Back to Work
Project / Finance Operations

AI-Augmented Finance Operations & Control System

A deterministic-first AI implementation for billing, reconciliation, forecasting, and source-verified financial reporting.

Built a 6-day no-code/low-code finance operations system using n8n, Airtable, Gemini, Nanonets, and Google Drive.

The system covers invoice processing, operational matching, AR collections, expense categorization, bank reconciliation, 13-week cash-flow forecasting, liquidity monitoring, and live-source verification.

Core principle

Deterministic logic owns financial truth. AI assists with extraction and grounded communication inside validated boundaries.

$10.4M

Balanced ledger

640

Ledger transactions

400

Invoices evaluated

578

Broken account relationships reconstructed

$6.815M

Outstanding AR cross-verified

11 / 11

Headline figures re-verified from live data

Overview

How can AI improve finance operations without becoming the authority for financial decisions?

Finance workflows are especially sensitive to silent errors. A system can generate convincing reports while still using the wrong amounts, incorrect classifications, broken relationships, or unsupported assumptions.

Instead of building an “AI does everything” workflow, this project separates the architecture into two responsibilities.

Deterministic systems
  • calculations
  • filtering
  • matching
  • classifications
  • reconciliation
  • thresholds
  • routing
  • Match / Mismatch decisions
  • source verification
AI-assisted tasks
  • document extraction
  • dunning-message wording
  • forecast narrative generation
  • source-verification reporting

Generative output cannot be persisted unless it passes deterministic validation.

Problem

Finance automation fails when the outputs look right but the underlying controls are wrong.

A staffing and professional-services finance workflow can fail in multiple places:

  • invoice data can be incomplete or inconsistent
  • transaction relationships can break during imports
  • expenses can be posted to incorrect accounts
  • duplicate transactions can enter the ledger
  • bank timing differences can look like unexplained variances
  • bank-only fees may never be recorded in the books
  • overdue invoices can be routed incorrectly
  • cash-flow models can silently depend on unsupported assumptions
  • LLM-generated reports can introduce numbers or claims that are not grounded in source data

The goal was not simply to automate finance tasks.

The goal was to build a system where automation remains traceable, testable, and auditable.

Architecture

System Architecture

Airtable acts as the operational data layer and n8n acts as the orchestration and deterministic control layer.

Financial decisions remain outside the LLM. Gemini and other AI services operate only in narrow extraction or language-generation tasks.

End-to-end architecture showing financial data flowing through independent deterministic controls, converging into the Key Figures Register, and finally producing a validated source-verification report.

Control Boundary

Deterministic vs. AI Decision Boundary

The most important design decision in this project was deciding where AI was allowed to act.

AI was useful for tasks involving language or unstructured information, but it was deliberately prevented from owning financial truth.

  1. 01

    Raw / Live Data

  2. 02

    Deterministic Controls

  3. 03

    Deterministic Fact Package

  4. 04

    LLM-Assisted Generation

  5. 05

    Deterministic Validation

  6. 06

    PASS / BLOCK

The LLM can assist with extraction and grounded communication, but calculations, Match/Mismatch decisions, reconciliation, routing, and publication controls remain deterministic.

Systems

Six finance workflows working under one control architecture

01

Invoice–Job–Time Entry Match

Operational three-way validation across invoices, jobs, and billable time.

400
Invoices evaluated
394
Exception Review
6
Pass
0
Workflow errors

The high exception rate is preserved as an unresolved data-quality issue rather than presented as 394 confirmed economic failures.

02

AR Dunning & Collections

Deterministic overdue classification with AI-assisted message generation.

240
Paid
20
Draft
140
Final Notice
$6,815,160
Outstanding AR
140 / 140
Required messages generated

Gemini generated collection-message wording only after deterministic eligibility and stage classification.

03

Expense Categorization

Rule-based expense classification with seeded-error testing.

62
Expense transactions processed
59
Auto Confirmed
3
Likely Misclassification
3 / 3
Seeded misclassifications surfaced

The benchmark is not fully independent because transaction descriptions contributed to both Account reconstruction and later categorization.

04

Bank Reconciliation

Amount, direction, normalized-description, and date-based matching.

172
Matched
2
Timing Differences
2
Possible Duplicates
2
Bank-Only Fees
0
Book Only
6 / 6
Seeded reconciliation artifacts surfaced
0
False-positive seeded exception classifications
05

13-Week Cash-Flow Forecast

Three-scenario deterministic liquidity forecasting.

$4,151,185.83
Opening cash
$6,126,693.59
Week 13 ending cash · Optimistic
$5,104,419.59
Week 13 ending cash · Base
$4,422,903.59
Week 13 ending cash · Pessimistic
60.37×–62.27×
Minimum cash coverage · known modeled fixed obligations

The forecast deliberately excludes new invoices, new revenue, operating growth, and staffing changes.

06

Live Source Verification

Independent recomputation of reported headline figures from live source data.

11 / 11
Real headline figures matched
Negative control
999
Reported Draft Invoice Count
20
Live Result
MISMATCH
Result

This demonstrates that the verification layer can reject an incorrect reported value rather than simply confirm what it is given.

Data Integrity

A Major Data Integrity Issue I Found Mid-Build

During an integrity audit, the system uncovered a structural issue that was not part of the planned demonstration.

Only 62 of 640 Transactions had populated Account links.

That meant: 578 missing Account relationships

Rather than manually editing the ledger, I built a deterministic reconstruction workflow that mapped transaction patterns back to the correct Chart of Accounts records.

578 / 578

Account links reconstructed

640 / 640

Transactions linked

0

Blank accounts

0

Multi-account errors

320

Debit rows

320

Credit rows

Debits

$10,402,848.17

Credits

$10,402,848.17

Automation was used not only to process data, but to investigate and repair a structural data-integrity failure.

Grounded Reporting

Grounded AI Reporting

Once deterministic finance calculations were complete, Gemini was allowed to generate human-readable financial narratives.

  1. 01

    Verified Data

  2. 02

    Deterministic Fact Package

  3. 03

    Gemini

  4. 04

    Deterministic Validation

  5. 05

    PASS / FAIL Gate

  6. 06

    Artifact

Gemini was not responsible for calculating forecast values or deciding whether a financial figure matched.

It could only communicate precomputed facts.

  • numeric grounding
  • required sections
  • citation fidelity
  • prohibited overclaiming
  • required limitations and caveats

AI Guardrails

The Guardrails Actually Blocked Bad AI Outputs

The report-generation guardrail was not theoretical. AI-generated outputs had to pass deterministic validation before any artifact could be persisted.

AI-generated reporting passes through deterministic checks for completeness, numeric grounding, citation fidelity, required caveats, and prohibited overclaiming before publication.

Execution #107Blocked

Gemini produced an incomplete report.

  • KF-11
  • Cross-Day Consistency section
  • KF-12 section
  • Source Tables Verified
  • Interpretation & Limitations

Validation failed.

No file was created.

No Drive upload occurred.

Execution #108Blocked

Gemini generated the complete report but paraphrased a required KF-03 limitation.

Validation failed again.

Instead of weakening the validator, deterministic caveat enforcement was added.

Execution #109Published

The final run passed:

  • required sections
  • numeric grounding
  • citation fidelity
  • overclaiming scan
  • KF-03 caveat validation

Only then was Day6_Source_Verification_Report.md created and uploaded to Google Drive.

Implementation Evidence

Evidence from the running system

578 / 578

Missing Account relationships reconstructed

6 / 6

Seeded bank-reconciliation artifacts surfaced

3 / 3

Seeded expense misclassifications surfaced

11 / 11

Real headline financial figures independently re-derived

999 → 20

Deliberately incorrect control figure rejected

2 BLOCKED → 1 PUBLISHED

AI report validation history before persistence

The implementation evidence comes from deterministic outputs, negative controls, seeded exceptions, live-source recomputation, and validation-gate behavior rather than presentation screenshots.

Key Results

Measured outcomes across the finance control system

$10.4M

Balanced debit / credit ledger

640

Ledger transactions controlled

400

Invoices evaluated

3,000

Time Entries incorporated into operational validation

578

Missing Account relationships reconstructed

$6.815M

Outstanding Final Notice AR independently cross-verified

6 / 6

Seeded bank-reconciliation artifacts surfaced

3 / 3

Seeded expense misclassifications surfaced

with documented benchmark limitation

13 Weeks

Three-scenario cash-flow forecast

11 / 11

Headline financial figures independently re-derived

Implementation Problems

Problems I Had to Solve

Broken Time Entry Relationships

Problem

All 3,000 Time Entries initially had missing Job links after API import.

Resolution

Created relationship-repair workflows using raw IDs and deterministic record mapping.

Gemini Parser Failure

Problem

The initial integration expected the wrong Gemini response structure.

Resolution

Inspected the live response and corrected extraction to:

content.parts[0].text

n8n Execution Limits

Problem

Large sequential writes exceeded execution time.

Resolution

Used controlled batches and idempotent reruns.

Airtable Output-Shape Differences

Problem

Some reads returned flattened fields while others returned item.json.fields.

Resolution

Implemented shape-safe field access.

578 Missing Account Links

Problem

A ledger integrity audit uncovered a large structural relationship failure.

Resolution

Created a deterministic Account Reconstruction workflow.

Floating-Point Validation

Problem

Correct two-decimal forecast values initially failed strict comparisons.

Resolution

Replaced naive equality with cent-safe / epsilon-safe validation.

Incorrect Live Dunning Value

Problem

The assumed source value was “Final.” Live data actually contained “Final Notice.”

Resolution

Performed read-only source inspection and changed verification to exact-match the live value.

Incomplete AI Report

Problem

Gemini stopped before completing all required report sections.

Resolution

Increased the output allowance while keeping deterministic validation unchanged.

AI Paraphrased a Required Limitation

Problem

Gemini rewrote an important disclosure instead of preserving it exactly.

Resolution

Added deterministic caveat enforcement rather than weakening validation.

Limitations

What this system does not prove

Three-Way Match

The 394 / 400 Exception Review rate is reproducible, but its underlying cause remains unresolved. The result demonstrates deterministic consistency with current source data, not that all 394 rows are genuine economic exceptions.

Expense Categorization

The expense-categorization benchmark is not fully independent because transaction descriptions contributed to both Account reconstruction and later classification.

Cash-Flow Forecast

The forecast excludes:

  • new invoices
  • new revenue
  • operating growth
  • staffing changes

It tests liquidity against the stated assumptions rather than predicting complete business performance.

Synthetic Environment

Several bank and classification exceptions were deliberately inserted to test whether the controls could detect them. This project demonstrates architecture, implementation discipline, and control behavior — not audited production-company results.

Skills

Skills Demonstrated

AI Implementation Architecture

Designed a system where generative AI operates inside explicit deterministic boundaries.

No-Code / Low-Code Orchestration

Built multi-stage workflows across n8n, Airtable, Gemini, Nanonets, and Google Drive.

Financial Operations

Worked with:

  • Chart of Accounts
  • Double-entry accounting
  • AR aging
  • Dunning
  • Expense categorization
  • Reconciliation
  • Cash forecasting
  • Liquidity monitoring
  • Financial controls

Data Engineering & Integrity

Implemented:

  • Linked-record repair
  • Deterministic mappings
  • Batching
  • Idempotent reruns
  • Audit logging
  • Source tracing

AI Guardrails

Implemented:

  • Structured fact packages
  • Numeric grounding
  • Citation fidelity
  • Post-LLM validation
  • Deterministic fallback
  • PASS-only persistence

QA & Validation

Used:

  • Sample validation
  • Negative controls
  • Planted exceptions
  • Independent recomputation
  • Live-source verification

Stack

Technology Stack

n8n

Workflow orchestration and deterministic control logic

Airtable

Operational finance database and control logs

Gemini 2.5 Flash

Document extraction and grounded language generation

Nanonets

Invoice extraction benchmark

Google Drive

Validated artifact storage

JavaScript / n8n Code Nodes

Lightweight deterministic transformations, aggregation, and validation

Artifacts

Project Artifacts

Day5_13_Week_Cash_Flow_Forecast_Summary.md

Day6_Source_Verification_Report.md

The implementation also includes production workflows for invoice extraction, dunning classification and messaging, expense categorization, transaction integrity auditing, Account reconstruction, bank reconciliation, forecasting, liquidity controls, figure verification, and final source verification.

Takeaway

What This Project Demonstrates

This project started as a finance automation exercise and became a broader demonstration of controlled AI implementation.

The most important result was not that AI could write a collections message or summarize a forecast.

The important result was demonstrating a system where AI could accelerate financial workflows without becoming the authority for financial truth.

  • headline metrics can be traced back to live source data
  • structural data-quality failures can be identified and repaired
  • deliberately planted reconciliation exceptions can be surfaced
  • intentionally incorrect reported values can be rejected
  • incomplete or unsupported AI-generated reports can be prevented from being persisted

Automation, intelligence, and control were treated as separate responsibilities.

That separation is the architecture I would carry into production AI implementation work.

Want to see another implementation?

View Conversational Analytics AgentLinkedInResumeContact Me
Back to Work