APA RPA Use Cases that do not require Co-pilot / chatobt

Hey Jason

Automation Co-Pilot is primarily a great way of bringing humans into an agentic process - human-in-the-loop as we call it - but there’s plenty of ways of accomplishing this. If you need human approval or a human review you could use a spreadsheet to track approvals. This was something I implemented at a customer many years back. An RPA bot would review a spreadsheet every n minutes and if an accountant marked a line as approved then the Bot would pick that data set up and post it to an Oracle General Ledger. Nowadays, you could accomplish something similar through MS Teams and the Graph APIs using API tasks.

In term of use cases, I’m a Chartered Accountant by background and have some strong views on APA in Finance and Audit...namely there’s not much it can’t do 😂. 

If I break Finance into Management and Financial accounting then I would say the big areas are:

• Month end - A lot of this has likely been automated with standard RPA. However, APA and specifically content generation and summarization allows for better and faster variance analysis and the production of reports. APA could also be used across business units giving a more holistic view of company performance rather than the current divisional/silo’d view. Impact of this is month end can close quicker, posting errors can be identified faster and variances against plan can be identified and understood better leading to corrective actions that much sooner.
• Reconciliations - One of my first ever jobs was reconciling £’millions of bank transactions across thousands of lines and this is something which should be done regularly to spot fraud or errors. An Agent combined with some tools that can match datasets and query transactions in a GL would be highly efficient at doing this and could run on schedule. I do believe this is something that could be built once and rolled out many times over. Impact of this would be huge amounts of time saved, reduction in errors.
• Financial statement production - Similar to month end, variances against plan could be queried and understood much faster. These reports will all be automatically produced but the understanding of why things have moved can be difficult because it’s at such a high level. For orgs that are heavily cash focussed right now, daily cash flow statements are not uncommon and these usually fall on a single person or team. Impact of this would be speed of production and better accuracy leading to improved decision making.

Automated audit of transactions is a use case where you may not need co-pilot. But you would need a way to highlight potential fraud / error transactions to users. Email could be used instead.
 

• Compare new invoices to past transactions for the same vendor. Do the new charges align (description, value and qty) with what you have purchased from them in the past?
• Compare new invoices to guidelines set out in a services agreement contract. Do the new charges align with the contract? Have they provided all required information on the invoice?

Another use case is assigning GL codes and recharge codes to transactions using past information or contractual agreements to guide the AI. Some companies have expressed how hard it s for field service teams to be experts in using project charge codes, when they are really only experts in whatever field service area they work in.

Hey ​@JasonD106468,

One really nice application of AI in general is Metadata creation.  This is something that’d fit really nicely in IT and Audit.  

Imagine in IT, trying to track down information on a specific incident or error, and it takes forever because you are relying on plain-text searches in your ITSM tool.  Or imagine an audit person is looking for evidence of a specific policy adherence, but it takes (again) forever to do so because of the organization of data.

In this situation, the current state is lacking and requires some help.  Adding AI skills to classify, extract data, summarize content, or do anything else to each record, while appending that information as metadata to the records, allows for easier search and organization of data later.

The reason this is so helpful and doesn’t really required HITL is because of the current vs future state nature of things.  Even you are only able to generate 80-90% of the useful metadata, you’ve potentially made a HUGE improvement.    If you happen to get some small percentage of tags incorrect, there’s no business impact because the people who do the real process of finding this data or using it can filter out the ‘wrong’ information just as easily as they already do in the current state.

Remember the concept of “Point of No Return” in my Managing AI Risk.  In this scenario, there is no single point where impact is unpreventable.  There’s almost nothing AI can do in this scenario that can break the existing process.  If EVERY tag of metadata was wrong and humans had to rely on the current methods instead, status quo persists.  


Hope this gives you some ideas!