Confirming a match also teaches the system new spellings: the counterparty name as your bank wrote it is saved as an alias for that vendor, so the counterparty signal recognises the vendor next time even when the bank mangles the name.
The guardrail: the trust boost never escalates a match
This is the most important property of the whole system. The learned boost never pushes a score across an automation boundary from below:
- A score below 70 can never reach the auto-link tier through the trust boost.
- A score below 90 can never become a silent auto-match through the trust boost.
- A match that is already strong gets no boost at all, because there's nothing to improve.
In other words: learning re-ranks and proposes. The trust boost never creates a link, and a suggestion cannot turn into a silent match because you've confirmed that vendor before. Spellings saved on a vendor (from confirmed matches, merges and close matches of scanned names) count in the counterparty signal like the vendor's own name, which is worth at most 15 of the 100 points, so learned signals alone never link a payment.
Unlinking a match or rejecting a suggestion counts against that vendor's trust; it does not block that payment and invoice from being paired again on a later run. See why this matched, and how to undo it.
Category learning for bank transactions
When you choose a category for a bank transaction, the system remembers it for that counterparty and money direction. After the same category has been chosen 3 times for a counterparty, it is suggested for that counterparty's open transactions that don't have a suggestion yet. Everyone in the company who can edit (every role except Viewer) and your tax advisor teach the same company memory, one at a time or in bulk; a bulk action counts once per counterparty. Accepting a suggestion counts as a choice too.
Suggestions only propose. On the web they appear as a chip above the category field reading "Suggested: {Label}" with an Accept button; on the phone the row shows the category as a one-tap chip. A suggestion is not booked until someone accepts it. If 3 or more different categories have been chosen for a counterparty, the system stops suggesting for it. A single different choice makes the newest choice win and restarts the count.
On bank transactions, a category is set without a click only by a matching rule set up in your company's account. Today our support sets rules up on request. Learned and AI categories are suggestions that you accept. Separately, a payment that matches an invoice with a high score is linked automatically. Links we are less sure of wait in Check our match, but they already count as paid until you reject them. You can undo any automatic link. When we read an invoice, the AI fills in its category. It counts in your books from the moment it is read, and you can change it at any time.
Two limits: a transaction that already shows a suggestion keeps it until someone categorizes it, so a newly learned category shows up on newly imported transactions. And the suggestion appears after the next processing run (a bank-statement import or the nightly run), not the moment you choose.
Categories set on invoices are not used for suggestions, neither on bank transactions nor on new invoices.
Precedence: Rules > Learned > AI
When the system decides what to suggest for a transaction, three sources are consulted in strict order:
- Your rules: a matching rule set up for your company always wins.
- Learned memory: your company's confirmed history comes next.
- AI suggestion: only for the long tail neither of the above covers, and low-confidence AI suggestions are dropped rather than shown.
You are always the top of the hierarchy: anything you've stated explicitly outranks anything the system inferred.
Everything stays in your company
- What the system learns from one company's decisions is used only for that company.
- It's deterministic counting, not model training, and no AI provider sees your confirmations.
- In the matching and category memory, a counterparty that isn't linked to a vendor record is stored only as a keyed fingerprint, never as readable text.
- When the company is deleted, everything learned for it is deleted with it, 30 days after you delete the company. When you delete your account as a company's only owner, its matching and category memory is deleted 30 days later; spellings saved on its vendor records stay with the company. See delete your account and export your data.
Related
- Why this matched, and how to undo it: the explanation behind every match
- How the matching pipeline works: the four signals and the tiers
- Creating a custom matching rule: the explicit layer that outranks learning
- Booking transactions for your clients: advisor actions teach the same memory
- Delete your account and export your data: what erasure covers