Natural language processing

These programs are built on a language model: a program trained on a large amount of text that understands what a sentence means and can write a reply. The model does not know your company, so we give it your documents, a glossary of your terms and limits: what it may do on its own and what it must pass to a person.

Document recognition turns the image of a scan into text and extracts the fields your system needs: invoice number and date, amounts, the parties to a contract, a deadline from a letter. The recognised fields go into your accounting system (ERP) or document workflow, each with a score showing how confident the program is.

Classifying and routing messages. The program reads an e-mail or request, assigns it a category from a list agreed with your departments and routes it to the right person. A message that fits two categories, or none, goes to a person.

Answers from company documents. An employee asks a question in an ordinary sentence, the program searches internal rules, contracts and procedures, and the answer cites the document, version and clause. When the documents do not contain an answer, the program says so plainly. An agent that prepares a decision in a process on this basis is a separate scope.

Summaries. A long letter or contract from a counterparty is summarised as a list of points: what the other side proposes, what has changed and what needs a decision, with a reference to the original.

What these have in common is a queue of uncertain cases: a result below the confidence threshold agreed with you is checked by an employee, and their correction stays in the log and is used to improve the program.

Before the program receives real documents, we measure its quality on a sample from your company: different document types, different suppliers, poor scans and every language your documents come in. Your employees prepare reference answers for the sample. We compare the program's output with the reference field by field: where it matches, where it is wrong and where it flagged the case as uncertain itself. On that basis we agree with you the confidence threshold and the order of implementation.

1. We collect a sample. Together with you we choose the types of documents and messages and take a sample from recent months, and your employees prepare reference answers. Result: a sample and a reference for measuring quality.
2. We measure quality. We run the program on the sample, compare the output with the reference and go through the errors with the people who read these documents today. Result: a sample report and a decision on which types we implement.
3. We connect the program to the system. We start with one document type or one inbox; results go into your system and uncertain cases into the queue. Result: the program in the daily work of one department.
4. We hand the program over to your team. We teach the person responsible to read the queue and the log, and to add categories and new versions of internal rules. Result: the program looked after by your team; measuring quality after launch is a separate scope.

1. On what sample of our documents will recognition quality be measured before implementation, and who will prepare the reference answers?
2. What happens to a document the program is unsure about: who will see it, and does the correction go back into the program?
3. Does every answer to an employee cite the document and clause, and what does the program reply when the documents contain no answer?
4. Where are our documents and personal data processed: in the European Union, on our premises or at an external language model provider, and does that provider train on our data?
5. How will the program cope with an invoice in a new layout, or a letter in a language that was not in the sample?
6. Who on our side will add a new message category or a new version of the internal rules without the contractor's involvement?

The answers for your company depend on which documents and messages pass through it, so the first step is a process analysis, which sets the scope and the order of implementation.