Treasury AI in 2026 guide: Build in house vs Buy AI layer on top
What is treasury AI?
Treasury AI is the use of machine learning, foundation models, and agentic workflows to automate cash forecasting, transaction categorisation, variance analysis, and liquidity optimisation. Unlike a traditional Treasury Management System (TMS), which is a system of record, a treasury AI platform sits on top of existing data sources, interprets that data, explains what changed, and recommends what to do next.
Most enterprise treasury teams in 2026 are sitting on a mature TMS and ERP that were designed to move money and book entries, not to think. That gap is what treasury AI exists to fill.
The three paths to treasury AI: a quick comparison
| Option | Time to value | Total cost (Yr 1) | Accuracy & audit trail | Best for |
|---|---|---|---|---|
| Buy a treasury-native AI layer (e.g. Palm) | 2–4 weeks | Predictable subscription | High. Explainable AI, full audit trail | Most treasury teams |
| Build in-house | 6–12 months for v1 | High and uncertain (engineers, infra, ongoing ops and maintenance) | Variable. You build the audit trail yourself | Teams with dedicated engineering resources |
| Extend a generic co-pilot (e.g. Microsoft 365 Copilot) | Immediate, but limited | Bundled licence cost | Low. No treasury context, no defensible forecast | Drafting, summarisation, board-pack prep |
Should you build treasury AI in-house?
Short answer: probably not. Building is the right choice only if you have a strong dedicated engineering team to help you out, an unusual workflow no platform can cover, and 12+ months of CFO patience. Fewer than 1 in 10 treasury teams actually meet that bar.
The reason is that the hard part of treasury AI isn't the model. It's everything underneath:
- Data fragmentation. Bank data often lives in various files (PDF, Excel, MT940, BAI2, CAMT), that are generated with different cadences and time zones. ERP exports drift as finance reorganises cost centres. AP runs, AR ageing, payroll and tax sit in separate systems owned by separate people, often in multiple ERPs after years of M&A.
- Foundational engineering. Before you train a single model you need entity resolution, currency handling, cash-pooling logic, ZBA and target-balance sweeps, intercompany flow recognition, and a self-learning categorisation engine.
- Ongoing operations. Models need continuous retraining. Pipelines need SRE attention. Every variance the CFO questions needs an audit trail you've built and maintained.
- Key-person risk. The first time an engineer leaves, the institutional knowledge of how the pipeline was wired walks out with them.
Industry benchmarks put the build at 6–12 months of dedicated engineering to reach parity with a packaged platform, and that's only version one. Most treasury teams systematically underestimate this ongoing operational cost and overestimate the strategic differentiation of a homegrown forecast.
There is also a budget reality. CFOs increasingly want a return inside a single fiscal year. A 12-month build with a 12-month payback is a 24-month bet, and most treasury transformation budgets do not survive that timeline.
Can Microsoft Copilot replace a treasury system?
Short answer: no. A generic AI co-pilot like Microsoft 365 Copilot or ChatGPT Enterprise is genuinely useful for summarising a board pack, drafting variance commentary, turning a question into a chart, or replying to routine emails. That's a productivity layer, not a treasury system.
Pointed at treasury, a generic co-pilot has no opinion on how to reconcile a ZBA sweep, no view of your 13-week forecast, no embedding of explainability throughout its choices, and no way to defend a forecast number to an auditor. It cannot tell whether the financial data it has been handed is accurate, fresh, or complete. That's pure garbage-in, garbage-out for finance.
Use co-pilots for what they're good at: drafting and summarisation. Don't expect them to deliver explainable and accurate cash forecasts.
What is Palm's AI layer on top of the TMS?
Palm is treasury AI infrastructure that sits above your existing TMS and ERP. It doesn't replace anything. It makes what you already run intelligent. Specifically, Palm ingests any data (bank feeds, APIs, files), normalises them into a single cash model, then runs an intelligence layer for forecasting, categorisation, variance analysis and liquidity signals.
This architecture matters because it directly addresses the failure modes of building in-house. It also protects the investment you've already made. Years of onboarding to your TMS, the consolidation work to get bank feeds flowing through SWIFT and middleware, the categorisation rules built in your ERP, the cash-pooling structures negotiated with your banks: none of that is wasted. Palm reads from the systems you already run, normalises the data you've already consolidated, and layers intelligence on top. You aren't starting from scratch and you aren't replacing the foundations. You are finally making sense of the data those foundations have been quietly producing for years.
1. You skip the foundational data work
Palm ingests MT940, BAI2, ISO 20022 CAMT, Excel, CSV and PDF files, and normalises them out of the box. Entity resolution, currency handling, cash-pooling logic, intercompany recognition and AI-powered categorisation are already done. The six months of foundational engineering required by an in-house build is replaced by a data feed.
2. Forecasts are model-selected empirically, not assumed
Palm evaluates 18+ models (classical statistical, machine learning, and foundation models) against every time series and selects the winner for that specific pattern. What works for payroll doesn't work for customer receipts, and Palm doesn't pretend it does. Models retrain daily on your data and incorporate covariates the spreadsheet can't see: holiday calendars, payroll cycles, planned tax payments, AP/AR feeds.
3. Every prediction is explainable and audit-ready
Treasurers cannot defend a number by saying "the AI said so." Palm produces plain-language variance explanations, transparent model selection (you can see which model won and why), and bottom-up traceability from forecast to underlying transaction. This is what makes the output defensible to auditors, the board, and the CFO.
4. Read-only, zero-write architecture clears security review fast
The Palm platform does not initiate transactions, move money, or modify source systems now. It ingests, it learns, it recommends. CISOs approve this faster than any TMS project. Average go-live is 18 days, versus 6–24 months for an in-house build or full TMS replacement.
5. Institutional knowledge is captured permanently
Every correction your team makes trains the AI. Your best treasurer's judgement on categorising an intercompany flow or recognising a seasonality pattern becomes organisational infrastructure. That knowledge no longer retires, takes holiday, or leaves for a competitor, directly mitigating the key-person risk that haunts homegrown systems.
6. Compounding advantage from day one
Because the model retrains continuously on your data, every day on Palm makes the forecast smarter. A team that started building a year ago is, at best, getting to v1 today. A team that deployed Palm a year ago has 365 days of compounding intelligence, a gap that widens over time.
Treasury AI: build vs buy decision framework
Use these four questions to decide:
- Do you have 6+ dedicated engineers and 12+ months? If no, don't build.
- Can your workflow be served by a packaged platform? If yes, the differentiation argument for building disappears.
- Will the CFO accept a 24-month payback? Most won't.
- Can you defend every forecast number to an auditor? Whatever you choose, the answer has to be yes, and that requires explainable AI, not a black box.