Guides→AI & ML

The finance professional's guide to everything AI, ML, and ChatGPT

Why AI in finance is different in 2026

Most finance functions now run AI somewhere. Gartner’s 2025 AI in Finance Survey found 59% of finance functions using AI, up from 37% in 2023, and KPMG reports that 93% of U.S. companies will be deploying or scaling AI in their finance functions within the next 18 months.

Treasury has followed. In the 2026 AFP Treasury Benchmarking Survey, 30% of treasury teams put AI and automation among their top five priorities, the first time the category has made the list. The same survey shows the strain that comes with it. 38% of respondents name AI opportunities and risks among the most challenging tasks they face, and only 34% consider themselves effective in AI knowledge.

Adoption has run ahead of impact. In the same Gartner survey, 91% of finance leaders reported low or moderate impact from AI initially. The Citi Institute’s October 2025 practitioner guide found 82% of treasury teams still in the early stages of identifying and exploring GenAI use cases, with only 3% claiming successful deployment in their treasury processes. Crisil Coalition Greenwich reports that fewer than 1 in 10 large corporates have AI in daily treasury workflows such as forecasting and fraud detection, even as about 60% expect to increase their AI investment.

The gap between the spending and the results comes down to data and know-how. An AI tool is only as good as the bank data underneath it, and most treasury data still arrives through portals, files, and spreadsheets in formats no model can read without cleanup. And few treasury teams have anyone whose job is to understand what each kind of AI does, where it fits in a treasury workflow, and what has to be true before it can be trusted.

This guide is a primer on both. It explains the vocabulary without the hype, shows where AI is working in finance and treasury now, lays out what to get right before switching it on, and describes how Trovata AI 2.0 applies it to corporate cash. For the hands-on library of agent patterns, read “The AI Agents Playbook for Treasury Teams.” This guide is the one to read first.

The vocabulary, plainly

Five terms cover most of what a finance professional will hear in 2026. Each one does a different job.

Machine learning

Machine learning (ML) is software that finds patterns in data without being programmed for each case. Feed it enough historical transactions and it learns what a payroll run, a vendor payment, or an intercompany transfer looks like. In treasury, the familiar example is tagging. Trovata AI learns from the tags a team applies and suggests tags for new transactions as they arrive, so a transaction set that took an afternoon to categorize by hand is categorized on arrival and reviewed in minutes.

Narrow or predictive AI

Narrow AI performs a single task well and does not carry what it learns to other tasks. Most of the AI that has been in finance for a decade is narrow, and much of it is predictive. It takes history and projects it forward. A forecast baseline is the treasury example. Predictive models read past inflows and outflows by category and produce a starting forecast the team adjusts for known events, which replaces the blank spreadsheet the team used to fill from scratch.

Generative AI and large language models

Generative AI produces new content, usually text, in response to a prompt. The models behind it are large language models (LLMs), trained on enormous volumes of text so they can read and write in plain language. For finance, the value is in the interface. A treasurer can ask a question in a sentence and get a table, a chart, or a summary back, with no query language and no report builder. The caution is that an LLM on its own has no access to your bank data and no way to verify a number. It has to be connected to a system of record, and the system has to be clean.

Agentic AI

An agent is an LLM given a trigger, a set of data it can read, a written prompt that describes how a practitioner would approach a task, and a defined output. Where a chat assistant waits for a question, an agent runs on a schedule or an event and delivers a draft. The treasury example is a bank fee audit. The agent reads the month’s fee statements against the negotiated schedule, flags lines that do not match, ranks them by dollar impact, and hands the treasurer a draft dispute list to review before anything goes to the bank. Policy and action stay with the person, while the production work moves to the agent.

Model Context Protocol

The Model Context Protocol (MCP) is an open standard that lets an AI assistant reach the data and tools inside another system in a controlled way. Without it, every connection between an assistant and a data source is a custom build. With it, an assistant such as Claude, ChatGPT, Gemini, or Copilot can read balances, transactions, and reports from a treasury platform through one published interface, with permissions set by the platform. MCP is why a treasurer in 2026 can ask a general-purpose assistant about last week’s cash position and get an answer drawn from live bank data.

Where AI is working in finance and treasury

The applications below are in production at banks and corporates today. Most run on predictive AI. The newer ones layer generative and agentic AI on top.

Forecasting baselines

Predictive models read historical balances and categorized transactions to produce a forecast baseline across short- and long-term horizons. The team then adds what the model cannot know, such as a planned acquisition or a customer that pays late every March. The result is a forecast built from data and adjusted by judgment, with variance analysis to show how close it landed each period.

Data normalization and tagging

Every bank formats data differently. Before any analysis can run, balances and transactions have to be pulled into one format with consistent fields. On a connected platform this happens as data arrives through APIs and file-based reporting, and AI then classifies each transaction by type, counterparty, and entity. Tags become the building blocks for reporting and forecasting, so a question such as “what did we pay in bank fees across all entities last quarter” is a filter the team can run in seconds.

Search

Natural language search lets a team find any transaction across every account in seconds. A search such as “all wires over $50,000 to vendors in Q3” returns a data set that can be saved, reused, and turned into a report or a forecast input. This is the capability that generative AI now extends into full conversation.

Reconciliation and invoice matching

Matching invoices from an enterprise resource planning (ERP) system to bank transactions confirms that the company paid for what it received and collected what it billed. AI matches on amount, date, counterparty, and reference, learns from the exceptions a team resolves by hand, and leaves the analyst with only the exceptions to work.

Anomaly and fraud detection

Models trained on a company’s own transaction history learn what normal looks like and flag what does not. A duplicate payment, an unusual bank fee, a wire to a new beneficiary at an odd hour, or a balance swing with no matching activity shows up as an alert for a person to investigate.

Payments fraud makes this the most urgent application on the list. In the 2026 AFP Payments Fraud and Control Survey, 76% of U.S. organizations experienced attempted or actual payments fraud in 2025 and 74% were affected by business email compromise (BEC). Only 17% use AI to combat it. The distance between those numbers is where most treasury teams have their clearest near-term case for AI.

Anti-money laundering at banks

Banks run ML models across transactions, customer records, and external data to identify patterns consistent with money laundering, and natural language processing across communications to flag suspicious language. The models generate alerts and human analysts investigate. It is the same pattern corporates now use for their own anomaly detection, applied at a larger scale and under a regulatory mandate.

Five ways finance teams use AI for better decisions today

1

Categorize data and deliver it where it is needed

The constraint on most finance analysis is access to the data. Teams that rely on spreadsheets still log into each bank portal, download, reformat, and consolidate before a single question can be answered, and every manual step is a chance for an error or an unconscious adjustment toward the answer someone expects. Bank APIs remove the person from consolidation. AI then classifies the normalized data and routes it to the people and systems that need it. Lemonade runs 39 accounts across five banks and 15 entities on Trovata, and a six-month transaction history report that took hours to assemble now takes under a minute.

2

Scenario planning and risk assessment

Once a forecast baseline exists, scenarios are a matter of changing inputs. A team can model a 15% drop in collections, a delayed customer payment, or a new debt covenant and see the effect on cash by week and by entity. AI makes the baseline fast to produce and the scenarios fast to re-run, which turns risk assessment from a quarterly exercise into a question the team can answer in a meeting.

3

Contingency plans for economic shifts

Contingency planning starts with the threats that can move cash, such as operational bottlenecks in logistics or revenue recognition, stale or inaccurate bank data that produces a confidently wrong forecast, and market changes that arrive without notice. This year has supplied several of the last kind. The Federal Reserve raised the federal funds target range to 3.75 to 4.00% on September 16 after cutting in late 2025, and Section 301 duties of 10% or 12.5% took effect on imports from 60 economies on July 24. A treasury team with a live forecast can model higher borrowing costs and higher landed costs the same week they take effect, instead of waiting for the next spreadsheet cycle.

4

Forecast accuracy and variance analysis

Predictive models learn the patterns in the data rather than requiring the user to specify them in advance, which is why they tend to outperform static statistical models and why they keep improving as more data arrives. Variance analysis closes the loop. Comparing forecast to actual by category each period shows where the model is reliable and where the assumptions need review. Eversana brought its forecast variance down to single-digit percentages and retired its Excel forecast in under four months, and the variance analysis that took one to five hours a month is now instant.

5

Confident decision-making

Models that review and cleanse their own inputs over time react to anomalies and get more reliable with each cycle. The treasurer’s role shifts from building the forecast to directing it. The person supplies the context, the model does the arithmetic across far more data than a person can process, and the two together produce a number leadership can act on with a stated accuracy rate behind it.

What to get right before AI

The data foundation

AI amplifies whatever it is fed. In Datarails’ 2026 CFO Sentiments Survey, 65% of finance leaders said their most common AI frustration is outputs that give confident answers based on the wrong data, and only 4% of organizations have achieved a single source of truth for finance and operational data. PwC’s 2025 Global Treasury Survey names poor data quality as the top forecasting challenge, cited by 76% of treasurers. The order of operations is fixed. Connect the banks, normalize the data, and categorize it. Then add AI. An assistant pointed at a stale export produces fluent, well-formatted errors.

Security and isolation

Bank data is among the most sensitive data a company holds. Before any AI tool touches it, ask where the model runs, whether prompts or data leave your environment, whether the vendor’s model trains on your data, and how access is scoped by user and role. A consumer chat tool with a pasted spreadsheet fails those tests, and a platform that runs its AI inside an isolated environment with no public internet access, under the same permissions as the rest of the application, is built to pass them.

Auditability

Every number an AI produces should be traceable to the transactions and balances behind it. A finance team has to be able to show an auditor, a board, or a regulator how a figure was reached. Favor tools that show their work, cite the underlying data, and keep a record of what was asked and what was returned.

Human review of agent output

An agent drafts and a person approves. Anything that touches a bank or a counterparty, such as a fee dispute, a payment, or a hedge, stays with the person, and so does anything that changes policy. Define the review checkpoints before the first run, run the agent alongside the manual process for the first few cycles, and tighten the prompt against the gaps. By the time the team trusts the output, the analyst’s job has moved from production to review.

Trovata AI 2.0

Trovata AI 2.0, released in March 2026, has three parts, all built on the normalized bank data that every Trovata product runs on.

Trovata AI Chat answers plain-language questions about cash. It is available from the first login, so a new user can ask about balances, transactions, and trends while learning the rest of the product.

Trovata AI Insights surfaces patterns and anomalies in the data without waiting for a question, such as an unusual fee, a balance outside its normal range, or a shift in a tagged category.

Trovata AI Agents run defined treasury work on a schedule or an event and deliver a draft for a person to review. Fee audits, variance attribution, runway projection, and daily position summaries are among the ten patterns documented in “The AI Agents Playbook for Treasury Teams,” which is the companion to this guide.

Trovata AI operates inside a fully airgapped environment within Trovata’s AWS architecture, with no public internet access. Prompts and data stay inside the platform. Trovata AI 2.0 is built on a dedicated MCP server, an industry-first implementation for corporate treasury. For teams that work in a general-purpose assistant, the Trovata MCP Connector gives read-only access to balances, transactions, tags, and reports from Claude, ChatGPT, Gemini, Copilot, and other assistants, so the answer to a cash question comes from the same data the treasury team sees in Trovata.

Questions a treasurer can ask

  • Show all cash outflows by cash flow type for the past quarter.

  • List cash inflows in USD above $10,000 for the past month, largest first.

  • What were our highest and lowest total cash positions across all banks and accounts in the last 12 months?

  • What is our burn rate for the past year?

  • Build a Report of Foreign Bank and Financial Accounts (FBAR) in a format ready to file.

Follow-ups work the same way. Ask for the table rounded to whole dollars, for the same data as a stacked bar chart, or for a year-over-year comparison of month-end cash balances.

Three steps to take advantage of AI

1. Treat it as a tool
AI is a capability to be scoped, tested, and governed like any other. Give each use an owner, a data source, and a review step. The teams getting results from AI in 2026 are the ones that stopped treating it as a research project and started treating it as part of the treasury operating model, with the same controls as a new bank connection or a new payment workflow.

2. Expect disruption and find your exposure
The work that repeats on a schedule and runs on structured data will be automated, inside your company or at a competitor. Inventory it. A useful prompt is to ask what a new analyst would have to learn in their first 90 days that then repeats every month. Fee analysis, variance commentary, daily position reporting, and reconciliation usually top the list. Those are the tasks to scope first, and the tasks where a treasury professional who understands the data adds the most value as reviewer and designer of the work.

3. Master the tools relevant to your job
Only 34% of treasury professionals consider themselves effective in AI knowledge. The way to join them is to use the tools on your own data. Learn how your platform’s assistant answers a cash question and where it draws its data from. Write one agent prompt for a task you know well and compare its output against your own for a few cycles. Read how MCP exposes your treasury data to the assistants your company has approved. A finance professional who can do those things is the person the CFO calls when the next AI decision comes up.

A customer example

Reiter Affiliated Companies moved its cash reporting from Excel to Trovata and now saves one day per week with 100% reporting accuracy. Rejina C., Global Treasury Manager, put it plainly: “Trovata is faster, more accurate than Excel, and has saved us a full day each week.”

Gibson reports 25 hours saved on cash reporting each week. In both cases the time came back once the bank data was connected, normalized, and categorized, which is the foundation this guide recommends before any AI tool is switched on.

Further resources

See Trovata AI on your own bank data

The fastest way to understand what AI can do for a treasury team is to ask it a question about your own cash. Book a demo and bring your hardest cash question.

Subscribe to our newsletter

Book a demo

Get a product walkthrough of Trovata with one of our experts.