6 things to know to make the best out of generative AI for Finance and Accounting

Don’t worry if you missed our AccountingGPT webinar in June because we took the time to break it down to its key insights and included some of questions viewers posted live. If you are a financial-, accounting- or a shared service center leader and are anxious to find the most seamless way to leverage the latest in AI and turn your operations into a truly no-touch workflow - keep reading! Here are the 6 things to consider for a successful APAI work frame.

Pick the right AI model: Large Language Models vs Specialized Transformers

When we talk about generative AI, we need to make the distinction between General Purpose Large Language Models (LLMs) and Specialized Language Models (Specialized Transformers) – even more so when it comes to employing them to serve specific business use cases. 

LLMs are built on publicly available datasets that include more than 200bn parameters and training them requires extensive and expensive fine tuning by people. They can be applied broadly – generate images, text, coding. Examples of LLM products are ChatGPT (OpenAI), Bard (Google), Luminous (Aleph Alpha).

Specialized Transformers are small to mid-size models (50 to 250 mn parameters), can be fine-tuned with historic data from an existing data base, like an ERP, without prolonged human intervention (initial training time between 4 and 24 hours). An example of a Specialized Transformer is our own AccountingGPT product.

At Hypatos we’ve tested the efficiency of LLMs against Specialized Transformers by to see if and how ChatGPT-4 can capture documents and process them end-to-end. The result was comparable to what a traditional OCR solution can achieve for data capture – in the neighbourhood of 85%, but 0% for further processing without additional training. The former is not low by any means, but it’s nowhere near enterprise standards.

Our Specialized Transformers on the other hand, besides allowing fast and comparably inexpensive retraining, deliver results in the 98 percentile or higher for data capture, and 95% or more for GL accounting.

You’ve asked:
Do I need to hire developers to adopt specialized transformer technology within my financial department?
We answer:
No, you don’t. You will however need to make sure you have a business user and a technical user who maintain and work with the specialized transformer. Those people are tech savvy, but they are not necessarily engineers or developers.

Work with the right data: Public vs Private data

A fundamental rule of performance is that any generative AI solution will only be as good as the quality of the data it’s fed. There are two types of data that go into AI models – public and private:

  • Public data – used to train general purpose LLMs includes publicly accessible information from the internet and more than 50 million invoices, sales orders and so on from organisations who share training data. Public training data teaches a model to understand language.
  • Private data – used on top of public data within Specialized Transformers to finetune performance. Private data includes the organization’s private accounting information from ERP and workflow systems combined with invoices, sales orders and so on from document management systems and document archives.

Enterprises looking to increase performance in their AP / AR departments are therefore looking to employ public and private data together to achieve concrete goals with a much narrower focus. In other words, they need a Specialized Transformer that understands and uses their private data to achieve end-to-end processing with high accuracy and confidence.

You’ve asked:
Will my data be used to train a model that competitors or other external companies can use?
We answer:
When it comes to your private data, the answer is a confident “no” for a very simple reason – no company does their accounting like you do. Your accounting coding, organization of cost and revenue centres, approvers and so on, are only your own, built around the logic that works for your enterprise. But since your specific transformer operates partly on public data – this has already been shared even before it became part of your solution. 

Clarify your focus: Document understanding, Processing automation or both

Earlier we mentioned that we’ve managed to achieve OCR levels of automation for data capture with ChatGPT-4. It’s important to highlight that this covers only the first part of a document’s processing, which includes document ingestion, document classification, information extraction and document validation.

If the goal is to deploy an end-to-end no-touch processing workflow, then enterprises must look beyond the document capture and understanding.

OCR technology can only go so far before it needs to be enhanced with a rule-based workflow solution in order to move you closer to an end-to-end processing. But even then, enterprises face blockers such as – dealing with exceptions, onboarding new vendor and client data, together with common challenges like duplicate payments, errors in document matching, manual coding, and other roadblocks that require a new rule implemented for each non-standard situation.

Here is how Specialized Transformers such as AccountingGPT truly shine because with them, you solve those issues differently.

Not only that they outperform OCR and workflow solutions in understanding documents, regardless of language or formatting, they can match the data from the incoming document to data that exists within the ERP and, without human intervention, apply the correct accounting codes and forward to the right approver, cost centre and so on in an instant. This is where their superpower lies – automatic master data matching, PO-matching, coding for whichever accounting attribute you need, and processing. The only pre-requisite is that the master data can be found within your ERP in sufficient quality.

You’ve asked:
Can specialized transformers process more than just invoices and POs, but also things like purchasing contracts?
We answer:
Yes. Any transactional document can be processed with Specialized Transformers. Our own solution, for example – AccountingGPT – can handle 20+ types of documents. 

Tell the difference: Accuracy vs Confidence

Machine-learning models are not perfect. Expecting 100% automation with a very high confidence level is not how this works from the get-go.

Within a generative AI model, the concept of “Accuracy” refers to the measure of how correct or accurate the generated output is in comparison to a reference or ground truth, i.e., the accessible ERP data. “Confidence” on the other hand, refers to the level of certainty or belief that the model has in its generated output. It indicates the model's internal measure of how reliable or confident it is about the generated result.

The way enterprises work with this peculiarity is to decide what confidence score to a specific answer is not high enough to be trusted, and to require human feedback. And because generative AI learns from what you've already done and from what you're doing, it gets better and better over time until it does not require a human in the loop. 

After your Specialized Transformer is implemented, it gets continuously better in using the information it has over time.

You’ve asked:
How do we avoid ChatGPT confidence about wrong statements with specialized transformers?
We answer:
The way we do it is we implement our specialized transformer models after initial training, and then allow for a post-implementation fine-tuning period that fills the gaps in the way it leverages enterprise data, e.g.: your private data, with feedback from operators. If a model’s confidence about a specific output is too low, it will not even suggest it within your tools, but it will learn what the correct decision is, once an operator provides feedback. Rinse and repeat for every such instance and over time you get a no-touch processing (or a very very low-touch one).

Decide for a standalone or an integrated AI solution

Generative AI tools are not here to replace or overhaul enterprise systems, nor force teams to manage tedious multi-month rollouts. They are here to make life within the finance and accounting department more transparent and organised, have positive impact on the enterprise’s bottom line, boost the efficiency of even legacy software, where needed and do that quickly.

You’ve asked:
Which ERP systems can be enhanced with generative AI technology?
We answer:
Connectors can be built for anything so the short answers is that sky is the limit. In terms of platforms where our AccountingGPT has native standard integrations - these currently are SAP S4HANA (private and public); Workday Financial Management, and a native integration with Coupa. We also work in the Oracle ecosystem quite a bit, and let’s not forget our Microsoft Dynamics connectors that we’ve deployed for a couple of clients.

Last, but not least - be clear on your objectives

Do you want to optimise how your team spends their time? Or maybe your number one priority is to eliminate processing errors? Thanks to our own generative AI solution – AccountingGPT – we can confidently say that enterprise success in the accounting department boils down to a few core indicators that impact the enterprise bottom line

  • 10x productivity – from 50 to 500 invoices processed per day per FTE
  • Up to 100x faster – from 12 days from capture to posting of an invoice, to mere hours
  • Up to 75% less errors – the error rate drops from 3-4% to below 1%


The success of your generative AI journey is a function of how well the available financial data is used, what the quality of this data is and the rate of its improvement, and how accessible it is within the enterprise systems. Choosing the right generative AI model is the second step and with it you make sure to get the right data, at the right moment, to the right person, immediately.

Reach out to test how AccountingGPT works with your specific business information.

Not ready for a demo yet? Watch our AccountingGPT webinar to learn more about generative AI in the finance department.