24 February 2025: Four Questions to Ask About AI Automation 

Shift your thinking from what AI can do, to what it should do. In a recent FT article (16 Feb 2025), business school professors posed four general questions to ask when contemplating automation. In adapting this useful article for finance professionals, I believe that asking these questions can help when creating your own automation and digitisation strategy. 

For each task, ask:

1) How complex is the task? First, analyse all of your tasks and processes, including outputs and inputs. Map them all out individually first, then connect them. Link to your policies. Then ask for each task: Does this task need to be done at all? If affirmative, then ask: Can it be simplified somehow? Finally, once simplified, ask: how complex is the process now? If it remains complex, then automation will be difficult because humans are still better at dealing with complex tasks (for now). However, if you have simplified the process, then AI may be able to automate in part or in full.

2) How frequent is the task? The higher the frequency, the more benefit to be gained from automating and the larger the efficiency savings potential.

3) How interconnected are the tasks? In many finance tasks, processes are interconnected and are often completed by different departments, people, and machines. Fragmentation costs arise from inefficiencies and errors when the baton of a task is handed over. A classic example is the cash forecasting process, in which there is a heavy reliance on disparate departments, workers, and machines. Time delays, errors, and inconsistencies can waste a lot of time and cause significant frustration, ultimately leading to poor strategic decisions.

It is recommended that the potentially high costs created by process fragmentation discourage finance professionals from dividing tasks between humans and GenAI, even if technically feasible. Automating potentially many ERP working capital ledgers into a consolidated group cash forecast, reviewed only by group treasury without local finance review, might appear cost-effective, but crucial information could be lost during the transition from GenAI to central treasury. Put simply, local and central finance teams should still be fully responsible for cash forecasting, with GenAI used only to speed up the process, reduce errors, and increase efficiency.

4) What is the cost of failure?  While GenAI can progressively lower fragmentation costs vs. traditional automation (e.g. RPA), it can also be less precise than some past forms of automation, leading to incorrect decisions and actions. Will you be comfortable if a full or part AI-automated task failed?  What would be the cost?

Summary: High frequency, repetitive, and simple finance processes should be considered for AI automation. However, fragmentation costs and the costs of failure will also need to be scrutinised to ensure balance between overt cost savings from automation and any hidden costs. 

6 February 2025: Your Data Strategy

I recently attended a data strategy webinar and thought it might be useful to summarise and add my thoughts and practical experience. 

There is much excitement about the potential of AI and the productivity gains that may be coming to a treasury department near you. However, as an experienced techie and TMS implementor, including intimate experience with ISO20022, data quality and accuracy – much of it stored in the ERP and TMS as ‘structured data’ - is paramount to realising these productivity gains. If not, you may find a productivity loss and therefore a project failure on your hands.

From practical experience, ensuring your data is complete, clean, and correct is relatively easy if you are creating a new treasury function from scratch, if you are fully aware of the data requirements, format, and accuracy, as per ISO20022. 

Achieving the necessary data quality and completeness becomes significantly more difficult when the treasury function has been around the block a few times and has old, incomplete, or/and inaccurate data. Legacy systems, information stored in long forgotten spreadsheets and banking platform can cause many unforeseen and frustrating issues when using this same low-quality data for your technology upgrade. 

A typical example, and one I have personally led on, is the automation of US Private Placement payments under ISO20022, with sFTP from TMS to ERP to banking platform. I have written a presentation on ISO20022 and its implementation on this website, which you can read (it’s all for free!). Just for the purposes of this article, note that to comply with ISO20022 and therefore to ensure automation of payments clear the bank with no fuss, the data must be present, accurate, and correct. 

While data maintenance and clean-up can be boring, it is vital if you wish to utilise the full potential of machine learning and AI (machine learning is a sub-category of AI). Missing data, data in the wrong format, incorrect data categorisation, bad spelling etc on potentially millions of lines can make this task seem daunting. But, if done right, it can be a one-off exercise to clean up the data, and then to create policies and processes to ensure the data is managed correctly.

You need a data strategy for treasury!

A data strategy is a long-term plan outlining how you will collect, manage, and use relevant data for your operational and strategic treasury goals. 

NOTE! Comprehensive data strategies are commonly conducted and applied organisation wide and regulated by IT. For our purposes, we are referring to treasury-related data required for automation of treasury-related operations. 

Here are a few tips and tricks to reduce the probability of project failure by ensuring your data is high quality, present, and correct:

1. Set out what you wish to get out of your data, your data goals and the desired outcomes. Be ambitious. Investigate best practice, e.g. what are the requirements of ISO 20022? 
   Practical tip: Align with the overall business strategy; consult directly with the relevant teams and departments; you will not be able to do this alone!
2. Investigate where you are now, what data you have, where it is stored, how accurate is it etc. Perform a data audit if needed. 
   Practical tip: This could take a while: your plan should include enough time to complete this part, plus additional labour may be required.
3. The gap between 1. and 2. feeds into your strategy. 
   Practical tip: Avoid collecting data that you do not need. Steps 1 and 2 should be performed to a high level of detail that you can state what data you need, why you need it, the source, how you will gather it, and where it will be stored.
4. Identify all owners of data and their data responsibilities, working with IT to do so.
5. If your organisation is complex, then consider setting up a data steering committee with appropriate executives as members who can wield the necessary influence.
6. Data security protocols are of high importance (e.g. GDPR compliance; archived data); ensure you consult with IT and other relevant departments in your organisation.
7. Data sources. Attempt to limit the number of sources and, where possible, automate the process of data transfer. 
   Practical tip: Keep a detailed record of all data sources in one place, how it is collected and transferred, definitions of any variables (e.g. tickers; RICs). 
   Practical tip: Do regular spot checks for errors and comparisons to ensure the correct data is being pulled through
8. Data processes and controls are paramount to ensure all the time you have spent to clean your data and get it ‘AI-ready’ does not go to waste by bad maintenance practices. 
   Practical tip: Map out the processes in full, include the controls, identify accountability points, and carry out formal testing to ensure the mapped processes work as intended.

You could also follow the PPCAT approach for your data strategy, which is a general problem-solving model used in statistics that I have adapted to construct a simple data strategy in powerpoint:  PPCAT Diagram

25 November 2024: GenAI in Corporate Treasury

I have recently completed the CFTE AI in Corporate Treasury course, link here: Generative AI in Corporate Treasury Course - CFTE.

The course consists of three chapters, starting from a short history of AI to how GenAI can be applied to treasury and indeed how it is currently doing so. Insightful interviews with treasurers at the forefront in this space are littered around the course, as are case studies, which include cash forecasting and investment analysis.

It was a good, organised and insightful course; however, I was already familiar with much of it. It appears that this new age of AI is not so new. For example, it discusses bank statement transaction categorisation and how it can be used to automate bank recs. Many TMSs have had this capability for a long time, and I have set these rules or algorithms up myself in IT2 and Kyriba. 

Data is the biggest challenge presented to treasurers: quality, completeness, integrity, location and format are some of the considerations required. AI needs lots of data to be of any use over and above excel- note that excel can also do regression but usually limited to a certain amount of data.