Digital Transformation · Field Notes
July 2026 · 5-minute read · Asseblia Insights
Every services company has a version of the same problem: documents arrive — invoices, contracts, onboarding forms, compliance filings — and skilled people spend hours reading them, retyping fragments into systems, and routing them to whoever acts next. Nobody budgets for this work. It simply absorbs whatever capacity exists.
Document intelligence — LLM-based classification, extraction, and validation of unstructured documents — is the most reliably deployable AI pattern in the back office today. It is also the pattern most often implemented badly, because teams optimize for the demo (one perfect invoice, parsed flawlessly) rather than for the operational reality (ten thousand imperfect documents, month after month).
The anatomy of a production pipeline
A document intelligence system that survives contact with real operations has four stages, each with a defined failure path:
Classification
Every inbound document is typed — invoice, contract amendment, purchase order, complaint — before anything else happens. Documents the classifier cannot place with high confidence go straight to a human queue. This single gate prevents most downstream errors.
Extraction
Structured fields — amounts, dates, counterparties, line items — are pulled from the document into a schema defined per document type. Extraction runs with confidence scores per field, not per document; one uncertain field does not discard an otherwise clean result.
Validation
Extracted data is checked against business rules and existing records: does the supplier exist, does the PO match, is the amount within tolerance? Validation is deterministic code, not an LLM — rules should never hallucinate.
Routing and action
Clean documents post directly into the target system — ERP, CRM, DMS — with a full audit trail. Everything else lands in a review queue with the extracted data pre-filled, so the human corrects rather than retypes.
Where the human belongs
The 70–80% touchless rate is not the interesting number. The interesting number is what happens to the remaining 20–30%: in a well-designed system, exception handling becomes the entire human job — and a better job than the one it replaces. Instead of retyping every document, the team adjudicates the genuinely ambiguous ones, with the machine’s best guess already on screen.
The threshold between automatic and human is a business decision, not a technical one. Sensible defaults we see work in practice:
→ Value thresholds — documents above a monetary limit always get human sign-off, regardless of confidence
→ Confidence thresholds — any field below the cut-off sends the document to review, tuned per field criticality
→ Novelty rules — first document from a new counterparty or a new document layout is always reviewed
→ Sampling — a fixed percentage of “clean” documents is spot-checked to detect silent drift
The failure mode to design against
Document pipelines rarely fail loudly. They fail quietly: a supplier redesigns their invoice template, extraction accuracy on that supplier drops from 98% to 60%, and nobody notices for a quarter because the documents still flow — just with wrong data in them. The countermeasure is unglamorous: per-source accuracy monitoring, the sampling rule above, and a monthly reconciliation between what the pipeline posted and what the ledger shows.
Deployment effort follows the same pattern we describe across every vertical: the model is the easy part. Connecting it to your ERP’s quirks, your approval matrix, and your auditors’ expectations is the project.
“Automation that removes the human entirely is a liability. Automation that moves the human to the exceptions — that’s an asset.”
Next step
Drowning in documents your team retypes by hand?
Book a 45-minute discovery call — we’ll map one high-impact automation opportunity. No commitment required.