AI Software / How the AI works
AIHow the AI works
Real AI, described exactly and checkable on your own screen
Every AppIQ AI product runs two kinds of model inside your browser: a language model that reads what each line says, and a machine-learning model trained on your own file. This page says what each does in each product, how we count what the AI found, and what the AI never does.
What the AI does
1. Reads meaning
A language model (all-MiniLM-L6-v2, Apache 2.0 licence, about 34 MB) turns each description into a vector of meaning. Two wordings of the same thing land close together, so "BRITISH AIRWAYS 1252" and "Flight to London" can be recognised as the same purchase. It runs in your browser with ONNX Runtime Web.
2. Suggests
For lines nothing could place, the same model finds the closest candidate by meaning and shows it with a confidence, such as "Probably VanArsdel Ltd (68%)". Suggestions are shown only at 50% or more and are never applied for you.
3. Spots the unusual
An Isolation Forest, a machine-learning model, is trained on your own file every time you run it. Lines that are easy to set apart from the rest score high, and each comes with the reasons in plain English, such as "6.3× this employee's usual card line". Nothing is pre-trained on anyone else's data.
What the AI does in each product
| Product | Language model | Suggestions | Machine learning |
|---|---|---|---|
| AI Bank Match Pro | Matches bank and ledger lines worded differently | Likely counterpart for open lines | Unusual bank and ledger lines |
| AI Supplier Match Pro | Matches statement and ledger lines worded differently | Likely cause for open lines | Price differences, duplicates, outsized invoices |
| AI Cash Flow Tracker Pro | Joins payee spellings; categorises payees by meaning | Recurring streams found through the joins | Unusual payments and recurring changes |
| AI Customer Payment Match Pro | Recognises customers written differently and applies receipts | Probable customer for unplaced receipts | Unusual receipts and customers |
| AI T&E Card Match Pro | Matches card wording to what employees typed; categorises spend | Likely category and claim for unclaimed lines | Unusual card lines; risk by employee |
| AI Intercompany Match Pro | Matches entries with no shared reference; classifies types | Likely counterpart for one-sided items | Unusual intercompany items |
| AI Journal Entry Tester Pro | Flags descriptions that don't fit the account they were posted to | The account the description fits better | Risk score for every journal entry |
| AI Balance Sheet Rec Reviewer Pro | Reads every reconciling item; flags vague and wrong-account items | The account a misplaced item fits better | Risk score for every reconciliation; review queue |
| AI Vendor Master Audit Pro | Finds the same supplier set up under different spellings | Which duplicate record to keep | Risk score for every vendor |
| AI Trial Balance Mapper Pro | Proposes a group account for every unmapped local account; checks existing mappings | Proposals below 65% confidence, for you to decide | Unusual balances and movements |
How we count what the AI found
Each product runs your files with the rules only, then with the rules and the AI. "AI matched 6 pairs the rules missed" is the difference between those two runs, worked out on your computer every time.
AI matches carry an AI badge and the similarity score. Suggestions say "Probably" with a confidence. Anomaly scores list their reasons.
Every product shows the model used, that it ran on your device, how many descriptions it read and what it added, on that run.
If your browser can't load the model, the product says so and shows the rules-only results. It never claims AI results it didn't produce.
What the AI never does
No upload, no AI service, no bank or system connection. The tool pages block all outgoing connections; turn off your Wi-Fi and they keep working.
Nothing is posted to your accounting system and nothing is signed off. You review the results.
A similarity is evidence, not proof, and an unusual line isn't necessarily wrong. That's why the scores and reasons are always shown.
The products support your review, audit and sign-off; they are not accounting, audit or legal advice.
Questions
Is it really AI?
Yes. Two kinds of model run in your browser. A language model, all-MiniLM-L6-v2, turns each description into a vector that captures its meaning, so two wordings of the same thing can be recognised. An Isolation Forest, a machine-learning model, is trained on your own file each time you run it and scores how unusual each line is. Neither is a list of keywords.
How do you know what the AI added?
Each product runs twice on your files: once with the rules only and once with the rules and the AI. The numbers you see, such as 'AI matched 5 card lines the wording rules missed', are the difference between the two runs. They are not estimates and not typed in.
Does the AI make decisions for me?
No. Matches the AI makes are marked with an AI badge and the similarity, so you can see and undo them. Suggestions are only suggestions and are never applied. Anomaly scores are prompts to look, not findings. You review and sign off.
Is my data sent to an AI service?
No. The models download from appiqfinance.com as files and run inside your browser tab. The tool pages carry a security rule (connect-src 'none') that stops them sending data anywhere, and you can prove it by turning off your internet connection: the products keep working.
Can the AI be wrong?
Yes, like any model. A similarity score is evidence, not certainty, which is why every AI match shows its score, doubtful pairs go to a To check list, and the reasons for every anomaly score are shown. If the model can't load on your computer, the product says so and shows the rules-only results.