For a short-term cash forecast, start with a spreadsheet or code-based model you can inspect. For repeated account forecasts, compare a finance platform with your current method on past periods. For connected agents, test the data updates as well as the numbers. These are different jobs, so they need different tests.
A client can pay late without changing your month-end total. Payroll may still arrive before the money does. A useful forecast must show that gap, its source, and the assumptions behind it.
This guide compares nine tools by the work they support. It includes a cash-timing exercise and a historical comparison you can copy. The examples are fictional. Provider sources were checked on 5 October 2026; we did not run a shared product trial or connect company accounts. This is a shortlist, not an accuracy ranking.
Choose the forecast before choosing the tool
Scroll sideways to see all columns.
| Your job | Required inputs | Result to review |
|---|---|---|
| Short-term cash | Opening cash, dated receipts and payments | Balances over time and any low point |
| Revenue or expense estimate | Historical actuals and relevant business assumptions | Forecast for a stated period, method, and error against a baseline |
| Business scenario | An approved model and a changed input | Which linked numbers change, by how much, and why |
| Full financial plan | Account rules, operational drivers, and source actuals | Linked profit, balance sheet, and cash forecasts |
A revenue forecast is not a dated cash schedule. Jirav's three-way financial guide gives an example where a sale is recorded in one month and collected in the next. Profit and bank cash answer different questions.
For a small cash task, an assistant may help build the model. For repeated planning across accounts and departments, financial planning and analysis (FP&A) software may help manage the process. The Association for Financial Professionals' learning resources cover rolling forecasts and baseline comparison. Choose the process your team needs, then assess the AI feature within it.
Nine tools at a glance
The categories below are our assessment of documented features. A product's presence here does not prove it will improve your forecast.
Scroll sideways to see all columns.
| Tool | Role worth testing | Evidence to request |
|---|---|---|
| ChatGPT data analysis | Calculate and explain a model from files | Code, source rows, and checked balances |
| Claude code execution | Create an editable forecast workbook | Working formulas in the saved file |
| Copilot in Excel | Build or edit an existing Excel model | A changed input recalculating the right cells |
| Gemini in Sheets | Build a formula-based Sheets model | Correct references and scenario changes |
| Jirav Auto-Forecast | Create an income-statement baseline | Account forecast plus reviewed cash-conversion assumptions |
| IBM Planning Analytics | Forecast regular time series inside a planning system | Method, suitable history, and error on unseen periods |
| Planful Projections | Generate account baselines from historical performance | Source actuals, model choice, and reviewed adjustments |
| Pigment Predictions | Add statistical or machine-learning forecasts to a model | Backtest measures and linked scenario results |
| Concourse | Maintain finance workflows across connected data | Source mapping, repeat-import handling, and review trail |
Shortlist two options for the same job. A cash spreadsheet and an account-forecasting platform should not receive one overall “accuracy” score from different tasks.
File and spreadsheet assistants
1. ChatGPT: calculations you can inspect
ChatGPT data analysis supports uploaded spreadsheets and CSV files. For some tasks, it writes and runs Python to calculate results and create tables or charts.
A useful trial: Ask for a weekly cash schedule, the input rows used, the running calculation, and a separate late-payment scenario. Inspect the code and outputs before using the result.
Main limit: The Python analysis environment cannot make external web requests or API calls. Provide files or an available connected source. That boundary applies to the calculation environment; it is not a claim that ChatGPT has no connected apps. Access depends on the account and workspace.
2. Claude: an editable workbook as the result
Claude's code execution and file creation can produce Excel files with working formulas. Anthropic also provides spreadsheet skills for file work.
A useful trial: Request separate Inputs, Base, Delayed Payment, and Checks sheets. Open the saved workbook, inspect formulas, and change a receipt yourself. The downstream balances should update without a new chat answer.
Main limit: A generated file still needs review. Confirm that code execution is enabled and that the workspace settings fit your data policy. Use the current file guide; Anthropic's older analysis-tool announcement notes its replacement by code execution.
3. Copilot in Excel: help within your existing model
Copilot in Excel can generate formulas, summaries, charts, and other analysis. Its Python-backed answers can expose code, while advanced analysis can create refreshable Python cells.
A useful trial: Work in a copy of your approved workbook. Ask for the cash formula, then move one receipt. Check whether later balances update and whether unrelated inputs stay unchanged.
Main limit: Microsoft documents static tables and charts in its direct-answer experience. A correct inserted result may need a new analysis rather than recalculating with the workbook. Check whether your result is static, a formula, or a refreshable Python cell. License and feature availability also need checking.
4. Gemini in Sheets: formulas for a shared forecast
Gemini in Sheets can create formulas, tables, charts, and analysis. It requires an eligible Google Workspace or Google AI plan.
A useful trial: Ask for separate input and calculated columns. Inspect references, copy formulas down, and change an amount and payment week. Check the whole schedule, not only the chart.
Main limit: Google's AI function generates text and other responses; it is not a substitute for normal arithmetic formulas. Keep money calculations in formulas you can test. Generated explanations should describe the model's output.
Finance platforms and connected agents
5. Jirav: account forecasts with separate cash rules
Jirav Auto-Forecast creates income-statement forecasts from historical account data. Its method uses the available history and can include trends and seasonality. Receivable and payable settings govern how sales and expenses become cash.
A useful trial: Forecast an account from past actuals, then show how collection and payment assumptions affect the linked cash plan. An account projection does not tell you the exact date a particular customer will pay.
Main limit: Jirav states two full years of usable history for trend and three for seasonality. Confirm which accounts have enough usable data. Its subscription guide uses Essentials and Enterprise, while its business pricing page lists Starter, Pro, and Enterprise, with Starter at $10,000 per year when checked. Ask the provider to map the features you need to your quote; do not assume the names match.
6. IBM Planning Analytics: forecasting inside a planning model
IBM Planning Analytics Workspace documents time-series forecasting and a natural-language Planning Analytics Agent. The agent interface and the forecasting calculation are separate capabilities.
A useful trial: Use a regular account series and inspect the forecast preview. Compare its output with actuals from periods hidden during fitting.
Main limit: IBM's requirements distinguish forecasts using one series from forecasts using additional variables. For a single series, they specify at least twice the forecast horizon in history and two seasonal cycles for detecting seasonality. These are setup requirements, not accuracy guarantees. Its current preview explanation also says the prediction-accuracy indicator reflects fit to historical data. That is not the same as performance on unseen periods. Check your deployed version.
7. Planful: historical baselines for a planning cycle
Planful Projections describes forecasts based on historical performance, with model selection and scenario changes. Its broader AI platform supports finance analysis and planning.
A useful trial: Select one account, trace its actuals, generate a baseline, and show a reviewed adjustment. Compare the forecast with a simple method over the same past periods.
Main limit: A product page's accuracy language is not evidence for your accounts. Ask which model and assumptions created the result, which features your plan includes, and how the team keeps an approved forecast separate from proposed changes.
8. Pigment: Predictions within a linked model
Pigment AI includes agents for model building and analysis. Predictions is a separate forecasting feature. Scenarios compare changed inputs or formulas within the planning structure.
A useful trial: Ask for the forecasting method, a backtest, and one scenario change through linked metrics. Keep a chosen business assumption separate from a learned prediction.
Main limit: The Predictions guide says activation requires a support request. Confirm access rather than assuming that access to AI agents includes Predictions. Its accuracy measures are optional and use a held-back portion of history. Request them during evaluation. Its model-comparison guide uses a seasonal baseline, but those published benchmark results do not rank models for your business.
9. Concourse: keeping a connected finance model current
Concourse describes finance agents that connect business systems, refresh forecasts, and run scenarios. It says calculations run in code.
A useful trial: Import one approved data set, change a source record, and trace the effect on the forecast. Import it again to check that nothing is counted twice. Ask who approves the changed version.
Main limit: Running code does not prove that the source mapping, formula, or forecast assumption is correct. The page is a provider description, not independent accuracy evidence. Judge the complete source-to-forecast path and its upkeep cost.
Test one: does the model calculate cash correctly?
Use this fictional four-week schedule. Opening cash is $20,000. All figures are dollars, with one currency and no unlisted cash movements.
Closing cash = opening cash + receipts − payments
Next week's opening cash = this week's closing cash
Scroll sideways to see all columns.
| Week | Base receipts | Payments | Base closing cash | Closing cash when $4,000 is delayed |
|---|---|---|---|---|
| 1 | $12,000 | $8,000 | $24,000 | $24,000 |
| 2 | $9,000 | $14,000 | $19,000 | $15,000 |
| 3 | $7,000 | $10,000 | $16,000 | $12,000 |
| 4 | $11,000 | $9,000 | $18,000 | $18,000 |
Move $4,000 from week 2 to week 4. Delayed receipts become $12,000, $5,000, $7,000, and $15,000. Total receipts remain $39,000 and payments $41,000. Both scenarios end at $18,000.
The lowest weekly closing balance changes from $16,000 to $12,000. Against a fictional $14,000 minimum cash reserve chosen by the owner, the delayed case falls below the reserve in week 3. The reserve is an extra review threshold; it is not another payment to subtract.
A passing result shows every week's calculation, preserves the delayed receipt, identifies the low point, and recalculates after an input change. Passing proves the model handles these supplied inputs. It does not prove it can predict payment dates.
Weekly closing balances can still hide a cash dip
In the base case, week 2 opens at $24,000. If all $14,000 of payments leave on Monday and the $9,000 arrives on Friday, cash falls to $10,000 before recovering to $19,000. The weekly closing figure alone misses that dip.
Use daily dates when the order of payments matters. For a longer planning format, AFP's cash-forecasting presentation, available as a current PDF, shows daily and weekly views over a roughly 13-week horizon. Four weeks here is an exercise, not a complete planning policy.
Copy this prompt for an assistant
Build a four-week cash model using formulas or executable code.
Opening cash: $20,000.
Receipts by week: $12,000; $9,000; $7,000; $11,000.
Payments by week: $8,000; $14,000; $10,000; $9,000.
Show opening cash, receipts, payments, and closing cash each week.
Create a second scenario: move $4,000 of week 2 receipts to week 4.
Keep total receipts unchanged. Reconcile the final balance to totals.
Report the lowest weekly closing balance and weeks below a $14,000 reserve.
Show formulas or code, input references, and assumptions.
Do not describe weekly closing cash as the lowest daily cash balance.
Do not claim these inputs predict real payment dates.
For a platform demo, ask the provider to show the equivalent result inside its product. If the product does not support dated cash, test its account forecast instead and record that scope limit.
Test two: are the estimates better than your baseline?
A backtest asks a model to forecast past periods using only information available before those periods. It checks estimation quality separately from arithmetic.
Pick one target and horizon, such as next month's sales or next week's closing cash. Test several past decision dates. For each date:
- Freeze the inputs. Keep later actuals, revised invoices, and plans agreed afterwards out of the data. For cash, preserve the payment expectations known then, not dates filled in after payment.
- Save the forecast before revealing actuals. Record the model, settings, assumptions, and input version. If you tune on these results, keep other later periods for a fresh test.
- Run the same baseline. For sales, this might be last month's value or the same month last year. For cash, use your existing dated schedule. Give it the same information and horizon.
- Compare error and missed risks. Record average error, largest miss, whether the forecast is usually high or low, and any missed reserve breach. Compare review time too.
Hyndman and Athanasopoulos' time-series evaluation guide explains testing at repeated past starting points without future observations. A model fitted to all past data has not passed this test.
A small example of how to judge the result
These three fictional forecasts of weekly closing cash illustrate the comparison. They are not product results or enough evidence to choose a model.
Scroll sideways to see all columns.
| Past forecast period | Actual closing cash | Candidate forecast | Existing-method forecast |
|---|---|---|---|
| A | $12,000 | $15,000 | $14,000 |
| B | $16,000 | $17,000 | $18,000 |
| C | $9,000 | $15,000 | $10,000 |
Mean absolute error (MAE) is the average size of the misses, ignoring whether they are high or low. The candidate's errors are $3,000, $1,000, and $6,000: MAE is about $3,333. The existing method's errors are $2,000, $2,000, and $1,000: about $1,667. The forecast-accuracy guide explains this measure and its limits.
With the same $14,000 reserve, actual cash is below it in A and C. The candidate flags neither period. The existing method flags C but misses A. In this example, the new tool gives larger errors and less useful warning. A more polished chart would not change that result.
Pigment's documented backtest uses the first 80% of a series for training and the last 20% for testing by default. Ask whether a provider's test matches your horizon and decision. A displayed score or confidence range alone does not answer that question.
Check the data and review process before a pilot
A forecast needs an opening balance reconciled to the chosen cash accounts, one currency or an explicit conversion method, and clear dates. Keep actual and planned rows separate. Once an invoice is paid, do not also leave its planned receipt in the future schedule.
For each row, keep an ID, amount, currency, expected cash date, actual-or-planned status, and source. Flag missing dates rather than guessing silently. Repeated invoice IDs need a rule: they may be duplicate imports or legitimate payment instalments. Ask the finance owner to define it.
Use sample or approved company data. Confirm file handling, retention, connected-source access, and your team's sharing rules before uploading financial records. Begin with read access and a separate working copy; payment execution is outside this forecasting trial.
Keep a compact review record:
Target and forecast horizon:
Input version and source cut-off date:
Model or calculation version:
Assumptions changed and reason:
Calculation checks:
Backtest result versus the existing method:
Largest error and missed reserve breaches:
Setup cost and recurring review effort:
Reviewer, approval date, and next review:
Choose the smallest setup that passes the relevant tests. A spreadsheet may be enough for one owner and a few dated cash lines. A planning platform may be worth testing when linked accounts, repeated updates, or several reviewers create work. A connected agent may help when collecting actuals is the main burden.
Start with one forecast and a named finance reviewer. Keep the approved model alongside the trial version. Continue only if the tool can explain changed numbers and improves the task enough to justify setup, subscription, and upkeep. Passing a past test does not guarantee the next forecast.