What problem does it solve?
Where cash forecasting runs on spreadsheets, someone pulls bank reports and extracts, maps transactions to categories by hand and rolls the result forward, cycle after cycle. The published cases describe exactly this. J.P. Morgan's case study on Prysmian says its two person North American treasury team relied on spreadsheets for cash forecasting and daily reconciliation, and its case study on Domino's describes manual data entry and categorisation that consumed valuable time. The work is slow, depends on a few people and can produce forecasts too coarse to act on, so companies hold buffers of idle cash, as Amtrak did before improving its projections.
Much of the data to do better already sits in the bank's systems, because the company's payments and receipts pass through them. The opportunity is to categorise those flows automatically, learn their patterns, and give the treasurer a forecast and a way to question it, while the decision to fund, sweep or invest stays with the treasurer.
How does it work?
- Connect the data. Account and transaction data from the bank's platform, and optionally other banks and the ERP, flow in daily.
- Categorise flows. Models sort every transaction into the company's own categories (payroll, suppliers, card receipts, taxes, intercompany), and the treasury team corrects the ones that are wrong.
- Forecast. Per category, account and currency, the tool projects positions over a chosen horizon and shows the forecast against actuals as they arrive.
- Ask in plain language. A conversational layer turns questions such as "show balances by account for the last three months" into queries and charts over the same data. This is the least mature step: J.P. Morgan describes its treasury analytics assistant as a prototype, while Bank of America offers CashPro Chat, a virtual service advisor, in its CashPro platform.
- Recommend within limits. J.P. Morgan describes, as a future direction, GenAI that gives treasurers recommendations and might one day act on their behalf within parameters they set. A design option that follows from this, not a feature any cited bank has shipped, is to run scenarios (a delayed receipt, a currency move) and propose sweeps or investment of idle balances for the treasurer to approve.
- Audience
- Employee facing
- Autonomy
- Assist
- Adoption
- Early adopters
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Productivity gain | Too few to pool | 50% to 90% Not pooled: up to 90% | 2plus 1 up to | 1 organization, 1 vendor |
| Cost savings | Not pooled | about USD 100,000 | 1 | 1 vendor |
| Users served | Not pooled | about 2500 | 1 | 1 organization |
Value drivers: Employee productivity, Speed and cycle time, Risk and loss reduction, Revenue growth.
Indicative value
A mid sized company with a treasury team of five
USD 40,700 to USD 310,400
Treasury time released plus yield on released idle cash per year
How this is calculated
Formula: teamSize * hoursPerWeek * weeks * gain * hourlyCost + releasedCash * netYield. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Treasury staff involved in forecasting teamSize, people | 5 | 5 | The reference company. |
| Hours per person per week on data preparation and forecasting hoursPerWeek, hours per person per week | 6 | 12 | Editorial assumption, replace with your own time study. |
| Working weeks per year weeks, weeks | 46 | 46 | Editorial assumption. |
| Share of that time saved gain, fraction of time | 0.25 | 0.4 | Editorial assumption, applied only to the data preparation and forecasting hours above. Set below the single cases on this page, which J.P. Morgan reports as half of one Prysmian team member's manual forecasting and reconciliation time and Domino's weekly manual data cleanup down by up to 90%. |
| Loaded cost of a treasury hour hourlyCost, USD per hour | 60 | 100 | Editorial assumption, replace with your own loaded cost. |
| Idle balance released for investment thanks to better forecasts releasedCash, USD | 1,000,000 | 5,000,000 | Editorial assumption. Amtrak describes investing balances it had set aside once forecasts improved; size this from your own buffers. |
| Net yield on the released balance netYield, fraction per year | 0.02 | 0.04 | Editorial assumption, replace with your own short term investment yield. |
What it leaves out: Leaves out the cost of the tool, the value of avoided overdrafts or short term borrowing, and any fees the bank charges. The released cash figure depends heavily on how conservative the current buffers are.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Bank of America
United States · Banking · 2022
Bank of America launched a cash forecasting solution that uses machine learning in January 2022, offered to business clients in its CashPro platform as CashPro Forecasting. The bank reports fast adoption rather than accuracy: from the first half of 2022 to the first half of 2023, new client enrollments rose 141%, active users 105% and sign ins among those users 375%. The same release describes enhancements to CashPro Chat, a virtual service advisor in CashPro that now uses the same AI and machine learning capabilities as Erica, the bank's consumer assistant.
No outcome disclosed.
Prysmian
Italy · Manufacturing · 2025
Prysmian, the cable manufacturer, used J.P. Morgan Payments Cash Flow Intelligence to automate cash visibility and forecasting for ten operating companies and thirteen bank accounts in North America, run by a treasury team of two. According to its treasurer, the forecast horizon grew from 30 to 91 days, manual reconciliation of more than 3,000 daily transactions was removed and answers to senior management questions came ten times faster.
- Cost savings: about USD 100,000, per year, estimated labour cost
"Contessa reports that, post-implementation, Prysmian maintains a <1% error rate, saves $100,000 annually and reduced workload by 10 hours weekly"
Claimed by: vendor - Productivity gain: about 50%, one treasury team member's manual forecasting and reconciliation time, about 10 hours a week
"Discover how Prysmian automated global cash flow forecasting, reduced manual work by 50% and saved $100K annually with J.P. Morgan Payments Cash Flow Intelligence."
Claimed by: vendor
Domino's Pizza
United States · Retail and ecommerce · 2024
Domino's treasury team adopted J.P. Morgan Payments Cash Flow Intelligence to aggregate, categorise and reconcile cash flows across a franchise model of more than 20,500 stores in 90 markets and a securitised debt structure. The team runs weekly cash reviews and forecast updates in the tool and reports forecasts that aligned closely with its 2024 budget.
- Productivity gain: up to 90%, weekly manual data cleanup
"According to Domino’s Treasury Team Leader, Nancy Romain, the team’s weekly manual data cleanup efforts were reduced by up to 90%."
Claimed by: vendor
Amtrak
United States · Logistics and transportation · 2023
Amtrak's treasury, which has to plan around the three large installments in which government funding arrives each year, went live with J.P. Morgan Payments Cash Flow Intelligence in May 2023 after testing it in beta. Separating daily card receipts, monthly partner receipts and infrequent federal receipts into their own categories improved projection accuracy, which let the team invest balances it had previously set aside as a buffer. No figures are published for the accuracy gain or the income.
No outcome disclosed.
JPMorgan Chase
United States · Banking · 2023
J.P. Morgan Payments offers Cash Flow Intelligence, a machine learning tool in its J.P. Morgan Access platform that categorises a corporate client's payment flows and produces cash forecasts. In a Bloomberg interview relayed by CTMfile, the bank's head of data and analytics for wholesale payments said that about a year after launch roughly 2,500 corporate clients used it free of charge, and Bloomberg reported that some had cut manual work in categorising and visualising payment flows by nearly 90%, while liquidity decisions stay with people. Separately, the bank built a prototype conversational analytics assistant that lets treasurers query their payments data in plain language.
- Users served: about 2500, corporate clients, about a year after launch
"He observed that since the AI tool was introduced about a year ago, approximately 2,500 JPMorgan corporate customers are currently using the product for free."
Claimed by: organization - Productivity gain: about 90%, some corporate clients, manual work in categorising and visualising payment flows
"Dubbed, Cash Flow Intelligence, the artificial intelligence (AI)-aided cashflow management tool launched by the largest US bank, JPMorgan Chase & Co. has helped some of its corporate clients vastly reduce their manual work by nearly 90%, as was reported last week by Bloomberg."
Claimed by: organization
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Transaction history of at least a year across the main accounts
- An agreed category scheme for inflows and outflows
- Known large or irregular flows (tax dates, dividends, funding installments) entered as events
Systems to integrate
- Bank cash management platform or multibank data aggregation
- ERP and treasury management system for payables, receivables and plans
- Market data for foreign exchange and investment rates
Complexity: Medium
The models are well understood; forecast quality depends on data coverage. Treasuries with many banks and entities need data from all of them, and categories must match how the company thinks about its cash.
- 1
Baseline the current forecast
Record how long the weekly forecast takes and how far it has been from actuals per category over the last quarters, so improvements can be measured.
- 2
Agree categories with the treasury team
Define categories that match decisions (payroll, suppliers, receipts, taxes, intercompany), and separate large irregular flows so they do not distort the daily pattern, as Amtrak did.
- 3
Run in parallel
Run the AI forecast next to the spreadsheet for several cycles and compare both with actuals before the team relies on it.
- 4
Add the conversational layer on governed data
Let treasurers query the same governed data in plain language, with every answer showing the query and data it used.
- 5
Keep actions behind approval
If the tool suggests sweeps or investments, route them as proposals with limits the treasurer sets, never as automatic instructions.
Guardrails
- Forecasts are decision support; no payment, sweep or investment is executed without explicit approval
- Forecast accuracy per category is shown next to the forecast, not hidden in a report
- Conversational answers show the underlying query and data source
- Drift monitoring on categorisation and forecast error, with an owner who acts on alerts
KPIs to instrument
- Hours per week spent on forecast preparation, before and after
- Forecast error per category and horizon against actuals
- Share of transactions categorised automatically without correction
- Idle balances invested or buffers reduced as a result of better forecasts
Human in the loop
Treasurers review forecasts, correct categories and decide on funding, sweeps and investments. Any automated action runs only within limits the client has explicitly authorised, with a reversible audit trail. The bank's model owner monitors accuracy and drift across clients.
Common failure modes
- Garbage categories
- Early miscategorisation trains the model on wrong labels. Review categories closely in the first cycles and lock the scheme once stable.
- Blind spots from missing banks
- Flows at other banks or in cash pools are missing and the forecast looks precise but is wrong. Show coverage and warn when material accounts are absent.
- Over trust in a single number
- The team treats a point forecast as certain. Show ranges and scenario results, and keep liquidity buffers under human policy.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
Forecasting a company's cash flows is not listed in Annex III and makes no decision about a natural person, so the forecasting model itself carries no obligations beyond AI literacy (Article 4). The conversational layer interacts directly with treasury staff, so under Article 50(1) they must be informed that they are dealing with an AI system unless that is obvious from the context. Without a conversational layer the use case is minimal risk.
Rules that apply
Guidance
- NIST AI Risk Management Framework (NIST, North America). A practical structure to map, measure and manage the accuracy and drift risks of a forecasting model offered to clients.
- MAS Guidelines for Artificial Intelligence (AI) Risk Management (Monetary Authority of Singapore, Asia Pacific). Consultation paper of November 2025 proposing supervisory expectations for AI inventories, risk materiality assessment, evaluation and testing, and monitoring at financial institutions.
Controls to put in place
- Model inventory entry with an owner, validation results and drift monitoring
- Documented limits for any automated sweep or investment, set by the client
- Audit trail of forecasts, overrides and approved actions
- Clear client terms that forecasts are informational and not advice
Frequently asked questions
- How much manual work does AI cash forecasting remove?
- The published client cases are case studies by the bank that sells the tool and should be read as such. J.P. Morgan reports that Prysmian halved the manual forecasting and reconciliation work of one treasury team member (about 10 hours a week) and saved an estimated USD 100,000 a year, and that Domino's cut weekly manual data cleanup by up to 90%. According to a trade press report of a Bloomberg interview, about 2,500 corporate clients used the tool a year after launch.
- Does the AI move money on its own?
- It should not by default. Forecasts are decision support, and any sweep or investment is a proposal the treasurer approves, or runs within limits the client has explicitly authorised.
- Is this only for large corporates?
- The published cases are all large companies (Prysmian, Domino's, Amtrak), so there is no public evidence yet for smaller firms. The teams can be small, though: J.P. Morgan's case study describes Prysmian's North American treasury as a team of two. For a smaller company, the deciding factors are whether its bank offers such a tool and how much of its cash flows through that bank.
How to cite this page
Blits.ai AI Use Case Library, "AI cash flow forecasting for corporate treasury", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/treasury-cash-flow-forecasting. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published