AI agents for finance: how agentic treasury works in practice
AI agents for finance are software systems that read your financial data, plan the next step within your limits and carry out tasks on their own. Typical tasks include categorizing transactions or drafting collection follow-ups, while a person still approves anything that changes records or moves money. In a Gartner survey of 183 CFOs published in June 2026, 84% of finance organizations had implemented AI or planned to. Only 7% reported a high or very high impact.
Treasury closes that gap faster than most finance areas. Bank data refreshes every day, and much of the work around it repeats from one week to the next. This guide defines agentic treasury and lists the use cases that work today. It then shows how agents reach your data, how to keep them under control and how one US healthcare group built the data foundation they need.
What are AI agents for finance?
An AI agent for finance is a program built on a large language model that pursues a goal across several steps. It calls tools to fetch or update data and asks for approval at defined checkpoints. Where a rule fires on one condition, an agent decides which steps a request needs. Ask it why cash dropped last week, for example, and it pulls the balances, compares them with the forecast, finds the categories behind the gap and writes a short explanation.
Three kinds of tools often get grouped under "AI in finance", and they fit different jobs.
Approach | What it does | Who starts the task | Treasury example | Fits when |
|---|---|---|---|---|
Rule-based automation | Applies a fixed rule to every matching item | A schedule or a trigger | Every outflow labeled "AMZ" goes to Office supplies | The pattern never changes |
AI assistant (copilot) | Answers questions and drafts outputs on request | A person, prompt by prompt | "What is our cash position per entity today?" | Ad-hoc analysis and one-off reports |
AI agent | Chains several steps toward a goal and stops for approval | A goal, a schedule or an event | Every Monday, pull the cash briefing, explain the variances and post it to the CFO's Slack | Recurring multi-step work with clear limits |
Most finance teams end up using all three. Rules handle stable patterns, an assistant answers the CFO's questions, and agents take over routines that used to need a person to copy data between tools.
What is an agentic treasury?
Agentic treasury means letting AI agents run parts of the daily cash cycle (monitoring, categorization, reconciliation, forecasting and reporting) under the treasurer's rules and approval. The treasurer sets objectives and limits. The agents do the preparation and flag what needs a decision.
Treasury teams already see this coming. In the 2026 AFP Treasury Benchmarking Survey, published in September 2026 from 425 practitioner responses, 30% of respondents ranked AI and automation among their top five priorities. And 38% named managing AI opportunities and risks as one of their most significant challenges. BCG's Global Payments Report 2026 describes the same shift from the banks' side. In its words, AI agents are "starting to operate directly on behalf of the CFO or treasurer."
What can AI agents do in finance and treasury today?
AI agents in finance work best on tasks that repeat, follow known rules and can be checked quickly. The six treasury use cases below are documented in Agicap today. Availability depends on your subscription, your banks and your settings.
Daily cash monitoring across entities
An agent can return the current bank balance across all your entities, per bank account or per entity, in the currency you choose. It can also project daily, weekly or monthly balances from actual bank movements and expected transactions, with or without credit lines. Instead of logging in to each bank portal, the treasurer asks one question and gets one consolidated answer.
Transaction categorization and bank reconciliation
Categorization makes a good first use case, because errors are visible and easy to fix. In Agicap, a global audit has the AI propose categories and the rules to create, modify or delete, along with how many transactions each change affects. The team accepts or ignores each recommendation, and it can teach the AI why it refused one so the next audit takes the reason into account.
Bank reconciliation follows the same logic. A machine learning model trained on anonymized historical reconciliations suggests matches between bank transactions from the last 60 days and expected transactions due in the next 60 days. That covers one-to-one and one-to-many matches as well as matches with an amount gap. Rejected suggestions are remembered and never proposed again.
Cash flow forecasting and variance analysis
Agicap AI can generate a cash flow prediction from your transaction history with three levels per period. Those levels are optimistic, median and pessimistic. For the short-term forecast, weekly snapshots let an agent compare what you forecast four weeks ago with what happened, category by category. The output shows which categories drove the gap, so the treasurer spends time on the cause instead of rebuilding the spreadsheet.
Accounts receivable and collections
On the receivables side, an agent can list a client's overdue invoices and compose a follow-up email with the invoices attached. It then sends the email through the Agicap Accounts Receivable module. Each follow-up stays logged in the client's history. Payment matching on the cash application side works the same way, and our guide on how AI cash application software matches payments to invoices covers it in detail.
Reporting and board packs
Reporting eats hours every month, which makes it an easy place to measure gains. An AI assistant can read a custom dashboard, including every chart, counter and table, and write a briefing on top of it. A recurring agent can then pull the latest figures each month and update the financial section of the board deck.
Anomalies and payment risk signals
Agents also help with control work. When the AI flags an anomaly, it brings the context the reviewer needs to decide. In the Payments module, each beneficiary carries an uncertainty flag when its name or bank account changed recently and no payment has gone through since. Agicap documents this as a fraud-risk signal, and an agent can list every flagged beneficiary before a payment batch goes out.
How do finance AI agents connect to your treasury data?
Agents reach treasury data in two ways. One is an assistant built into the treasury platform. The other is a standard connector that lets an external AI assistant query the platform with your permissions.
Agicap offers both. The Smart Assistant, currently in beta, runs inside Agicap with no separate AI subscription and is powered by Claude, by Anthropic. Users choose between three models depending on the depth of analysis they need. The Agicap MCP Server uses Model Context Protocol, an open standard developed by Anthropic. It gives AI assistants such as Claude live access to cash positions, transactions, forecasts, payables and receivables through one secure endpoint.
Access follows the rights you already manage in Agicap. An organization administrator enables the MCP and chooses between read and write mode. Each user then gets read access (queries only) or full access (write actions). The AI never sees data that the person using it cannot see in Agicap. For setup and real prompts, see a step-by-step look at the Agicap MCP inside Claude.
How do you keep AI agents under control?
You keep AI agents under control by limiting what they can see, separating read from write access, and requiring human approval for every change to records or money. BCG's 2026 report recommends "authorization, monitoring, and audit controls designed specifically for nonhuman actors," and the same logic applies inside a finance team.
A practical way to set this up is to decide, task by task, what the agent may do alone and what a person must approve. The split below is a recommended starting point.
Task | The agent can do alone | A person approves |
|---|---|---|
Answer a question about cash position or balances | Yes | No |
Explain forecast variances | Yes | No, the treasurer reviews the conclusion |
Recategorize transactions in bulk | Propose the changes | Accept or ignore each recommendation |
Reconcile bank transactions | Suggest matches | Accept or reject the suggestions |
Replace imported transactions | Prepare the import | Explicit confirmation before anything is overwritten |
Send a collection follow-up | Draft the email and attach invoices | Check the recipient and the tone before sending |
Release a payment | Track remittance status | Your usual signatories |
Two further habits help. Start every new use case in read-only mode for a few weeks, then grant write access to that use case alone. And keep a weekly review of changes made by AI, since a colleague may not realize an agent recategorized transactions on someone else's behalf. On data protection, Agicap's built-in AI does not train on customer data. When you use an external provider such as Claude, that provider's terms apply, and Agicap recommends an enterprise plan.
Customer case: how built the data foundation
University Medical Partners (UMP) shows what agents need before they can help. The community medical group partners with Stanford Medicine, runs 40 clinics across two entities and generates $160 million in revenue.
Before Agicap, the finance team exported data by hand from Wells Fargo and Goldman Sachs and forecast in static Excel files. Because those forecasts were unreliable, UMP held high balances in non-interest-bearing accounts to protect payroll and accounts payable. Agicap now connects both banks directly, categorizes transactions with AI based on general ledger logic, and builds 13-week and 12-month rolling forecasts that update from actual bank movements.
The results are measurable. Categorization time fell by 85%, from about two hours to 10 to 15 minutes a month. Automated dashboards save the Director of Finance and Analytics more than four hours of reporting each month. Better forecasts also let UMP run leaner cash buffers and move more money into its investment portfolio. "So we've completely removed our Excel-based reporting," says Ben Courtney, Director of Finance and Analytics. Clean, categorized and connected data of this kind is what any AI agent for finance reads first.
Worked example: what a better forecast frees up
Take a mid-market company that keeps a $6 million cash buffer in operating accounts that pay no interest. The buffer exists because its forecast regularly misses payroll and vendor payment swings. After three months of weekly variance analysis, the treasurer sees that the worst week of the quarter needed $4 million. The team moves the other $2 million into an investment account.
Assuming an illustrative 4% annual yield, that move earns the following.
$2,000,000 × 4% = $80,000 a year
Time adds up too. UMP's categorization dropped by about 1 hour 45 minutes a month, or 21 hours a year (1.75 hours × 12). Agents give that time back to the treasurer for analysis.
How to start with AI agents in finance: a four-step plan
Starting with AI agents in finance takes four steps. Fix the data, pick a few use cases, open access gradually, then measure.
Fix the data foundation. Connect your banks, set categorization rules and reconcile daily, because an agent reasoning on messy data produces confident wrong answers.
Pick three to five use cases. Gartner's June 2026 advice is to pilot three to five at a time. Choose tasks that repeat weekly and that a person can check in minutes, such as the Monday cash briefing or bulk recategorization.
Start read-only. Run each use case with query access first, then grant write access to that use case alone, with approval on every change.
Measure and extend. Track hours saved, forecast variance and idle cash moved, then add the next batch of use cases.
Plan for training as well. The 2026 AFP survey found that only 34% of treasury practitioners consider themselves effective in AI knowledge, while 50% say it matters. Give the team time to practice writing prompts and reviewing agent output before you extend write access.
See AI agents work on your own treasury data
The fastest way to judge finance agents is to run them on your own cash data for one week. Pick the routine your team dislikes most, often the Monday cash briefing or month-end categorization, and measure the hours it takes today.
Frequently Asked Questions (FAQs) about AI Agents for Finance
Which AI agent is best for finance?
The best AI agent for finance is the one connected to your actual bank and accounting data with your existing permissions. A general chatbot without that connection can only reason on what you paste into it.
For treasury, look for three things:
The agent should read live cash positions across entities
Respect user rights by entity
Require approval before changing records.
Those criteria matter more than which model sits underneath.
Is there an AI tool for treasury teams?
Yes. Treasury platforms now include AI for categorization, reconciliation suggestions and cash flow prediction, and some expose their data to AI assistants through Model Context Protocol. In Agicap, the Smart Assistant works inside the platform, and the MCP Server connects external assistants such as Claude. Spreadsheets and ERP budgeting modules rarely offer this, because they hold no live bank data.
Are AI agents safe to use with bank data?
AI agents are safe with bank data when access is scoped and changes need approval. Check how the connector authenticates users, whether the AI inherits each person's permissions, and whether write access can be switched off. Also check whether your data trains any model. Agicap's built-in AI does not train on customer data, and the platform is ISO 27001 certified.
