What Treasury Is Building With AI: NeuGroup 2026 H1 AI Workbench Report

# AI

# Cash and Working Capital

# Fx

# Technology

Insights and use cases from in-person peer group meetings about the state of AI in treasury—what NeuGroup members deployed, what stalled and what they learned.

INTRODUCTION — From AI Curiosity to AI Construction

From São Paulo and New York to Singapore, Amsterdam and San Diego, talk of AI swept through the meetings of more than 20 NeuGroup peer groups this spring. AI not only dominated discussions devoted specifically to treasury technology but shaped sessions focused first on liquidity, cash forecasting, risk management and talent. How NeuGroup members are responding to the pressure to leverage generative and agentic AI showcases and validates the power of exchanging knowledge with peers to achieve outcomes sought by all—the core of the NeuGroup Network experience.

AI discussions and presentations in H1 meetings demonstrated how finance teams are already using AI, quantified ROI when possible, stated plainly what broke and drew a clear line around what AI can and can't do. This report unpacks member use cases while exploring five AI lessons learned:

Members are building their own AI tools instead of buying them.

Data challenges determine whether an AI solution works.

The ROI decides whether it lasts.

The walls: governance, access and cost.

Human judgment: what members won't hand off.

This is a deep dive into the AI Workbench , NeuGroup's members-only collection of the AI and automation use cases members are building across the network. The Workbench is part of NeuGroup for Technology Transformation (NGTT), the cross-network group open to every member for comparing notes on technology across the Office of the CFO, where members can explore the use cases, add their own and join the conversation through NeuGroup Connect.

1 — AI got real, and members built much of it themselves

At the Tech Treasurers meeting, the treasurer of a large technology company put numbers on the table that may have sounded like the stuff of vaporware a year earlier. Her team built an investment agent designed with Concourse , trained on the company's treasury policy, duration targets and portfolio parameters, that recommends bond-rotation trades every morning. Acting on its recommendations has added an estimated $2.5 million in net interest income over two years. A second agent watches global cash pools and pings analysts when idle balances could be put to work, worth another $500,000 annually. The agent runs its analysis in 10 minutes; the same work took an analyst up to 90 minutes on a Bloomberg screen. A person still signs off, but the agent is generating the trade.

This team and others empowered to build AI tools—many of which are beyond testing and in production—have put themselves ahead of the curve and impressed peers. At the Cash Investments meeting, one member described an agent that a junior staffer built to handle the flood of KYC requests from banks—pulling the data and filling in the forms, with the goal of one day letting banks query the corporate's agent directly. Speaking in São Paulo, a member of NeuGroup for Latin America Treasury talked about an agent that runs every new debt agreement against the legal team's checklist. One of the FX risk managers gathered in San Diego explained a heavily automated hedging process assembled largely through " vibe coding " with large language models.

Those projects didn’t depend on treasury-specific AI software. Members built many of them with the general-purpose tools already in the finance stack: Microsoft Copilot, OpenAI's Codex, Anthropic's Claude. AI enables treasury to develop what it needs now, rather than waiting for a vendor or IT to provide the solutions.

2 — Deal with data now to avoid AI bottlenecks later

Many of the teams making solid AI progress point to the same prerequisite: getting your data house in order. That includes data cleanliness, organization and access. But improving data infrastructure can be a long, resource-intensive process, as one large technology company learned during a three-year effort to build the foundation its AI work required.

The company, a member of NeuGroup for Global Cash and Banking , started with dozens of systems and a cumbersome, manual routine: Treasury team members pulled files from each source, reworked them by hand and loaded them somewhere else—and the same data threw off different numbers depending on who reported it. The fix was establishing one home for the data: a Snowflake data lake, with systems connected to it over time, starting with Kyriba , since "day-to-day cash positioning, FX data, it's all in Kyriba," a member said. After the lake was in place, treasury built on top of it. A machine-learning model now produces the cash forecast directly from the data, replacing a quarterly process where many people gathered inputs by hand. AI agents check the work—reviewing the inputs and reconciling the forecast against actuals.

Members hitting that same AI data wall in meeting after meeting reached a shared conclusion: Teams that treat data governance as a first step gain a real edge as agentic AI becomes standard.

The plumbing is the project. At the Growth-Tech Treasurers meeting, members spoke candidly about needing to solve data issues first to fulfill their AI ambitions. The projects they described weren't models; they were foundations: finance data lakes, consolidating bank and ERP data, enriching treasury data in Snowflake, upgrading API connections so forecasts and reports have something clean to run on. Several said flatly that their AI efforts are on hold, waiting for improved data quality and plumbing.

The same connectivity question emerged among Life Sciences teams weighing whether AI can be wired directly into SAP, Bloomberg, Kyriba, Clearwater and internal data lakes to automate work that's still manual. Mega-Cap treasurers sounded the same theme, with one member noting that using AI in forecasting and automation "depends on foundational investments like clean data and robust data lakes.”

Members of NeuGroup for Latin America Treasury validated the importance of well-connected data. While the bulk of their cash flows through global banks like JPMorgan and Citi, a slice of payments—local taxes, judicial and government payments in Brazil and Argentina, low-volume operations not worth integrating—sits in regional bank portals. If a portal can't be automated into the ERP, that slice stays manual and invisible to any AI forecasting tool. The group's clear verdict is that AI still can't do predictive cash forecasting—what everyone wants—because the inputs aren't connected. The gap is starting to close as global banks partner with local institutions to absorb payments they couldn't previously process.

Scott Dunphy, managing director and AI lead at MetLife Investment Management , opened the DB & DC Plan Management meeting by describing how he built a proof of concept in which 21 AI agents worked in sequence to produce a full investment memo. The result, he said, was "pretty amazing." The harder problem was upstream: His group's real estate business has run for more than a century, and much of its institutional knowledge sits on shared drives that AI can't effectively search, so the team is undertaking a large effort to centralize data.

NeuGroup members can read and download the full report, NeuGroup 2026 H1 AI Workbench: What Treasury Is Building With AI, by clicking here .