How AIFT Learns From Your Corrections
A correction you make today becomes an example the AI follows tomorrow. This is about what that means in practice, because a matched example does not always do the same thing.
What A Correction Becomes
You do not have to do anything to build this up. Whenever someone changes a category the AI picked, the change is recorded as an example: for this partner, or this wording, the right answer is X.
Filling in a value the AI left empty counts too. A human supplying an answer the AI missed is at least as strong a signal as a human overruling it.
Four things are deliberately not learned:
- a line whose category came from a partner rule rather than from a person,
- anything the pipeline itself wrote,
- a "no invoice needed" mark with no specific transaction type,
- a record with neither a partner name nor a description to key on.
What A Learned Example Then Does
This is the part worth understanding, because there are two very different outcomes.
It can book the line outright, with no AI call at all. On invoice and cash categories, a matched example is applied directly, at very high confidence, and the record is marked as having come from a correction rather than from the AI.
Or it can simply ride along as a strong hint. On the other surfaces, and whenever the direct match does not apply, the example is put in front of the AI as an example and the AI still decides.
Two conditions govern the direct route, and only for captured examples:
- The amount has to be in the neighbourhood. A captured example remembers the range of amounts it was learned at, and only books by itself within roughly four times that range.
- Never across currencies. If both currencies are known and differ, the direct route is refused.
Outside those bounds the example is demoted, not discarded: it still rides the prompt as a hint.
A hand-added example has no amount range at all, so it books unconditionally, at any amount and in any currency. That is the sharp edge of pre-loading examples for a new client. It is often what you want, but it is stronger than what a captured example does.
The Four Surfaces
They do not behave alike, and the differences matter more than the similarities.
| Where | What the example sets | Books on its own? | Matches on the partner? |
|---|---|---|---|
| Accounting categories → Examples, Invoice examples | the invoice category | yes | no, description only |
| Accounting categories → Examples, Cash examples | the transaction category | yes | yes, then the description |
| Bank Transaction Types → Examples | the transaction type, and whether an invoice is needed | no, it only informs the AI | yes |
| VAT codes → VAT Examples | the product/service VAT group | no, the line still goes to the AI | only to break a tie |
A few consequences of that table:
- A partner-only example is useless for invoice categories and for VAT, because those match on the description. Give those a description to match on.
- Transaction types have no direct route at all. The deterministic layer there is the separate pattern list on the transaction type itself, not these examples.
- VAT examples supply one element out of five. The rate and the calculation method still have to come from somewhere, so the AI is always called. What gets replayed is the group, not the resolved code, deliberately, so a code learned before a rate change is not applied after it.
Custom category lists are not a fifth surface. Corrections there are recorded but nothing reads them.
Reading The List
| Column | Meaning |
|---|---|
| Partner | The partner the example matches on, where that surface uses it |
| Description | Text the record should contain |
| Category | What to apply. Reads Transaction type or Product / service group on those surfaces |
| Resolved code | VAT examples only, read-only, showing what the group resolved to |
| Verdict | Transaction-type examples only: whether an invoice is needed |
| Source | Auto-learned, or Manual if someone added it |
| Times seen | See below |
| Last updated | When the example last changed |
Times seen does not mean the same thing everywhere. On transaction-type and VAT examples it counts how often the pattern recurred. On category examples it counts how often that same record was corrected again, so a pattern recurring across different invoices produces several separate examples sitting at one rather than one example with a high count.
A hand-added example starts at three, deliberately above the noise floor of a single organic correction.
You can search by partner or description, filter by category, and filter by source. The list is ordered by how often an example has been seen.
Adding, Editing And Removing By Hand
+ Add example teaches the system something before it has seen it organically, which is what you want when onboarding a client.
Two behaviours to know:
- On transaction-type and VAT examples, adding one that matches an existing pattern converts the learned row to Manual and keeps its accumulated count, rather than creating a second row.
- On category examples there is no such merge, so two identical hand-made examples can coexist.
Delete removes an example for good. The AI stops using it on the next run, and if the same thing is corrected again later a fresh example appears by itself.
Nothing here re-processes what you already have. Adding, editing or deleting affects future runs only. To revisit existing records, use Re-categorize on the Invoices or Bank transactions list, which leaves anything a person set alone.
One thing to watch: an example pointing at a category you have since archived can still be applied on the cash side without being flagged for review. Delete examples whose target you have retired.
How Much The AI Sees
Behind the Settings button on an Examples tab there is a single dial called AI learning, running from Off up to 100 in steps of five, at 30 by default.
It governs all four surfaces at once. Turning it down to control cost also gives the AI fewer transaction-type and VAT examples to work from. More examples can improve accuracy but cost more per request, and very long lists give diminishing returns, so 30 is a reasonable starting point.
Off does more than stop the examples reaching the prompt. It also switches off the direct-match shortcut, so nothing books by itself from a correction either.
Changing the dial needs the "Edit workspace settings" permission.
Who Can See And Change This
The Examples tabs are accounting-side only; client users, including a Finance admin, see none of them.
Adding, editing and deleting examples needs "Edit master data". That is a permission rather than a job title, so it can be granted or withheld per member. See Roles and Permissions.
Examples never cross workspaces. A correction you make for one client is used for that client only, and nothing about it reaches another. (The company register is the one thing in the product that is deliberately shared between workspaces, and that holds public company data, not your corrections.)
Related
- Categorising Invoices & Transactions - how a category is assigned in the first place
- Transaction Types & No Invoice Needed
- VAT master data - the VAT Examples tab in its own context