ES
 work / ai-accounting-agents

Taking an accounting team from manual drudgery to AI agents it supervises.

Industry
Agricultural inputs, import and wholesale
Client
Not disclosed
Team
Accounting department
Duration
3 months, ongoing
Role
AI automation, end to end
01The problem

An accounting team drowning in manual work, and skeptical of AI.

The company imports seeds, fertilizers, and other agricultural inputs and sells them wholesale, only to authorized distributors. Its accounting department spent hours every day keying in supplier invoices and weeks on bank reconciliation. They had turned down automation and AI before. When global pressure reopened the door, they expected little: their own attempts had gone AI-first, throwing everything at a model with poorly framed asks.

Supplier XMLInvoice dataProduct linesChart of accounts01Extract02Classify03Validate04Human review→Posted entrySupplier XMLInvoice dataProduct linesChart of accounts01Extract02Classify03Validate04Human review→Posted entry
Supplier invoice intake: the XML each supplier sends goes in, a validated entry that a person approved comes out.
02Architecture

Before and after.

Before
Supplier invoicesKeyed in by hand, 2 to 4 hours a day
Bank reconciliationLine by line, two people, two weeks
Translation
Rules + ClaudeDeterministic where possible · AI where it pays
After
Invoice intakeAutomated, reviewed, approved by people
Reconciliation inventoryAnomalies · possible causes · suggested actions
6 matched2 flagged for review
Illustration of the monthly inventory: every bank line is matched against the books or flagged, with a possible cause and a suggested action for the team.
03Decisions

  1. Rules first, AI where judgment is needed

    Invoice data comes from the XML each supplier sends along with the invoice it files with the SRI, Ecuador's tax authority. Extraction, capture, and storage follow strict rules. The hard part is mapping each supplier's varied products to valid accounting categories and accounts.

  2. Humans approve, the system escalates

    Every entry is reviewed and validated. When something fails or lacks context, the system escalates instead of guessing. Observability, auditability, traceability, and human approval and supervision are built in from the start.

  3. Learn from the team's own history

    For bank reconciliation, deterministic logic and Claude work together, comparing the bank statement against the books every month. Claude learned from six months of past reconciliations how this team operates, which errors are common, and which patterns repeat.

  4. An inventory, not just a match

    The result is an exhaustive, detailed inventory: anomalies, failures, errors, and items not yet recorded, each with possible causes and suggested actions. The team decides; the system does the digging.

04Stack
ClaudeDeterministic rulesSRI e-invoice XMLHuman-in-the-loop approvalObservability and audit trail
05Results

Supplier invoice entry

Before2 to 4 hours a day
After15 to 30 min a day

Monthly bank reconciliation

Before2 weeks, 2 people
After4 hours

Bars drawn to scale using the upper bound of each range and 8-hour working days.

  1. Invoice entry: from hours to minutes

    Supplier invoice entry went from 2 to 4 hours a day to 15 to 30 minutes at most, with people still approving every entry.

  2. Reconciliation: from two weeks to four hours

    The monthly bank reconciliation used to take two people two weeks. It now takes 4 hours.

  3. From skeptics to the next project

    The team that once turned down AI is now extending the same approach to reconciling the company's 8 authorized credit cards.

Next step

Is your finance team buried in manual work? Let's talk.