Custom AI document processing, for documents a tool cannot read
When an off-the-shelf IDP product stops at your fourteenth document format or your fifth language, the answer is a pipeline built for your documents: ingestion, classification, extraction, validation and a review step, written into the system you already run. The one we shipped for a logistics operation took 87% of the manual processing out at 99.2% extraction accuracy.
In production in 10 weeks · a platform of that shape sits in our $40,000–$120,000 AI product band · or hire AI engineers to build it inside your team from $4,000 a month
Should you buy an IDP product or build the pipeline?
Buy first, if a product fits. Intelligent document processing products and cloud document services handle common formats well, and a custom build only earns its cost where they stop.
A product is the right answer when
Your documents are common types — invoices, receipts, standard forms — in languages the product supports, the fields you need are the fields it extracts, and the result can live in its interface or a standard export.
A custom pipeline earns its cost when
Documents arrive from many systems in many shapes, in several languages, with fields specific to your business, and the values have to be validated against your own reference data and written into your ERP rather than exported.
And when the data cannot leave
If documents may not reach a third-party API at all, the pipeline is designed self-hosted. That is a constraint to state on the first call, because it changes the architecture and the price.
What a document processing pipeline is made of
Six parts, taken from the one we shipped for a multilingual logistics operation — client anonymised, every figure on the case study.
Ingestion and normalisation
PDFs, scans, structured exports and inbox attachments from 14 sources, normalised before anything is extracted. Invoices, customs paperwork and shipment documents all arrived in different formats.
Classification and extraction
Document type first, then the fields that type carries, across 5 operating languages without a separate workflow per region.
Validation against your own data
Extracted values are grounded against client-approved reference material and routing rules through a retrieval layer before an analyst ever sees them.
A review workspace
Confidence-based review with the source highlighted, exception handling and approval. Below the threshold a person decides; above it the value moves on.
Writing into the ERP
Validated values are synced into the ERP the operation already runs, with approval checkpoints kept for edge cases rather than a parallel system beside it.
A golden dataset from week one
An evaluation set across languages and document types, built before the pipeline, so accuracy is a measurement from the first slice rather than an impression at the end.
What it measured in production
And one thing it deliberately does not claim.
87% less manual handling and 99.2% extraction accuracy on production flows, measured by the evaluation harness rather than estimated. The corpus behind it is roughly 2,500 documents — about 15,000 pages — reindexed daily. The system answers in 1.9 seconds on average and 4.2 at p95, to the last token, at about $0.011 per answer.
What is not claimed: that the system is more accurate than the people it took work from. Nobody measured the manual process before it was replaced, so there is no before figure. 99.2% is what the system scores; the improvement that is measured is the workload. If you are about to automate document handling, measure the current process first — it costs an afternoon and cannot be bought back later.
The full case study · Why retrieval and not a fine-tuned model
How it rolls out without stopping operations
In slices, highest volume first, so the team keeps working while the system takes over.
Workflow audit
Map the highest-volume document flows and decide where automation saves the most analyst time first.
Golden dataset
Real documents with known correct values, across every language and type, before any pipeline code.
Pipeline in slices
One document flow at a time rather than a big-bang release across every business unit.
ERP integration
Validated values written into the existing workflow, with approval checkpoints for the exceptions.
Team enablement
A review workspace analysts trust, so adoption does not need heavy retraining.
Launch with monitoring
Accuracy dashboards and a feedback loop for prompt and retrieval improvements after go-live.
Common questions
What comes up first when the documents are on the table.
Have documents an IDP tool cannot read?
Tell us what arrives, in which formats and languages, and where the values need to go. We come back with how we would build it and what it costs, in writing.
Message received
We’ll review your enquiry and respond within one business day.