AI in Treasury: The Use Case Map for Cash, Capital, and FX

Treasury sits on structured, high-frequency, numerical data with clear feedback loops. That makes it one of the more tractable finance functions for AI. But the work is also unforgiving: a bad cash forecast means an unfunded payroll, a missed FX hedge means a P&L hit at quarter-end, and a debt covenant breach can trigger downstream credit consequences. This page maps where AI is being applied across the treasury stack, what data each use case requires, and what to evaluate before buying.

The map is organized around five treasury domains: cash management, in-house banking and intercompany, foreign exchange, debt and capital markets, and short-term investments. Within each, AI shows up in three forms — predictive models for forecasting, Generative AI for document and communication workloads, and Agentic AI for multi-step workflows that previously required a treasury analyst to stitch together systems.

Why treasury is moving on AI now

Three pressures are converging. First, interest rates have made idle cash expensive again, which raises the value of accurate forecasting and active liquidity management. Second, treasury teams are thinly staffed relative to the data volume they now process — TMS platforms, ERP modules, bank portals, and trading venues each produce their own feeds. Third, the data is finally in usable shape: API-based bank connectivity, ISO 20022 migration, and consolidated TMS reporting have removed much of the manual reconciliation that used to dominate the day.

What this means in practice: the bottleneck has shifted from data acquisition to data interpretation. That is where AI earns its keep.

Domain 1: Cash management and forecasting

Cash forecasting is the most mature treasury AI use case. The work is well-defined — predict inflows and outflows across entities, currencies, and time horizons — and the data is structured. Machine learning models replace or augment the spreadsheet-driven categorization that treasury analysts have historically maintained.

1. Short-horizon cash forecasting (1–30 days). Predicts daily inflows and outflows by bank account and currency. Models learn from historical bank transactions, AR aging, AP run schedules, and payroll cycles. Outcome: tighter target balances and less precautionary cash.
2. Medium-horizon forecasting (1–12 months). Blends transactional history with budgets, sales pipelines, and seasonality. Used to plan revolver draws, term debt issuance, and capex funding.
3. Transaction categorization. Classifies incoming bank transactions into customer receipts, intercompany flows, tax, payroll, and other categories. Reduces the manual coding burden on cash accountants.
4. Variance explanation. Generative AI summarizes why the actual cash position differs from the forecast — which customer paid early, which subsidiary drew on a credit line — by reading transaction data and posting a daily commentary.
5. Liquidity stress testing. Simulates downside scenarios (delayed receipts, supplier prepayment demands, market closures) to size committed liquidity buffers.

Domain 2: In-house banking and intercompany

Multinationals run internal banks that net intercompany positions, pool cash across entities, and settle in fewer external currencies. The work is repetitive and rule-heavy, which makes it well-suited to agentic AI.

1. Intercompany netting orchestration. Agents reconcile intercompany invoices across ERPs, propose netting settlements, and route exceptions to humans.
2. Cash pooling optimization. Models suggest target balances and sweep amounts across pool participants based on local funding needs, withholding tax considerations, and minimum operating balances.
3. Transfer pricing documentation. Generative AI drafts intercompany loan agreements and supports the documentation trail for tax authorities — with human review before signing.
4. FBAR and entity-level reporting. Automates the assembly of regulatory filings that require pulling balances across dozens of legal entities and jurisdictions.

Domain 3: Foreign exchange

FX is where treasury AI gets the most attention and the most skepticism. Two things are happening: AI is improving exposure identification and hedge execution workflows, and separately, vendors are pitching alpha-generating directional models. The first is real and increasingly mature. The second deserves caution.

1. Exposure aggregation. Pulls FX exposure from ERP order books, intercompany positions, and forecasted commercial flows into a single net exposure view by currency pair and tenor.
2. Hedge ratio recommendation. Suggests hedge ratios based on policy, exposure confidence, and forward point cost. Used to support, not replace, the treasurer's decision.
3. Execution timing analytics. Analyzes intraday liquidity and spread patterns to suggest better execution windows for large notional trades.
4. Hedge effectiveness testing. Automates the retrospective and prospective effectiveness tests required for hedge accounting under IFRS 9 and ASC 815.
5. Confirmations and settlement matching. Generative AI parses incoming trade confirmations and matches them to internal trade tickets, flagging discrepancies.

Domain 4: Debt and capital markets

Debt workflows are document-heavy and infrequent, which makes them a strong fit for Generative AI. The bulk of value here is in reducing the analyst-hours spent assembling, reviewing, and monitoring credit documentation.

1. Covenant monitoring. Extracts covenant definitions from credit agreements and tracks compliance against actuals each reporting period.
2. Loan documentation review. Compares term sheets, draft credit agreements, and final executed versions to flag changes. Speeds the legal review cycle.
3. Rating agency Q&A preparation. Drafts responses to rating agency questionnaires by pulling from prior filings, financial statements, and management presentations.
4. Debt portfolio analytics. Models the impact of refinancing decisions, prepayment options, and rate changes on weighted-average cost of debt and maturity profiles.
5. Investor communications. Generates first drafts of bondholder updates and roadshow materials, with treasury and IR retaining editorial control.

Domain 5: Short-term investments

Investment policies typically constrain corporate cash to highly rated, short-duration instruments. AI's role here is narrower — there is less judgment to automate when the policy already restricts the universe — but it shows up in portfolio monitoring and reporting.

1. Counterparty exposure monitoring. Tracks deposits, MMF holdings, and repo positions against counterparty limits. Alerts when a bank rating action or news event affects exposure.
2. Yield optimization within policy. Suggests allocations across MMFs, term deposits, and government securities that improve yield while respecting duration, credit, and concentration limits.
3. Investment policy compliance. Automated checks that flag any proposed trade that would breach policy before it executes.

Vendor categories to evaluate

What to ask in vendor demos

Common pitfalls

1. Buying a forecasting model before fixing the data. If bank transactions are still being categorized inconsistently across entities, no model will produce a stable forecast. Categorization and master data work comes first.
2. Treating directional FX or rate prediction as a use case. Corporate treasury is a risk management function, not a trading desk. Models that promise alpha on macro variables rarely survive contact with policy review.
3. Ignoring the audit trail. SOX, internal audit, and external auditors will ask how AI-influenced decisions were made. If the system cannot produce a defensible record, it cannot be used for control-relevant work.
4. Skipping tax review on in-house banking changes. Pooling and netting structures have tax consequences. AI-suggested optimizations must go through tax before execution.
5. Letting one bank's copilot become the system of record. Convenient at first, painful at the next RFP. Keep the analytical layer independent of any single banking relationship.