Paperwork arrives, checked data leaves. Your staff settle the exceptions.
Applications, supplier invoices, claims and correspondence reach you as scans, PDFs and e-forms. We build an intake that recognises each document, pulls out the fields you need, tests them against your rules and passes only the doubtful ones to an officer.
For Maltese public entities and companies that receive paperwork from residents, clients and suppliers in Malta and abroad, and need to show where every figure came from.
The system collects documents from all channels, recognises their type, reads out the fields and tests them. Any document with a doubtful field or a broken rule waits for an officer in a review queue. Checked data moves on to the business system, and each correction an officer makes is added to the test set.
Before and after
How the working day changes
Today
With the system
Today
Today: Keying data in by hand
Someone opens every scan, PDF or e-form and types names, amounts and reference numbers into the case file or the ERP.
With the system
With the system: Fields come in already structured
Each value is stored with the spot on the page it was read from, so the officer checks the figure and its source in one view.
Today
Today: Mistakes surface too late
A missing page or a wrong figure only shows up weeks afterwards, when a decision is appealed or a bank transfer is returned.
With the system
With the system: Checks happen on arrival
Completeness, matches against your registers and your business rules are tested before a file is opened, while there is still time to go back to the sender.
Today
Today: Simple and messy files cost the same
A complete, legible invoice ties up as much staff time as an unclear one with pages missing.
With the system
With the system: Staff time goes to the odd cases
Clear documents go through on their own. Officers work through a queue of doubtful and unusual items, each tagged with the reason it was held.
Today
Today: Nobody can say where a figure came from
If an auditor asks about a value, the reply depends on which colleague happened to type it.
With the system
With the system: Each figure can be traced
For every field the system keeps the source document, the reading, its confidence, the rule outcomes and any change an officer made.
Workflow
The workflow, step by step
Seven steps and a single exception route. Your process owners agree the thresholds and rules with us, and we tune them on your own documents before anything goes live.
Illustrative workflow
7 steps · 1 human checkpoint
Legend
System
Person
Human checkpoint
Exception path
Document intelligence with its exception route (illustrative)
Read the diagram as text
The main route leads from collection to recognition, reading and checking, then to the business system, and closes with the system improving from corrections.
Once the checks have run, a document with a doubtful field or a broken business rule is taken off the main route and placed in the review queue, where an officer accepts or corrects the values. From there it returns to the main route at the step that passes data to the business system.
The review queue is where a person steps in: no doubtful value enters the business system without an officer.
01
Step 1: Collect from all channelsSystem
Paper scans, email attachments, portal uploads and e-forms all enter a single intake queue, each with a file reference, details of the channel and a checksum. Structured e-invoices in Peppol UBL skip the reading step and move directly to the checks.
02
Step 2: Separate and recogniseSystem
Bundles holding several documents are broken into their parts. Each part is labelled by type and language, and that label decides which extraction schema and which set of rules will be applied.
03
Step 3: Read out the fieldsSystem
OCR that understands page layout reads the text, then a language model fills the agreed fields into a fixed schema. Every value comes with a confidence score and the place on the page where it was found.
04
Step 4: Test against your rulesSystem
The system compares values with your master data (clients, suppliers, VAT numbers and IBANs), applies rules that span several fields and looks for duplicates. Values above the agreed confidence level that clear every rule continue without stopping.
05
Step 5: Officer review queuePerson
The officer sees the original document next to the values read from it, with the broken rule or the doubtful field marked, and accepts or corrects each one.
Exception path: Doubtful field or broken rule· Returns to step 6
Human checkpoint
No doubtful value enters the business system unless a person has looked at it. The officer may also refuse a document and ask the sender to submit it again.
06
Step 6: Pass to the business systemSystem
Checked data creates or updates the file in the case system, CRM or ERP through its API, with the original document attached and a retention category assigned.
07
Step 7: Improve from correctionsSystem
Each correction is stored as a labelled example and added to the test set. A new threshold or prompt goes live only once the updated version clears that test set.
Components
The building blocks
01
Intake and channel links
Connections to mailboxes, scanners, portals and e-forms that bring every document into one queue, with a file reference and its details.
Layout-aware OCR combined with model-based reading into a versioned schema for each document type, recording a confidence score and page coordinates per field.
Deployment in your own cloud tenancy or chosen EU region, with encryption, retention categories and scheduled deletion that follow your records policy.
Master-data registers for clients, suppliers, VAT numbers and IBANs
Identity provider for officer roles
Controls
Controls built in from the start
Access boundaries
Officers see only the document types and files that their role permits. The extraction services hold no permanent access to the business system; each integration passes data on through its own narrowly scoped service account.
Human review
A low confidence score or a broken rule sends the item to an officer. Thresholds are set field by field rather than per document, and are lowered only when the test set shows that this is safe. Decisions on applications remain with your case officers.
Auditability
Per field we record where on the page the value was found, what was read, the confidence score, each rule outcome, the model and prompt version in use, and any officer correction with its reason code.
Data protection
We collect only the personal data the process needs, offer redaction for access requests and apply retention and deletion according to your records schedule. Processing stays in the region agreed with you, within the EU.
AI transparency
Each model and prompt version is documented together with its test results. Where an output affects people, the AI use is entered in your AI register and explained in the information given to applicants.
Measures
Indicators we would track
We settle these indicators with you in discovery, after first recording how your current process performs.
Indicators we would track
Metric
Why it matters
How we would measure it
01Pass-through rate
Why it mattersShows, per document type, how much routine work no longer needs an officer.
How we would measure itThe share of documents that clear every threshold and rule without review, taken from the workflow logs.
02Exceptions and their causes
Why it mattersReveals whether exceptions stem from poor scans, missing information or rules that should be revised.
How we would measure itQueue volumes broken down by reason code, channel and document type.
03Accuracy per field
Why it mattersThe quality figure that counts for later decisions and for payments.
How we would measure itPrecision and recall for each field on a labelled test set agreed with you, repeated with every release.
04Processing time
Why it mattersShows the effort saved and the effort that has shifted to review.
How we would measure itTime from receipt to hand-off and the active review time per exception, both read from queue timestamps.
No targets are set before a baseline exists.
Rollout
A phased rollout
Phase 01
Discovery and baseline
Two or three document types through a single channel, with today’s volumes and typical errors. We build a labelled test set from genuine documents with personal data removed.
Exit criteria
Fields, rules and quality targets agreed
Current performance recorded
Data protection impact assessment under way
Phase 02
Shadow pilot
The system reads and checks documents alongside the existing process; staff carry on as before and compare the results.
Exit criteria
Accuracy per field reaches the agreed targets
A threshold set for each field
Officers trained on the workbench
Phase 03
Supervised go-live
Pass-through is switched on for the document types from the pilot, and a sample of automatically accepted items is checked.
Exit criteria
Samples show no errors beyond the agreed tolerance
Runbook and monitoring transferred to your team
Rollback rehearsed
Phase 04
Extend
Further document types, channels and languages are added, and each must clear the same test gate before going live.
Permit, grant and benefit applications submitted with supporting documents, including those coming in through servizz.gov; requests for access to documents that need redaction first.
Delivery notes, certificates of conformity and customs papers for an export manufacturer, matched against purchase orders.
Illustrative example
Applications with supporting documents at a public entity
Situation
A Maltese public entity receives applications through servizz.gov, by email and over the counter, each with annexes such as ID copies, certificates and receipts, and staff copy the details into the case system.
System
A single intake recognises each annex, reads out the agreed fields and checks them against the application and the entity’s registers before a case file is opened.
Human control
Files with a missing annex, a doubtful field or a mismatch go to a case officer, who accepts, corrects or asks the applicant for a new document. The decision on the application stays with the officer.
What we would measure
Pass-through rate per channel and annex type, the causes of exceptions, and accuracy per field on a labelled test set.
Malta
Structured where possible, read where not
A Maltese organisation receives documents from residents and local firms as well as from suppliers and clients across the EU, each with its own layout. We build one intake and extraction layer with a schema and rule set per document type, so a Maltese application form and an Italian delivery note follow the same controlled route.
Government in Malta already accepts e-invoices in EN 16931 over Peppol, although suppliers are not yet legally obliged to send them; structured invoices of this kind bypass extraction completely. Where extracted data feeds decisions on people, we design for GDPR and the EU AI Act, whose Maltese oversight sits with the IDPC and the MDIA: documented processing, review by a person and logs your auditors can follow.
That depends on your documents, so we test it instead of promising a figure. From your own documents, with personal data removed, we build a labelled test set and report precision and recall field by field. Extraction models do get things wrong, and that is exactly why a doubtful field is sent to an officer rather than into the business system.
02What about handwriting, poor scans and documents in Maltese?
Handwritten forms and weak scans can be read, but with lower confidence, so a larger share ends up in the review queue. Each document is first sorted by language. Documents may reach you in English, Maltese or another language; for forms in Maltese we test reading quality on your own samples before the system relies on it.
03We already run OCR or a capture tool. Must it go?
Not as a rule. Where your capture layer does its job, we add checking, review and hand-off around it, and bring in model-based reading only for document types where it makes a real difference.
04Will the system decide on applications or claims?
No. It prepares and checks the data. Any decision with a legal or financial effect on a person is taken by your case officers, as Article 22 GDPR and your internal policies require.
05Where will our documents be processed and kept?
In the cloud region or data centre of your choice, with data kept within the EU. We describe every processing step, any external model service included, so that your data protection officer can assess it before go-live and answer questions from the IDPC.
06Where should a first project start?
With one document type that arrives in volume and follows clear rules, for example supplier invoices or a single application form. That is enough to prove the workflow, the review screen and the indicators before further types follow. See how we take projects into production.
Show us your document flows
Bring two or three document types and the systems they should land in. We will sketch a realistic first release for your organisation in Malta.