AI cash flow forecasting: what treasury and cash management tools actually predict (September 2026)

Why cash forecasting is still the hardest job in treasury

The Association for Financial Professionals (AFP) surveyed 425 treasury practitioners in May 2026 and published the results on September 15, 2026. Cash management and forecasting came out as the top priority, named by 63 percent. Cash and liquidity forecasting was also the most challenging activity, named by 49 percent, and AI and automation entered the top 5 priorities for the first time, cited by 30 percent.

The same survey shows where teams feel unprepared. 38 percent called managing AI opportunities and risks a challenge and 35 percent named using AI to automate manual processes. Respondents gave their policies on AI and emerging technology an effectiveness score of 2.9 out of 5, the lowest of all policy areas, against 4.4 for cash management policies.

Customers who pay late make the forecast harder still. In the 2026 Atradius survey of Western European companies, nearly 4 in 5 reported late payments and about a quarter of B2B invoices were paid after the due date. Each delayed payment moves a line in the forecast, which is one reason why receivables and collections tools now feed their payment predictions into cash planning.

Sources: [1][2]

What AI does inside a cash forecast: tag, model, explain

The first layer sorts the cash. Bank transactions arrive with cryptic reference texts, and the software assigns each one to a category such as payroll, supplier payments or tax. Treasury4 tags every inflow and outflow as operating, investing or financing cash flow the moment it arrives, and Tidely assigns bank transactions to categories based on historical patterns.

The second layer builds the forecast. According to its April 2025 announcement, Nomentia generates reference forecasts from bank and ERP history, selects and trains suitable models by itself and corrects anomalies and seasonal patterns. Palm says it combines more than 25 AI models on ERP and bank data, TIS recalibrates its forecasts from historical data as conditions change, and Treasury4 trains its models only on each customer's own history.

The third layer explains. Agicap describes an assistant that detects recurring patterns and identifies variances between actual and projected values, and Kyriba's TAI answers natural language questions about treasury data and analyzes variances. In a finance meeting, the useful answer is usually why the number moved since last week, and this layer is built to give it.

A fourth step stays with people: deciding what to do about a gap. Commitly states that its AI never replaces the official plan without being asked, and the agent products described further down follow the same rule of proposal first, approval second.

Sources: [3][4][5][6][7][8][9][10]

From forecasts to agents: what changed in treasury AI since 2025

Between 2025 and September 2026, treasury AI moved from better forecasts to software that proposes actions. Larger companies also bought their way into the category: Tipalti acquired the AI treasury start-up Statement in June 2025, and in October 2025 Ripple announced the $1 billion acquisition of GTreasury, a treasury platform with more than 1,000 customers in 160 countries.

The agent features follow one pattern: the software prepares, a person approves. Kyriba describes TAI as recommend, then decide, with explicit sign-off for sensitive transactions. Atlar runs agents for cash positions, payment briefs, reconciliation and cash sweeps that reference uploaded treasury policies and take no action without approval. Ripple Treasury's GSmart agents, presented on September 10, 2026, propose a specific action, cite the policy clause behind it and wait.

Some vendors let customers build their own agents or bring their own model. Nilus turns a plain description into a scheduled agent with its Agent Builder and lists more than 160 prebuilt agents. Treasury4 and Commitly offer an MCP server, a standard connection through which general AI models such as Claude or ChatGPT can query the treasury data, with access control staying inside the treasury tool.

Sources: [11][12][9][13][14][15][16][3][10][5][1]

Treasury tools in our directory: enterprise system or midsize cash tool

Our directory lists 26 tools built mainly for treasury and cash management, plus accounts payable and receivables suites that carry a treasury or cash module. Most vendors in the category are headquartered in the United States (24 of 52); the rest come mainly from Europe.

The group splits by buyer. Enterprise treasury management systems such as Kyriba, TIS and Ripple Treasury cover payments, bank connectivity and risk alongside forecasting, and TIS says it connects to more than 11,000 banks. Nomentia reports more than 1,400 business customers in more than 80 countries. Midsize tools such as Agicap, Commitly and Tidely concentrate on cash visibility and planning.

Data location differs as much as scope. 7 tools in the category can host data in the EU, among them Tidely, which says it is developed and hosted in Germany. Only 17 serve companies with fewer than 50 employees according to our data, so small firms have a short list to compare.

Architecture is another dividing line. Treasury4 runs each customer on a single-tenant setup with a dedicated Snowflake environment, so its models learn only from that customer's transactions. Palm does not replace a treasury system at all: it sits on top of an existing TMS or ERP as an AI layer for forecasting, categorization and liquidity alerts. For a company that already runs a treasury system, a layer like that can be the smaller project.

Sources: [17][7][5][4][3][6]

Cash forecasting software pricing: 3 treasury specialists publish list prices

Among the tools built mainly for treasury, only Trovata, Tidely and Commitly showed prices in September 2026. The others quote after a demo; Atlar, for example, lists 4 plans without amounts and includes payments and cash flow forecasting only from its Professional plan.

Trovata's base package costs $24,000 per year and includes 1 bank connection, up to 100 accounts, 1 million stored transactions and 10 users, with machine learning forecasts and Trovata AI. Its full treasury management system is quoted separately.

Tidely and Commitly price for smaller companies. Tidely charges 55, 105 or 265 euros per month for Starter, Professional and Business with monthly billing, or 39, 75 and 189 euros on a 2-year term, with AI categorization and AI forecasting in every plan; its DATEV connector for the German accounting system costs 35 euros extra per month, or 25 euros on the 2-year term. Commitly charges 55, 105 and 195 euros per month plus VAT for Basis, Business and Professional with monthly billing and adds its MCP server for 15 euros per month below Professional.

Enterprise vendors such as Kyriba, Nomentia, TIS and Ripple Treasury publish no amounts, and 33 of the 52 tools in the category quote on request.

Trials are rare at the enterprise end but common among the smaller tools. Tidely offers 7 days and Commitly 14 days without a credit card, and Commitly cuts its prices by 40 percent for 6 months on a 12 month term or by 50 percent for 12 months on a 24 month term.

Sources: [18][19][4][20][17]

What to check before you buy a cash forecasting tool

Start with the data the forecast needs. Ask how many of your banks the vendor connects to directly, how it reads your ERP or accounting system and how often balances refresh. A forecast that starts from last week's statement cannot warn about this week's shortfall.

Then test the model on your own history. Ask for a back test over the past 12 months that compares forecast and actuals by week and by category. Vendor figures such as Palm's claim to cut variance by more than 60 percent, or the forecast accuracy gains of more than 30 percent reported by Nilus, come without a published method and cannot replace that test.

Finally, check control and security. Agent features should show the data and the policy behind each proposal and wait for approval. Since respondents in the AFP survey rated their own AI policies lowest of all policy areas, write that policy before switching agents on. Certifications vary too: 26 tools in the category list SOC 2 and 16 list ISO 27001.

Sources: [6][16][1][17]