Rebuilding Finance Operations as an Agent Layer
How Stewdio Agents turned a fragmented, spreadsheet-heavy finance department into a continuously reconciled operating system.
Company Profile
The Challenge
Every acquisition had added another ledger, another chart of accounts, and another set of side spreadsheets. Controllers were closing the books on adrenaline, exception queues grew faster than they cleared, and leadership saw finished numbers weeks after the decisions that needed them. The CFO did not want more dashboards - she wanted an operating layer that did the reconciling work itself and escalated only what deserved human judgment.
The Audit (5 Weeks)
A Stewdio deployment engineer sat inside the finance org for five weeks, tracing 31 workflows across 7 sub-functions from source document to posted entry. One theme dominated: skilled people were operating as human middleware between systems that never learned to talk to each other.
Repeat exceptions - Roughly two-thirds of AP exceptions traced back to just 9 recurring failure modes - the same vendors, the same mismatch types, every cycle.
Silent intercompany drift - A 3.1% intercompany mismatch rate between SAP and the two Dynamics entities, typically caught quarters later during external review.
Late variance visibility - FP&A could not deliver a variance narrative until day 9-11 of the following month, after the operational moment had passed.
The morning cash scramble - A treasury analyst spent the first 90 minutes of every day stitching together a cash view that was outdated by 11:00 AM.
Aging-only collections - AR worked the oldest invoices first, ignoring clear payer-behavior signals sitting unused in three years of payment history.
Card spend re-coding - Brex categorization never mapped cleanly to SAP GL structures, forcing a manual re-coding sweep that consumed two full days of every close.
The Architecture
Stewdio did not ship isolated bots. We built one coordinated agent layer covering every sub-function, sharing a common data spine and a single audit trail.
Accounts Payable
An intake agent captures invoices from email, Coupa, and supplier portals, extracts line items, and three-way matches against SAP purchase orders. A triage agent resolves the 9 known exception patterns automatically and hands genuinely novel cases to a named owner with the full evidence chain attached. A disbursement agent sequences approved payments against terms and available cash.
Accounts Receivable
A collections agent scores every open invoice on payer behavior, dispute history, and account health, then drafts outreach matched to how each customer has actually responded before. A cash application agent auto-matches remittances across all 9 entities, including partials and short-pays.
Close Management
A reconciliation agent matches intercompany activity across SAP and Dynamics using probabilistic matching on counterparty, amount, date, and entity. A journal agent drafts recurring and accrual entries with built-in anomaly checks, while a close-status agent gives the controller a live task board across every entity.
Expense Management
A Brex-to-SAP mapping agent translates card categories into the correct GL structure, reconciles transactions against reports, and flags policy breaks before posting. The two-day re-coding sweep disappeared from the close calendar entirely.
Treasury
A cash agent aggregates balances from 11 bank accounts in 4 currencies and publishes the consolidated position before the workday starts, alongside a rolling 13-week forecast. A currency exposure agent watches cross-entity balances and surfaces hedge-worthy positions.
FP&A
A variance agent begins drafting the monthly analysis the moment entries post, attaching driver-level explanations to every material swing. A reporting agent compiles the leadership package straight from SAP, Brex, and Power BI models with no export-and-paste cycle.
Tax and Compliance
An estimates agent assembles entity-level results, applies jurisdiction rates, and produces the consolidated quarterly workbook. A compliance agent logs every agent action, human override, and data transformation so the department is audit-ready by default, not by fire drill.
Implementation (4 Months)
Each month shipped a working slice into production, with humans reviewing every agent decision until accuracy earned autonomy.
Invoice intake and exception triage live. Match accuracy opened at 85% and reached 96% by week four. Brex-to-SAP mapping removed the manual re-coding sweep from close.
Collections agent re-prioritized the full receivables book on payer behavior. Cash application auto-match climbed from 81% to 91%. Intercompany reconciliation moved from two dedicated staff to review-only.
First full close on the agent layer landed in 10 days, down from 15-19. The consolidated cash position began arriving before 6:30 AM, and payment sequencing started capturing early-pay discounts previously left on the table.
End-to-end integration pass across all agents, tax estimates validated against the prior two quarters, and escalation thresholds tuned. Second close on the full system: 6 days.
Results
Measured after the second full close cycle on the agent layer, against the trailing-year baseline.
This case study is an illustrative scenario: a representative engagement model, not a named client. We publish real deployment numbers as engagements complete.