Technology
Machine Learning
Domain-specific models for classification, extraction and matching, measured on your documents rather than a benchmark.
The concept
Models trained for one job each, measured on your documents
Models that each learned one narrow task from labelled examples. None of them is trusted until it has been measured on documents of yours.
Machine learning is the part of the platform that learns patterns from examples instead of following instructions someone wrote. In PTAS AI products it does four kinds of work. Classification settles what a document is. Extraction models find fields and table rows on layouts they have not seen before. Ranking models propose which vendor, item or purchase order line something most likely belongs to. Anomaly models flag what looks unusual for this vendor or this account. Each model has one job, so each can be tested and replaced on its own.
Every answer comes with a confidence score, and the score is often more useful than the answer. It is attached per field: the invoice number can score high on the same page where a handwritten quantity scores low. Thresholds on those scores decide what proceeds and what waits for a person. A score earns that role only if it is calibrated, meaning fields turn out right about as often as their scores say, so a low score really does mark a doubtful value. We check that on a sample of your documents, field by field, before a single threshold is set.
The models also learn from the people who correct them. A correction in the review queue is kept as a labelled example: the field, what was read, what it should have been and the page region it came from. Those examples feed the next training round, and the new version is tested against the current one before it replaces it. Nothing retrains itself in production overnight. A model that changed without a test would make yesterday's thresholds meaningless.
Where it sits
Trained models across the five stages
PTAS AI workflows run as Perception, Intelligence, Orchestration, Action and Verification. Trained models do most of their work in the second stage, and their scores travel into the third.
In the products
What the models do inside DocPro and FinSuite
In DocPro the models read and score. In FinSuite they rank and flag, and a rule decides what follows.
DocPro
Document classification
Tax invoice, proforma, credit note, delivery challan, bank statement, KYC document or a custom type, settled before any field is read, because the type picks the schema and the checks. A close call goes to a person.
DocPro
Field and table extraction
Models trained across many layouts find the invoice number by context, wherever it is printed, and keep line rows together across pages. A new supplier's first invoice is read without a template.
DocPro
Confidence per field
Each value carries its own score. Thresholds are set per field and document type, so the grand total and the GSTIN can demand more certainty than a remarks line.
FinSuite
Match candidates ranked
The invoice says Housing, bearing and the PO says BRG-40-HSG. A model ranks the candidate PO lines with a score; the tolerance check that follows is a rule.
FinSuite
Coding suggestions
Ledger account and cost centre proposed from the line description and your prior postings to that vendor, with a score and the evidence attached to the draft entry.
FinSuite
Anomaly flags
An amount far outside this vendor's history, a bank detail change followed by a large invoice, a run of invoices just under an approval band. Each is flagged for a person with the pattern named. A flag never blocks or pays anything by itself.
Worked example
One file classified, then a month of corrections put to work
A scanned PDF holding three documents, classified page by page. Then what became of the corrections reviewers made on one vendor's invoices over a month. Values are illustrative.
One PDF, three documents. Illustrative example
Classification is scored like any field. When two types are close, a person settles it before the wrong schema is applied.
Type before fields
Schema, checks and thresholds all depend on the type, so the type is settled first.
Close calls go to a person
When two types score close together, a reviewer picks before any field is read.
Bundles split by score
Page boundaries are predicted and scored too, so one PDF can become three records.
Over the following month, reviewers settled the fields the models were unsure about on that vendor's invoices. Each correction was kept. Read together, they say different things: some are examples the next model version should learn from, one is a pattern a rule should handle, and one is a threshold set higher than it needs to be.
One vendor, one month of corrections. Illustrative example
Thirty-eight confirmations of the same field is evidence that its threshold can come down. Eleven identical corrections is a rule nobody has written yet.
Every correction labelled
Field, value read, value corrected and page region, kept together as one example.
Patterns become rules
The same correction for the same cause is proposed to the rule owner as a rule.
Released after a test
A new version runs on held-back documents of yours before it replaces the current one.
Limits and controls
What a trained model cannot tell you
A model is only as good as the examples it learned from. A supplier who redesigns their invoice, a scanner that starts producing darker images, a document type nobody labelled: each shows up as falling scores before it shows up as errors, and only if someone watches scores by source. One average across the whole queue hides it.
A model also cannot explain itself the way a rule can. It can say how sure it is and point to the region of the page it read. It cannot cite a policy. That is why its answers go into rules as inputs and never stand as decisions. And a model measured on someone else's documents says little about yours, which is why the measurement that counts is done on a sample of your own, and why you can test a document before any conversation about thresholds.
Measured on your sample
Before go-live, a sample of your documents is read by the models and checked by hand, field by field. Thresholds come from that measurement, per field and document type.
Calibration watched
Scores are compared with reviewer outcomes and sampled checks over time, by vendor and document type. A score that stops meaning what it meant is raised as drift.
Versioned releases
A retrained model is a new version, tested against the current one on held-back documents before release. The previous version stays available to roll back to.
Model version on the record
Each extracted field records which model version read it, so a question about last March's invoice is answered with last March's model.
The queue where corrections are made is described under Human-in-the-Loop. How a new model version is approved for release is under AI Security and Governance.
FAQ
Questions evaluators ask about machine learning
What accuracy should we expect?
We do not quote a figure on this page. Accuracy depends on your document types, your suppliers, the quality of your scans and which fields you count, so a number measured on someone else's documents tells you little. What we do is measure it on yours: a sample read by the models and checked by hand, field by field, during the assessment or onboarding. You see the result for each field before anyone decides where a threshold sits.
Do the models learn from our corrections automatically?
They learn from them, but not on their own in production. Each correction is stored as a labelled example with the page region it came from. Those examples go into the next training or calibration round, the resulting version is tested against the current one on held-back documents, and it is released through change control. Where a correction reflects a pattern rather than a misreading, such as a supplier who prints the PO number in the remarks, it becomes a proposed rule instead.
Are our documents used to train models for other customers?
How your documents and corrections may be used for training, and whether that use stays within your own deployment, is set out in your agreement, and we do not go beyond it. In a private cloud or on-premise deployment, where supported, the labelled examples stay inside your boundary along with the documents.
How is this different from the language models on the generative AI page?
They are different tools. The models on this page are trained for one task each and return a label or a value with a score: this is a credit note, this is the invoice number, this PO line is the likeliest match. A language model reads and writes free text, and its output has to be forced into a schema before anything can use it. Both feed rules, and neither decides alone.
Keep reading
Where the models connect
Document AI
The reading stack these models sit in: OCR, layout, extraction and validation.
Read more →Computer Vision
The models that work on the image before any text is read.
Read more →Human-in-the-Loop
The review queue where corrections are made and kept.
Read more →Generative AI
Language models: where they read and draft, and where they are kept out.
Read more →Next step
See the scores on your own documents
Upload one document and read the confidence on every field. For a proper measurement, bring a sample and we will check it by hand with you, field by field, before any threshold is set.