Explore/Academy/AI in Treasury Academy/Module 3

MODULE 03 OF 10 · ~31 MIN · NO CODE

AI Workflows That Replace Manual Work

From asking questions to running processes.

What you'll cover

01A 90-minute manual reconciliation turned into an 11-minute AI-assisted one
02A full reconciliation workflow in 4 steps, with 4 planted errors to find
03The Automation Ladder: 4 levels of AI maturity

You walk away with

INGEST > PROCESS > VALIDATE > OUTPUT
The Automation Ladder
The "My First AI Workflow" canvas and readiness scorecard

A prompt is a question. A workflow is a process.

Prompt (Module 2)

You ask AI one thing. AI gives you one answer.

"Analyze this statement and flag exceptions over EUR 10K."

You still

· Open the file manually
· Copy-paste the data
· Format the output yourself
· Repeat tomorrow

Workflow (Module 3)

AI runs a process. You validate the result.

Upload → AI parses → AI matches → AI flags → you review exceptions only

AI handles

· File parsing (structured extraction)
· Data matching (rules-based comparison)
· Exception flagging (threshold + pattern)
· Report formatting (ready to send)

The goal is not to use AI faster. It is to use AI instead of you for the repeatable parts.

Before: manual reconciliation, 6 steps, all human

15 min
Download statement
Log into bank portal. Export MT940/CSV.
210 min
Open in Excel
Import, clean headers, format columns.
35 min
Export ERP data
Run AP/AR report. Export to Excel.
445 min
Manual matching
VLOOKUP/eyeball. Row by row.
515 min
Flag exceptions
Highlight mismatches. Add notes column.
610 min
Write report
Summary email. Attach spreadsheet.

Total: ~90 minutes of human work. Every. Single. Day.

Boring. Error-prone. Unscalable. Dependent on one person knowing the process.

After: human reviews, AI works

Human · 11 min
Upload statement
Drag and drop. AI detects format.
AI · 210 sec
AI parses
Structured extraction. Any format.
AI · 315 sec
AI matches vs ERP
Amount + date + ref. Fuzzy matching.
AI · 45 sec
AI flags exceptions
Rules + thresholds. Categorized.
Human · 510 min
Human reviews
Only exceptions. Approve or reject.
AI · 65 sec
Report generated
Auto-formatted. Ready to send.

Total: ~11 minutes (was 90). The human touches the data twice: upload and review.

The human shifts from operator to validator. From doing the work to checking the work.

What makes a process automatable with AI?

Not everything should be automated. These are the five signals.

✓

Repetitive. Same steps, same logic, different data. Daily/weekly/monthly cycle.

✓

Rule-based. Clear rules exist: if amount > X, if date mismatch > 2 days, if field is blank.

✓

Data-heavy. Hundreds of rows. Human fatigue = errors. AI does not get tired at row 347.

✓

Time-sensitive. Morning position report, payment deadline, forecast submission.

✓

Low-risk validation. Human reviews exceptions, not every row. Errors are caught, not catastrophic.

If your process has all 5: automate it. If it has 3 or more: start with AI-assisted.

The 4-step workflow architecture

Every AI treasury workflow follows this pattern. Learn it once, apply it everywhere.

1INGEST

Upload data. AI detects format and extracts structured fields.

File upload + AI parsing prompt

2PROCESS

AI applies rules: matching, calculation, categorization, comparison.

Structured prompt with T.R.A.C.E.

3VALIDATE

Human reviews AI output. Approves, corrects, or sends back.

Exception review (not full review)

4OUTPUT

Final report, email, or data export. Ready to distribute.

Formatted table, email, PDF

Live build next: a complete bank reconciliation with this architecture. No special tools: an AI chat (Claude, ChatGPT or Copilot) and the data. Run it along with the course.

The data: one bank statement, one ERP export

A Société Générale statement for a French subsidiary (12 transactions: suppliers, customers, payroll, tax, intercompany, fees) and the SAP export for the same day. There are 4 realistic problems hidden in the two files. See if the AI finds them before you open the answer keys.

Data set 1 · MT940 (Société Générale, Entity France)

:20:STMT2025031001
:25:SOGEFRPP/FR7630003000700012345678901
:28C:00047/001
:60F:C250309EUR892341,67
:61:2503100310D247500,00NTRF20250310-001//SUP-FR-0891
:86:Fournisseur - Alstom SA - Facture #FAC-2025-1847 - Equipement industriel
:61:2503100310C185000,00NTRF20250310-002//CLI-FR-0445
:86:Client - Schneider Electric - Commande #PO-2024-9921
:61:2503100310D89400,00NTRF20250310-003//SAL-2025-03
:86:Paie Mars 2025 - 62 salaries - Entite France
:61:2503100310D34200,00NCHK20250310-004//TVA-2025-T1
:86:Tresor Public - Acompte TVA T1 2025
:61:2503100310C62750,00NTRF20250310-005//CLI-FR-0446
:86:Client - Legrand SA - Facture #INV-2025-0234
:61:2503100310D12800,00NTRF20250310-006//SUP-FR-0892
:86:Fournisseur - Bureau Veritas - Certification annuelle ISO 9001
:61:2503100310C450000,00NTRF20250310-007//IC-DE-0089
:86:Intercompany - Entity Germany - Cash pool sweep Q1
:61:2503100310D8750,00NTRF20250310-008//ASS-2025-Q1
:86:AXA Corporate - Prime assurance RC Pro trimestrielle
:61:2503100310D1890,00NTRF20250310-009//FEE-SG-03
:86:Societe Generale - Frais tenue de compte + commissions SWIFT
:61:2503100310C28900,00NTRF20250310-010//CLI-FR-0447
:86:Client - Valeo SA - Avoir sur retour marchandises #CN-2025-0018
:61:2503100310D175000,00NTRF20250310-011//SUP-FR-0893
:86:Fournisseur - Saint-Gobain - Materiaux construction projet Lyon
:61:2503100310D5600,00NTRF20250310-012//DIV-2025-03
:86:Divers - Abonnement logiciel SAP - Licence mensuelle Mars
:62F:C250310EUR844851,67

Data set 2 · ERP export (AP/AR ledger)

ERP Export - Entity France - 10 Mar 2025
Ref          | Type    | Counterparty        | Amount EUR  | Due Date  | Status
SUP-FR-0891  | AP      | Alstom SA           | 247,500.00  | 10-Mar-25 | Approved
CLI-FR-0445  | AR      | Schneider Electric  | 185,000.00  | 08-Mar-25 | Invoiced
SAL-2025-03  | Payroll | Payroll March 2025  | 89,400.00   | 10-Mar-25 | Processed
TVA-2025-T1  | Tax     | Tresor Public       | 34,200.00   | 10-Mar-25 | Filed
CLI-FR-0446  | AR      | Legrand SA          | 62,750.00   | 05-Mar-25 | Invoiced
SUP-FR-0892  | AP      | Bureau Veritas      | 12,600.00   | 10-Mar-25 | Approved
IC-DE-0089   | IC      | Entity Germany      | 450,000.00  | 10-Mar-25 | Confirmed
ASS-2025-Q1  | AP      | AXA Corporate       | 8,750.00    | 10-Mar-25 | Approved
CLI-FR-0447  | AR      | Valeo SA            | 28,900.00   | 12-Mar-25 | Credit Note
SUP-FR-0893  | AP      | Saint-Gobain        | 175,000.00  | 10-Mar-25 | Approved
DIV-2025-03  | AP      | SAP SE              | 5,600.00    | 01-Mar-25 | Approved
SUP-FR-0891  | AP      | Alstom SA           | 247,500.00  | 10-Mar-25 | Approved

INGEST: parse the bank statement

Raw data extraction. No analysis yet, just structure, plus one check: does the statement balance?

Step 1 prompt · paste the MT940 where indicated

You are a bank statement parser for a Senior Treasury Analyst performing month-end reconciliation.

TASK: Parse this MT940 bank statement into a structured table. Extract every transaction.

DATA: This is an MT940 from Societe Generale for Entity France (EUR), dated 10 March 2025.

[paste MT940 here]

CONSTRAINTS: Extract ALL transactions. Do not skip any. Do not summarize. Validate that opening balance + sum of transactions = closing balance. Flag if mismatch.

OUTPUT: Table with columns: # | Date | Reference | Description | Amount EUR | Direction (D/C) | Running Balance
After the table: Balance validation line.

Answer key · open after you run it

All 12 transactions parsed. Credits total 726,650; debits total 575,140; net +151,510.

Expected closing: 892,341.67 + 151,510 = 1,043,851.67. The statement says 844,851.67.

Problem #1: a closing balance mismatch of EUR 199,000. This is why INGEST includes validation: you catch data issues before you even start matching.

If the AI misses it, say so in the prompt ("Validate the balance and show your calculation") and rerun. That is the lesson: constraints must be explicit.

PROCESS: match against the ERP export

In the same conversation (so the AI keeps the parsed data from Step 1), paste the ERP export with this prompt. This is where the reconciliation logic lives.

Step 2 prompt · paste the ERP export where indicated

Now match the parsed bank transactions against this ERP export.

ERP DATA:
[paste ERP export here]

MATCHING RULES:
- Primary match: Reference number exact match + amount exact match
- Secondary match: Reference match + amount within EUR 500 tolerance (flag as PARTIAL MATCH with variance noted)
- Flag any bank transaction with no ERP match as UNMATCHED BANK
- Flag any ERP entry with no bank match as UNMATCHED ERP
- Flag duplicate ERP entries

CONSTRAINTS: Never force a match. If uncertain, mark as REQUIRES REVIEW. Show the match confidence: HIGH (exact ref + exact amount), MEDIUM (ref match + amount variance < 500), LOW (amount match only, no ref).

OUTPUT: 5 sections:
1. MATCHED (High confidence): table with Bank Ref | ERP Ref | Amount | Status
2. PARTIAL MATCHES (Medium): Bank Ref | ERP Ref | Bank Amount | ERP Amount | Variance | Issue
3. UNMATCHED BANK: transactions in bank with no ERP match
4. UNMATCHED ERP: entries in ERP with no bank match
5. DUPLICATES DETECTED: any repeated ERP entries
Summary line: X matched, X partial, X unmatched bank, X unmatched ERP, X duplicates.

Answer key · open after you run it

Matched (10): Alstom, Schneider, payroll, TVA, Legrand, Entity Germany, AXA, Valeo, Saint-Gobain, SAP licence.
Problem #2, partial match: Bureau Veritas, bank 12,800 vs ERP 12,600, a EUR 200 variance.
Problem #3, unmatched bank item: Société Générale fees, 1,890, with no ERP entry.
Problem #4, duplicate ERP entry: SUP-FR-0891 Alstom appears twice at 247,500.
Bonus: a sharp model may question Valeo. The ERP calls it a credit note, yet the bank shows an incoming credit. Worth a look.

VALIDATE: the human reviews exceptions only

The critical shift. You are NOT reviewing everything. You are reviewing what the AI flagged. At scale, it looks like this:

347

transactions parsed by AI

312

auto-matched (high confidence)

possible matches (review)

unmatched (investigate)

Your job: review 35 items, not 347. A 90% reduction in review scope. Spot-check 5–10 high-confidence matches, confirm or reject the possible matches, and find a valid explanation for each unmatched item (timing, pending, manual entry).

In our exercise: 4 items, about 5 minutes

Alstom duplicate (247,500). The critical one. The bank shows only one debit, so this is an ERP booking error, not a double payment. Correct it before month-end close.

Balance mismatch (199,000). Raise it with the bank or investigate: a statement error or a missing transaction.

Bureau Veritas (200). Check whether the bank added a commission or the invoice changed. If it is a bank commission, note it and approve the match.

Bank fees (1,890). Expected: fees are rarely pre-booked. Post a journal entry for account maintenance and SWIFT fees.

OUTPUT: generate the exception report

One last prompt in the same conversation turns everything into a report you can email to your manager, attach to the recon file, or archive for audit.

Step 4 prompt

Generate a reconciliation exception report from everything above.

FORMAT:
HEADER: Entity France | Societe Generale | 10 March 2025 | Prepared by: [Analyst Name]

SECTION 1 - SUMMARY
Total transactions: X | Auto-matched: X | Partial matches: X | Unmatched: X | Duplicates: X
Balance validation: PASS or FAIL (with variance amount)

SECTION 2 - EXCEPTIONS DETAIL
Table: # | Reference | Counterparty | Issue Type | Bank Amount | ERP Amount | Variance | Recommended Action

SECTION 3 - SIGN-OFF
Line for Treasury Manager signature, date, and comments.

CONSTRAINTS: Include ONLY exceptions. Do not list matched items. Sort by impact (highest variance first).

What good output looks like

Summary: 12 transactions, 10 matched, 1 partial, 1 unmatched, 1 duplicate, balance validation FAIL (199,000).

Exceptions sorted by impact: Alstom duplicate → balance mismatch → bank fees → Bureau Veritas. Sign-off line at the bottom.

Total time: ~12 minutes versus 90 manual, and the manual version would probably have missed the duplicate.

If something goes wrong: if the AI loses track in a long chat, start a new one and paste the Step 1 output. In production, you save each step's output and feed it fresh. If the answer is cut off, type "continue".

Same pattern, different workflows

INGEST > PROCESS > VALIDATE > OUTPUT works for any treasury process. The framework does not change; the data does.

Workflow	Ingest	Process	Validate	Output
Cash forecasting	Upload actuals + AR/AP aging	AI projects 13 weeks, flags shortfalls	Review outlier weeks	Forecast report + CFO email
Payment validation	Upload SEPA XML / payment file	AI checks duplicates, IBANs, amounts vs approved list	Review blocking errors	Validation report: pass/fail
FX exposure report	Upload deal register + forecast	AI nets exposures, calculates hedge ratios	Confirm netting logic	Exposure dashboard + policy flags
Bank fee review	Upload quarterly fee statements	AI categorizes, compares vs agreed schedule	Review overcharges	Negotiation brief for RM
Board pack prep	Upload month-end KPIs + data	AI generates narratives, calculates trends	Review tone and accuracy	Board-ready treasury section

The Automation Ladder

Four levels. Where are you today? Where do you want to be?

Level

Autonomous AI agentsHuman: Set strategy

AI runs processes end to end. Monitors, decides, escalates. Human sets policy, not steps.

2–3 years out for most teams

Level

AI executes with human approvalHuman: Approve

AI runs the full workflow. Human approves at checkpoints. Exceptions escalated.

Early adopters are here now

Level

AI processes documentsHuman: Review

AI handles parsing, matching, flagging. Human still initiates and reviews.

What you just built in this module

Level

AI answers questionsHuman: Ask

You ask, AI answers. One prompt at a time. No workflow.

Where most people are today

Level 1 to Level 2: where 90% of the value is

Achievable this month. No code, no special tools: just structured prompts chained into a sequence.

You open the file→AI opens the fileUpload + parse prompt

You eyeball the data→AI structures the dataT.R.A.C.E. extraction prompt

You do VLOOKUP→AI matches with rulesMatching prompt with criteria

You highlight exceptions→AI flags and categorizesThreshold + classification rules

You write the summary→AI drafts, you editOutput format specification

Level 2 to Level 3: the building blocks

Scheduled triggers

Process runs automatically at 8am.

Difficulty: Medium

API connections

AI pulls bank data directly. No manual upload.

Difficulty: Medium-High

Approval workflows

AI queues output; human approves via button or email.

Difficulty: Medium

Exception routing

Critical items go to the manager; routine items auto-approved.

Difficulty: Medium

Audit logging

Every AI decision logged: who ran it, when, what it decided.

Difficulty: Low

Treasury platform

All of the above integrated in a system built for treasury.

Difficulty: Depends

Exercise: map your first AI workflow

Pick ONE process you do regularly. Write down its name, frequency, the minutes it takes today and its pain points. Then map it into the four steps.

1. INGEST

What data do you start with? What format (MT940, CSV, Excel, PDF)? Where does it come from (bank portal, ERP, email)?

2. PROCESS

What logic do you apply? What are you comparing, calculating, matching or categorizing? What rules exist?

3. VALIDATE

What would you need to check in the AI output? What could go wrong? Which exceptions need human judgment?

4. OUTPUT

What does the final deliverable look like? Who receives it? In what format?

Automation readiness scorecard

Score your process on each criterion (1 = low, 5 = high).

Repetitive Same steps, same logic, different data each time.Rule-based Clear rules exist: thresholds, matching criteria, validation checks.Data-heavy Hundreds of rows. Human fatigue causes errors.Time-sensitive Deadline-driven: morning reports, payment cutoffs, period close.Low-risk validation Errors are catchable. A human reviews exceptions, not every row.

Total: 15 / 25

Prefer a full canvas? Try the AI Use Case Canvas.

✓ Module 3 complete

You now think in processes, not prompts

01
90 minutes became 12 on a real reconciliation, and the AI found a duplicate a tired human would likely miss.
02
INGEST > PROCESS > VALIDATE > OUTPUT. Works for any treasury workflow.
03
The Automation Ladder. The biggest jump is Level 1 to Level 2, and it needs no code.

AI does not replace treasury professionals. It replaces the parts of your job you wish someone else would do.

Next: Module 4, Governance, Security & Getting Buy-In. How to protect your organization and convince your CFO.

Built by a treasurer, for treasurers. · treasuryease.com