Skip to content
Zoevin

Document and invoice intake

Get the fields off the PDF without retyping them

Invoices, statements, and forms arrive as documents and leave as somebody's afternoon. Extraction can read the recurring ones reliably enough to be useful and never reliably enough to be trusted blindly. The design that matters is the one that decides what a low-confidence read does — and the answer is always that it stops for a person.

The same office worker copies between two systems, tabs through a spreadsheet, then clicks down an inbox.
Trace the work
  • Confidence has consequences

    A field the extraction is unsure of routes to a review queue instead of being written as if it were certain.

  • Approval stays human

    Nothing is paid, released, or committed. The workflow prepares a record for a person to approve.

  • The source is kept

    The original document stays linked to the record it produced, so any figure can be traced back.

  • Client-owned handover

    You receive the workflow, the field mapping, the runbook, and the rollback path.

Chapter 01

Map the work

Start with the process, not the tool

  1. 01

    Sort documents by how alike they are

    A handful of senders in a stable format is a very different problem from arbitrary documents. The first is worth building for; the second usually is not.
  2. 02

    Decide which fields actually matter

    Most documents contain far more than the system of record needs. Extracting fewer fields well beats extracting everything badly.
  3. 03

    Find the approval that already exists

    Someone already checks these before they count. Identify who, what they look at, and what would make their check faster rather than replacing it.

Zoevin’s automation service is new and has not delivered a client engagement yet. My experience comes from building automation in salaried in-house roles.

Chapter 02

Build the right path

A bounded workflow with an owner

  1. 01

    Map document to record

    Define, per document type, which fields are extracted, what they map to, what makes a value valid, and what happens when validation fails.
  2. 02

    Build the extraction and the gate

    Implement extraction in the platform that fits your stack, with a confidence threshold, a review queue, and a written record of what was read versus what a person corrected.
  3. 03

    Prove the awkward documents

    Test a clean document, a re-sent duplicate, a changed layout, a missing field, and one that is simply the wrong document — before anyone relies on it.

Chapter 03

Prove and hand over

Acceptance is more than a happy path

  • A clean document posts once

    The record appears with the mapped fields and a link back to the document it came from.
  • An unclear read stops visibly

    Low-confidence fields reach a named reviewer rather than being written silently as a best guess.
  • Nothing moves money

    The workflow prepares records for approval and has no ability to pay, release, or commit on its own.

You receive the process map, decision memo, workflow export, access and secrets design, failure register, runbook, rollback, and handover record. See a fictional example of that artifact set.

Before we start

Know the boundaries

When this is the wrong buy

  • Every document is different

    Extraction earns its keep on repetition. If no two documents look alike, a person reading them is the honest answer.
  • Your accounting platform already does it

    Several platforms include document capture in a plan you may already pay for. If yours does and it meets the written need, configure that instead.
  • You want approval removed

    The gate is the design. If the goal is to have documents post without anyone looking, this is the wrong engagement.

Next step

Bring the real workflow to a fit consult

Bring the systems involved, the person who does the work, and an example of an exception. The first answer may be a setting you already own. If mapping is worthwhile, the next step is a fixed-scope discovery.