From receipt photo to Omie: an evidence-based accounts payable workflow

Define the outcome that completes the process

Zild can receive a receipt photo, extract data with OCR, classify the expense and route the entry to Omie accounts payable. The expected outcome is a checked, traceable ERP record linked to its original document. A response saying the receipt was understood does not complete the work.

Start with one company, one expense group and an approval policy. Define who submits, checks, resolves classification questions and authorizes posting. This design also supports field teams and contractors: the person submits evidence through an authorized channel and tracks an identifiable exception without needing to know ERP codes.

Prepare credentials and master data

In Zild Admin, configure the Omie connector with the company application’s App Key and App Secret in ApiKey and ApiSecret. Check permissions and contact association when required by the agent. Saving a connector does not prove the operation will work: validate the intended access with test data.

Before classifying documents, make approved suppliers, categories, cost centers or departments and accounts available. Record the source and version of that reference. The agent should select valid codes; similar names do not authorize creating a supplier, inventing a category or changing the destination account.

Capture the original and wait for extraction

Preserve the received image, its identifier, authorized sender and intake time. Cropped photographs, glare, low resolution and multiple receipts in one image need handling. If a value is unreadable, request another image or review; do not fill missing digits from context.

In the Zild documents API, upload creates a record and processing continues in the background. Correlate the returned identifier with DocumentCompleted and check Completed status before consuming the result. Distinguish received, extracted, under review, approved, sent to ERP, confirmed and failed. These are proposed implementation states, not mandatory screen names.

Separate extraction, classification and approval

Extraction answers what is written: supplier, document identity, dates, amount and description. Classification relates that content to master data and company policy. Approval decides whether sufficient evidence exists to record the obligation. One stage does not replace another.

Synthetic example: a BRL 128.40 transport receipt may contain an issue date but no due date. The agent should not automatically copy it into the due-date field. Apply an approved rule or raise an exception. Similarly, a generic description such as services may map to different categories depending on contract, purpose and cost center. Present the rationale and source passage to the reviewer.

Post with confirmation and duplicate protection

Omie documentation identifies IncluirContaPagar as the creation operation and returns codigo_lancamento_omie for later reference. The codigo_lancamento_integracao field relates the integration record to the ERP entry. Check which actions are enabled for the agent in the connector and deployed version.

In the workflow design, retain a stable key per obligation, company and document. A retry should not create a new key. Also compare supplier, number, amount and dates; an image hash alone does not identify the same obligation photographed twice. After a timeout, query the record before retrying creation. Preserve the ERP response and treat rejection as an exception, not completion. Creating a payable does not execute payment or confirm settlement.

Build a sample representing operations

Separate development documents from a held-out evaluation sample. Include clear and poor photos, known and new suppliers, ambiguous dates, allocations, repeated documents, different numeric separators and evidence of already-paid expenses. The finance owner should define expected outcomes before evaluating the agent.

Compare manual and assisted workflows using the same criteria. Measure critical-field accuracy, classification accepted without correction, confirmed entries without rework, duplicates and review minutes. Record correctly rejected cases too. A workflow that posts every document may be worse than one that routes uncertainty to a person.

Calculate cost per correct entry

Add attributable subscription, OCR, model, integration, storage and human-review costs for the period. Divide by confirmed entries verified as correct. Include failed attempts and rejected documents in costs; excluding them makes automation look artificially cheap.

Freed hours represent recovered capacity. Record cash savings only when an expense actually decreases, and do not count the same hour as both cost reduction and additional capacity. Before expanding to more areas, the pilot should demonstrate time to confirmation, quality and cost per valid outcome.

Bring a concrete process to the assessment

Prepare anonymized receipt examples, allowed categories, due-date rules, approval owners and estimated monthly volume. Define the intake channel, available integration and acceptance criteria with Zild. Begin with review of every entry; increase autonomy only for validated document classes while retaining the ability to suspend ERP writes.