AI in Treasury: Promise, Preparation and Precaution

Published: February 20, 2025

AI’s unstoppable move into treasury is happening now, but what are the rules of engagement? TMI seeks the advice of an expert witness.

The rapid evolution of AI is set to fundamentally reshape treasury technology. While TMSs have gradually shifted from on-premise solutions to SaaS models, we’re now facing a more dramatic transformation.

At its core, a TMS performs four key functions: it creates, reads, updates, and deletes records in structured databases to manage cash, investments, debt, and risk. AI models have now demonstrated the capability to perform all these functions through natural language processing (NLP) and generation, potentially eliminating the need for traditional database structures entirely.

AI is already making inroads into treasury operations in practical ways. Treasury teams are using AI to extract insights from policy documents, identify FX exposures, analyse bank fees, and support investment decisions with market analysis. The future will ultimately see AI included in some part of all processes.

The treasury technology landscape remains a mixture of on-premise and SaaS solutions, though the trend strongly favours cloud-based systems. The pace of innovation in AI is remarkable, with significant improvements in capability coupled with reductions in cost and environmental impact. Major cloud providers Microsoft, Amazon, and Google are integrating AI throughout their platforms, seeing it as a crucial opportunity to strengthen their enterprise ecosystems. However, the roll-out has been deliberately measured, particularly given the sensitive nature of financial data, while few TMSs have yet to integrate any generative AI (GenAI) capabilities.

Bank of America’s implementation of Erica, a chatbot handling transaction searches and balance inquiries, demonstrates how FIs can successfully implement AI with appropriate safeguards. This careful approach to innovation while maintaining security will be crucial as AI capabilities expand.

The future could look radically different: Microsoft’s CEO Satya Nadella, outlined on the BG2 podcast in December 2024 that AI agents could replace traditional SaaS software solutions entirely. The alternative vision would see solutions providers creating task-specific agents, with the business logic currently performed by TMSs instead being carried outby agents that can interact with multiple data sources.

For treasury departments, this would represent both an unprecedented opportunity to achieve much greater levels of automation and but also a significant challenge in maintaining security and control. If this prediction is accurate , the transition is likely to be gradual but profound – moving from today’s structured systems to more flexible, intuitive AI partners that can handle everything from daily cash management to complex hedge accounting.

The case for human-AI collaboration

Recent research has revealed both the promise and perils of AI systems. New performance and intelligence benchmarks are constantly being achieved, such that in most disciplines there is an AI model that comfortably ranks in the top 1% for understanding and performance.

However, TIME magazine covered a recent study by Anthropic and the non-profit Redwood Research that exposed the fact that advanced AI models can strategically mislead their creators, checking for oversight before acting and simulating compliance while pursuing their own objectives. For treasury operations, where accuracy and trust are paramount, this reinforces the necessity of human oversight. While AI can process vast amounts of data and suggest actions, human judgment remains essential for validating recommendations and ensuring decisions align with organisational strategy and risk appetite.

Treasury departments are typically lean operations where team members juggle multiple critical responsibilities. The challenge isn’t overstaffing, and therefore productivity gains from technology shouldn’t be about cutting heads. Instead, this is an opportunity to add more value, as experienced professionals spend considerable time on necessary but routine tasks such as data gathering, report creation, and transaction processing, leaving limited capacity for strategic work.

AI’s role is to create capacity for these value-adding activities. By handling routine tasks and providing initial analysis, AI can release treasury professionals to focus on strategic decision-making, market analysis, and stakeholder engagement. For instance, by using AI treasurers can leverage technology to examine a much wider range of investment options than through a manual process and then use the extra time gained to provide the crucial business context and risk assessment to produce an optimal outcome.

The path forward

The transformation of treasury technology will not be overnight, but adoption is likely to accelerate significantly over the next three to four years, such that within five years, treasury technology is likely to look notably different. TMS providers are already experimenting with basic AI implementations, focusing on data retrieval and analysis. More sophisticated applications are expected within 18 months, with those providers that are early adopters potentially gaining significant advantages in efficiency and capability.

James Kelly will be presenting a pre-conference training session covering AI in Treasury at this year's EACT Summit, taking place from 10-11 April near Brussels.

Please contact Livi Miller at livi@tmi.co to register your interest.

Starting small: practical firsts

  • Treasury teams can begin their AI journey by drawing up a manageable roadmap
  • Identify one regular spreadsheet-based report that pulls data from multiple sources
  • Document the current process and data sources
  • Consider automating the data-gathering process using simple tools
  • Standardise data formatting across related spreadsheets
  • Create a simple data dictionary for key treasury metrics

These initial steps build the foundation for future AI implementation without requiring major infrastructure changes. As teams become comfortable with basic automation and standardisation, they can gradually move towards more sophisticated data management approaches.

The other key area to prioritise is in security. While AI assists us to become more productive, it also enables those who employ it to do harm. AI also enables increasingly sophisticated phishing attempts and realistic fakes, and so treasury teams will need robust vulnerability management systems, an area we can expect to see a growing focus for advisers and fintechs.

Shaping the relationship

The potential to improve treasury performance and working lives through AI is enormous, but treasury heads shouldn’t feel pressured into changing overnight. Successful adoption will require careful balancing of innovation and control. Taking a measured approach, starting with small, manageable steps will enable departments to learn what works for them and build confidence, while maintaining strong security protocols.

By beginning this journey now, treasurers can proactively shape their own relationship with AI technology, rather than wait for external consultants to impose a ‘one-size-fits-all best practice template’, ensuring it evolves to truly serve their specific needs.

Helpful checklists:

Essential Controls for AI Implementation

Ensuring that AI can be understood and is producing reliable results is critical, so taking the following steps helps ensure that the output can be trusted:

  1. Data Validation
  • Ensure that all decisions are explained with workings. If using GenAI, this can be achieved through prompting, while if using ML, the inputs and outputs can be validated to show how a decision is reached
  • Maintain audit trails of AI-assisted decisions in the same way as you would for the process today
  • Regular testing of AI system accuracy. Spot checks and regular reviews help ensure reliability
  • If anomalies are identified, such as AI trying to solve problems by using new techniques, these should be quickly followed up on to identify whether these are in the interests of the company.
  1. Security Protocols
  • Multilayer authentication for AI system access: this is particularly important for the logic and data being used
  • Ensure that GenAI is used with care: the risk of hallucinations (when a large language model, or LLM, perceives patterns or objects that are non-existent) or unintended consequences means that rules-based automation should be preferred to AI, so that AI is used only for challenges to which automation is not well suited (it is difficult to create rules when producing commentary or forms for individual banks, for example).
  • Regular security assessments: just as there are regular audits and IT penetration tests, expect that Big Tech - Alphabet, Amazon, Apple, Meta, and Microsoft - and others, will be busy performing AI security assessments
  • Clear segregation of duties: for small teams this will be harder, but just as changing payment details requires higher levels of security, the same applies for changes to AI logic once deployed (especially if it’s working autonomously)
  • Incident response procedures: preparing for the risk of downtime, errors or the system being compromised is particularly important
  1. Governance Structure
  • Define clear ownership of AI systems: –if something goes wrong, it is important that there’s clear accountability to enable swift resolution
  • Establish review and approval workflows. Best practice would be to ensure at least one human review of any AI-enabled decisions.
  • Regular performance monitoring to facilitate early warning of any issues is essential
  • Compliance documentation requirements: control and validation will be important for treasury, given the sums of money typically involved

Evaluating TMS Provider AI Readiness

When discussing AI capabilities with TMS providers, consider these key areas:

Strategic Direction:

  • Timeline for AI feature integration
  • Team experience both in AI and in treasury
  • Development approach (proprietary versus partnerships)
  • Priority areas for AI enhancement
  • Pricing models for AI capabilities

Technical Implementation:

  • Integration with existing workflows
  • Security infrastructure
  • Data segregation approach. No data should be included on a shared model or database
  • Offline functionality: enables a system to run even when the internet is not available. This can be useful given recent ChatGPT outages

Security Considerations:

  • Data protection measures: no data should be used for training the model unless it is a private model and the company should be able to request that all data be deleted
  • Audit trail capabilities: any actions should be documented
  • Access control systems: strict limits should be in place
  • Compliance frameworks: ISO 27001 and SOC 1 and SOC 2 should be followed
  • Testing protocols: regular testing should take place
Article Last Updated: February 20, 2025