← Back to portfolio
O2C Reconciliation Modernization
L5 — Business Case & Enterprise Readiness
What Would It Take to Use This in a Real Organization?
L4 established where AI belongs in the workflow and where it deliberately doesn't. L5 answers what a hiring manager or budget approver actually asks next: what does this cost and return, who's accountable for each decision, what could go wrong, and what's still missing between this prototype and a production system.
Source: L2/L4 verified figures + labeled assumptions No production system — case study on public data Built: Jul 2026
This is not intended to demonstrate that AI should replace financial analysts. It demonstrates how a Business Analyst redesigns a financial review workflow so AI handles repetitive preparation while humans retain accountability for every business decision — and what it would take to run that redesign for real.
Contents
01
Executive business case
Actual figures from L2/L4, kept separate from labeled assumptions
There's no production system here to benchmark against, so this section is built two ways on purpose: numbers that are Actual (verified against L2's SQL output) and numbers that are Assumption (a labeled, sourced estimate standing in for a measurement that doesn't exist yet). Presenting both honestly — instead of one blended "ROI" figure — is the point of this section, not a hedge.
Recoverable cash
R$199.52
Actual — UNDERPAYMENT only (83 orders), the one category that's genuinely money owed back
Risk-managed exposure
R$268,659.73
Actual — CANCELED_WITH_PAYMENT + MISSING_ITEMS + OVERPAYMENT (refund liability & data-integrity risk, not new cash in)
Total exception exposure
R$271,891.01
Actual — verified vs. L2/L3/L4, includes expected installment interest
Exceptions in scope
1,618
Actual — verified vs. L2/L3/L4
The honest headline number
Recoverable cash is R$199.52 — not R$271,891.01. Most of the exposure is a refund liability (CANCELED_WITH_PAYMENT) or a data-integrity gap (MISSING_ITEMS), not cash walking in the door. The real business case for this workflow is faster, more consistent resolution of that exposure — not cash recovery.
L4 Step 02 already states its manual-review baseline as 10–15 min/case, "industry-typical estimate — not measured on this dataset." This section reuses that exact range (midpoint 12.5 min) rather than introducing a new number, and assumes 4 min/case for the AI-assisted review (reading a drafted package and validating it, vs. building one from scratch).
inputvaluesource
Manual review time12.5 min/caseAssumption — midpoint of L4's stated 10–15 min range
AI-assisted review time4 min/caseAssumption — validate a drafted package vs. build from scratch
US loaded hourly rate$58.50/hrAssumption — entry BA/Data Analyst, SF Bay Area (ZipRecruiter, Salary.com) ×1.3 benefits/overhead
Brazil loaded hourly rateR$67.60/hrAssumption — entry Financial Analyst, Brazil (ERI SalaryExpert) ×1.3 benefits/overhead
229.2 hours saved clearing the full 1,618-case backlog once
Assumption — 8.5 min saved × 1,618 cases
$13,409 / R$15,495 labor value of that one-time saving
Assumption — hours saved × loaded rate, both currencies
This project's real build hours weren't tracked cleanly enough to report as a verified figure — commit and session timestamps could confirm some sessions but not others, and inventing precision there would fail the same no-fabrication bar as inventing a dollar figure. Instead, build cost below models a different, clearly-labeled question: if an organization hired an analyst to build this from scratch, how long would it reasonably take? One input has a direct external citation; the rest are scope-based estimates.
layerhoursbasis
L0 — Discovery14Scope estimate — requirements/data-profiling norm (~1–2 days) extended for SQL exploration
L1 — Process mapping6Scope estimate — two BPMN diagrams + documentation
L2 — SQL classification20Scope estimate — 6 exception types, validated against two known bug classes
L3 — Dashboard12Cited — 8–15 hrs for a moderately complex dashboard on clean data (Elevated Signal), midpoint
L4 — AI workflow design20Scope estimate — two-axis priority framework + interactive demo build
Total72~9 working days
$4,212 / R$4,867 scenario build cost (72 hrs × loaded rate)
Assumption — both currencies, see rate basis above
0.31 cycles to break even on labor savings alone
Assumption — build cost ÷ one-time backlog labor savings, same in both currencies
Why 0.31 isn't the number to lead with in an interview
The 1,618 exceptions are a multi-year historical backlog (2016–2018 order data), not a monthly volume. "0.31 cycles to break even" describes clearing that backlog once — it says nothing about ongoing monthly ROI, which would need an actual monthly exception volume to compute honestly. Presented without this caveat, the number overstates the case; presented with it, it's a defensible one-time efficiency gain plus an open question for further analysis.
02
AI governance & controls
Who decides what, how the design prevents the AI from acting on its own, and what's still risky
These three tables answer three different questions on purpose: who is accountable for each decision, how the workflow physically prevents AI from acting unsupervised, and what could still go wrong even with those controls in place. All three describe the actual L4 design — none of this is aspirational.
decisionaihuman
Detect & classify exception (SQL rules)✓
Assign risk tier / priority✓ (SQL, not AI)
Draft summary, cause, action✓ Step 4
Approve / modify / reject the draft✓ Step 5
Escalate a case✓ Step 5
Write to the ledger / system of record✓ (manually, outside this tool)
Log what happened✓ Step 6, independent of AI's draft
Reading the table above literally: AI participates in exactly 1 of 6 workflow steps, and that step produces a draft, not a decision. Every other row is either a fixed SQL rule or a human action.
1
Row count verified vs. L2/L3 (1,618)
→
2
SQL business rules only, no ML training
→
3 · AI
Fixed prompt, grounded strictly in SQL fields
→
4
Human validation (Approve/Modify/Reject/Escalate)
→
5
Decision logged to audit trail
→
6
No automatic write-back to ledger
The last link is a design constraint, not a policy note: this workflow has no code path that writes to a system of record. L4's demo literally cannot do it — the validation buttons log a decision string and stop there.
riskmitigation
AI hallucinates a cause or amount not in the dataHuman validation required on every case (Step 5); prompt grounded only in SQL-computed fields, no external facts
Wrong dataset / schema silently usedRow-count verification against L2/L3 (1,618) before any AI step runs
Missing or null fields break classificationCASE WHEN business-status flags evaluated before amount-based flags; COALESCE handles missing joins (see L2)
Prompt injection via order/customer text fieldsFixed prompt template; inputs are structured numeric/categorical SQL fields, not free-text customer input
Reviewer over-trusts the AI draft ("automation bias")Four explicit decision options (not just Approve) — Modify and Reject are first-class actions, not friction
Sensitive/PII data exposureDataset is public/synthetic (Olist); no real customer PII was ever in scope
03
Business analyst readiness questions
What I'd need answered before recommending this be built for a real organization — not built yet, deliberately
This section is explicitly hypothetical: a checklist of what a BA should ask before scoping this for a real company, not a claim that these questions were asked or answered here. The point is the instinct to ask business questions before technical ones — not the specific answers, which don't exist for a synthetic case study.
Questions I'd ask before scoping this for real
Which ERP or order-management system is the source of truth today, and does it expose an API?
Who currently has approval authority over a refund or write-off, and at what dollar threshold does that authority change?
Is there an existing SLA for exception resolution, and what does "resolved" mean today?
Which exception types legally or contractually require escalation beyond finance (e.g. to legal, or to the customer directly)?
Where do audit logs live today, and who has access to them?
Which KPIs does finance leadership actually review weekly — exposure dollars, case age, or resolution rate?
What regulatory regime applies (SOX, LGPD given this is Brazilian data, PCI if payment data is involved)?
What's the organization's acceptable AI risk tolerance — is a drafting-only AI acceptable, or is even that too much for this workflow?
04
Enterprise deployment roadmap
The honest gap between this prototype and a production system
Today — this portfolio
Static CSV dataset (Olist, public/synthetic)
DuckDB, run locally, no persistence layer
20 AI packages generated once, embedded as static JSON
No authentication — page is public
Validation decisions logged to the browser only, not a database
Production — what it would take
Live connection to the ERP / order-management system
Scheduled or event-driven pipeline (not a manual DuckDB run)
Live API calls per case, with rate limiting and cost monitoring
Authentication + role-based access (who can Approve vs. who can only view)
Decisions written to a real audit-trail database, queryable for compliance
1
CSV / static demo (today)
→
2
Persistent database (Postgres/Snowflake)
→
3
ERP / API integration
→
4
Authentication + role-based access
→
5
Audit database + monitoring/alerting
Scope honesty
This portfolio demonstrates the concept and the design reasoning behind it — not a production-ready system. Every item in the "Production" column above is real, scoped engineering work that hasn't been built.
05
Executive conclusion
What L0–L5 together are meant to prove
layerwhat it demonstrates
L0Understanding the business problem before touching a tool
L1Analyzing and designing a process, not just describing it
L2Turning business rules into verified, defensible SQL
L3Communicating findings through a dashboard a non-technical stakeholder can read
L4Deciding where AI belongs in a process, and designing human-in-the-loop control around it
L5Thinking like an enterprise: governance, controls, risk, and what's still missing before this is real
This portfolio is not intended to demonstrate that AI should replace financial analysts. It demonstrates how a Business Analyst can redesign a financial review workflow so AI handles repetitive preparation while humans retain accountability for business decisions.
★
Lesson: A business case that's honest about which numbers are measured and which are assumptions is more defensible in an interview than a bigger number that isn't. The R$199.52 recoverable-cash figure looks unimpressive next to R$271,891.01 — and stating that difference plainly is what makes the rest of this section trustworthy.