AI & Document Intelligence
We build AI into specific document tasks: pulling fields out of invoices and forms, answering questions from your own files with source links, and routing incoming work. Each is measured on your documents before launch.
AI and document intelligence means applying language models and document recognition to one task in your operations, such as extracting data from documents, searching internal knowledge or routing incoming work, with an accuracy target you can check.
It is for teams where skilled staff spend hours reading similar documents or answers sit in files nobody can search, and for organisations whose AI pilot stalled on questions of security, cost or accuracy.
Is this the right fit?
You may need this if
Staff retype document data
Invoices and forms arrive as PDFs and scans, and people key the same fields in every day.
Answers are buried in shared drives
Finding the right paragraph in policies, contracts and past cases depends on knowing who wrote it.
The AI pilot stalled
It worked on sample files, then questions about security, cost and accuracy on real documents stopped it.
An AI mandate with no use case
Leadership has asked for AI, but no workflow, owner or success measure is attached.
What we deliver
Document data extraction
Fields from invoices, forms and contracts, including Thai documents and scans, with uncertain values flagged.
Internal document search (RAG)
Answers with links to the passages used, drawn only from files the user is allowed to open.
Classification and routing
Incoming documents sorted by type and sent to the right queue, with unclear cases passed to a person.
Summaries linked to the source
Short versions of long documents, each linking back to the passages it draws on.
Review and evaluation
An accuracy baseline before launch, review screens for uncertain results, and logged corrections to track accuracy afterwards.
How we work
Every engagement runs through the same six-stage fitting process, from the first consultation to long-term care. The stages stay in the same order; how deep each one goes depends on the project.
See the full fitting process- ConsultationDiscover
- MeasureAnalyze and architect
- PatternDesign
- FittingBuild with you
- FinishingHarden and launch
- AftercareMaintain and support
How we scope and test an AI project
We scope each project to a task with a result we can measure. For example: extract six named fields from supplier invoices, and send any value the system is unsure of to a person. That gives an accuracy target, a path for errors and a cost to compare with doing the work by hand.
Before building, we test on a sample of your real documents, including poor scans and pages that mix Thai and English. The accuracy target and the review threshold are agreed with you from that test.
Language models can give wrong answers that read as confident. Drawing answers from your documents reduces errors, and showing the source passages makes the rest easier to spot. Outputs that feed a decision still need a person to check them. Search also has to respect permissions: the system retrieves only passages the user may open, before anything reaches the model.
Organizations we work with
Engagements span build, integration and long-term support. Named with permission; clients under NDA are not listed.
Related services
- Data & Analytics: when the need is reporting on structured data.
- System Integration & API Development: when the documents sit in systems that need connecting first.
- Enterprise Software Development: when the workflow around the AI needs building too.
- Solution Architecture: when you are still deciding whether AI suits the problem.
Frequently asked questions
How much does an AI document project cost?
Build cost depends on the number of document types, volume, language and scan quality, and the systems the output feeds. Running cost depends on the model, hosting and volume. We give a written estimate of both after a short scoping call, or price against the TOR for public-sector work.
How accurate will it be?
That depends on your documents, so we measure before promising a figure. A pilot on a sample of real files, including the messy ones, gives accuracy per field or question type. You then set the target and review threshold knowing roughly how much work will still go to people.
How well does this work with Thai documents?
Usually less well than with English. Thai handwriting, pages mixing Thai and English, and poor scans all reduce accuracy, and published benchmarks are often measured on English text. We test on your own documents before agreeing a target, and report Thai and English results separately.
Can we use commercial AI models with personal or sensitive data?
Sometimes, depending on the data's classification and the provider terms, including whether your data may be used for training. If not, sensitive fields can stay in your environment or go to a self-hosted model, with trade-offs in cost and capability. PDPA shapes the design; your advisers interpret the law.
Should we build this or buy a product?
Buy, if a product fits your workflow. Many document tasks are well served by existing tools, and we will say so. Custom work is worth its cost when the documents are unusual, when Thai handling matters, or when connecting a product to your systems would become custom work anyway.
Next step
Start with a conversation
Tell us about the problem you are trying to solve. A senior person reads every inquiry and replies with our view of fit, scope and cost, including when another option would suit you better.