LLMs for Automation in Accounting and ERP

1. Abstract

In this article, we will understand the reason behind the risk of job loss in the Lower Level of Management of the Accounting division of any organization. We will learn how the holistic implementation of Artificial Intelligence can empower an organization in catalysing the growth amidst the increasing competition through practical experimentation with real-life use cases. We will learn how the implementation of LLMs will help in increasing the efficiency and accuracy among the teams. Also, using the experiments we will try to understand how capable this technology is and where it lacks. We will also get to look at how the global enterprises are implementing this technology and how it is benefiting them.

2. Introduction

LLMs for Automation in Accounting and ERP, we are familiar with the word Accounting, ERP & Automation, but what does LLM mean, well, LLM is an abbreviation for Large Language Model.

Basically, it is a model, for easy reference, let's say Google's Gemini or OpenAI's Chat GPT, these both are Large Language Models. These models are trained on vast amount of text data to understand, generate and interact using human language

We are now using these LLMs in our daily life, knowingly or unknowingly, these technologies have found their way not only into our Personal Computers and smartphones but with the ever-increasing use of these, these technologies are now being implemented on our Television Sets and Refrigerators as well.

But, how can it help in an Accounting and ERP Context? Well in that case, let us take a step back and have a look at the "Future of Jobs Report 2025" issued by the World Economic Forum which has laid down a list of Largest Declining Jobs by 2030.

Extract of Top Largest Declining Jobs by 2030, which revolve around Accounting and ERP Environment are as follows, Out of 15 Jobs enlisted in the Report, almost 50% (7 Jobs) are somehow revolving around ERP or Accounting.

What all these jobs have in common is, Manual or Semi-Automated Data Entry of Highly Structured Information in to the ERP System.

Mainly these jobs involve identification of characteristics such as, "Who?', "What?", "How Much?" or any mix of these questions.

Let's have a look at some of these jobs and identify these traits,An LLM is capable to read / scan the data / document, extract these relevant information and report it in a formal structure like a Spreadsheet, Charts, PowerPoint Presentation or even an Excel File.

We are very much familiar with the insights being shared by the CEOs of Swiggy, Zomato or Blinkit on huge events such as any Festival or a Cricket Match between India & Pakistan, if I list a few, these insights are like No. of Pizzas sold on the New Year's Eve by Swiggy and Zomato, No. of Diyas ordered on Diwali at Blinkit, No. of iPhones Sold at the Lauch Day in Couple of Minutes at Blinkit etc. This type of information is strategically important for any Organization in identifying the Consumer Trends, such information not only enables the organization to plan Operations in a better and eƯective way but also paves the way to identify and allocate the resources on the products with maximum returns.

Basic set of resources required for any Organization to identify the insights like above are, an ERP Software and Qualified & Trained Personnel who can understand the data captured by the system and then help the management with such insights for decision making process, among other constraints acquiring both of these resources is costly.

In these cases, LLMs can be very useful for an organization for both Accounting and Management Information System.

3. Methodology

Let's begin with accounting first, just like at a departmental store the cashier scans the product and the invoice is generated with correct inventory and value, what if on scanning such Invoice we can capture all the very basic fields such as GSTIN of Vendor, Name of the Vendor, Date of Invoice, HSN Code effecting the Inventory, Number of Quantity, Rate per item of Inventory, Rate and Amount of GST and Invoice Value. These are very basic particulars of any GST Invoice and using the LLMs this data can be easily tabulated in a spreadsheet and feed to the ERP & Accounting Software in the matter of few seconds. Such technology will not only help in increasing the efficiency of the existing team but will also assist the team to accurately record the transaction.

LLMs are capable enough to read and analyse the data, identify the trends, summarise the data and present it in a report form. Let's understand it with some use cases, on a simple prompt to the LLM regarding the best performing Sales Person of the Month, the LLM will access the Sales Data for the month, it will summarise the data Sales Personwise, arrange it in an orderly format and then will answer the user all within faction of aminute, not only that on subsequent changing the parameter of the prompt the new response would still remain quick, say now the information is sought regarding SKU wise or the Region wise or some another parameter, the LLMs are capable of amending its results considering the new prompts, and presenting all of them in the chronological order, thereby ensuring that the reporting of results is not Selective or Biased towards correct Results only.

Use of LLM in Production and Planning Department of an Enterprise, Production and Procurement activity of any organization are dependent on the Purchase Orders issued by its Customers, all the Customers may not have the same form and structure of a Purchase Order Form, also each customer may issue more than one purchase order for a single product based on warehouse locations, hence it is the primary task of the Planning and Production team to estimate the total units of the Goods to be Produced and Procured, ordinarily these Purchase Orders are recorded in the ERP and using an ERP, the Production and Planning team can decide the future course of action. Now if we assign this task to LLM simply by uploading all the Purchase Orders (can be in either pdf, text or excel format) and asking it to summarise the same, it will not take more than a minute to summarise all the Purchase Orders SKU-wise, this will help the team in increasing the efficiency by eliminating the manual data entry of the Purchase Order into the ERP for Production / Procurement Planning.

The next question which arises is how to use existing ERP Database with any LLM for MIS Reporting? Primarily there are two ways, first is to manually feed the raw data to the LLM and proceed with prompting, second is the automatic route which enables Realtime access of ERP data to the LLM, in this case the LLM will be able to produce results based on the whole of ERP Data, hence the LLM will have enhanced scope of source informationto process the Prompt and furnish the results. This automation is achieved using the AI Agents. AI Agents are not only used for fetching the data from the ERP but can also be trained to input the data into the ERP based on the outcomings of LLM Prompts. As per Gemini,

"An AI agent is a software application that uses artificial intelligence to make decisions and perform tasks independently with minimal human input. They can reason, plan, and act to achieve goals, often in complex or dynamic environments. AI agents can learn from experience, adapt their behaviour, and even collaborate with other agents to accomplish complex tasks".

Best examples of AI Agents which are encountered by us all in our daily lives are Customer Service Chat Bots and our Digital Personal Assistants such as Apple's Siri and Amazon's Alexa. Such AI Agents are developed and trained for user specific needs.

4. Results

Experiment No. 1 - Invoice Details Extraction

Objective - Automate the Purchase Invoice Data Entry into the Accounting Software

LLM - ChatGPT

Workflow - We trained the LLM to look for specific details on the Invoices, Extract & Tabulate them, then we Uploaded Certain Purchase Invoices and received an Excel Spreadsheet as an Output.

Observations - LLM was able to gather the information as sought and present the same in the tabular format however there were certain instances where it failed to provide correct / accurate output.

Positive Outcomes - It was able to identify the information which was missing in the Invoice, for example it was trained to extract IGST, CGST & SGST from the Invoice, when the uploaded Invoice was for IGST it marked both CGST & SGST as "0" in the table and vice versa.

It was able to identify that the uploaded document has one irrelevant document, for example a pdf uploaded was along with the copy of e-Way Bill but the LLM Extracted all the data from the Invoice Only and didn't seek any further user input for that specific pdf.

Optical Character Recognition (OCR) capability, one of the pdfs uploaded was not a native pdf but was an image converted to the pdf, LLM notified the issue to the user and sought permission to run OCR and thus was able to extract the relevant information from the file.

Negative Outcomes - One of the Invoice had two separate HSN Codes, it failed to identify both of them separately, it summed up all the values against one HSN Code Only, which was factually incorrect.

In the instance with OCR Invoice, it failed to recognise the two separate Invoices and Added both of them together, resulting into incorrect Data Extraction.

Experiment No. 2 - Sales MIS Reporting

Objective - To identify, Top 10 Customers, each Region-wise Top 3 Customers, Weekly Sales, Region-wise Sales Comparison, Last 10 Customers from the data set having 4 Columns with headings as Date, Customer Name, State and Taxable Amount. Data Set contained 153 Rows of data, it comprised of 65 Unique Customers and 18 Unique States.

LLM - ChatGPT 4o

Workflow - Trained the LLM to identify the above referred parameters and report the same using Charts where-ever applicable. Uploaded the excel file and received the Output in less than a minute.

Observations - The LLM was able to perform the analytics and report it in the Tabular and Chart Format.

Positive Outcomes - LLM was able to perform all the analytics and report the same all within short of a minute, whereas if all these tasks are performed and reported manually it would take 20-25 Minutes, thereby depicting a very sharp advantage in eƯiciency.

The outcomes in both the tabulated form and in chart form were precisely accurate.

Even though the version ChatGPT 4o was chargeable, the cost of subscription is much lower against the cost of setting up the ERP, hiring and training the Personnel.

Negative Outcomes - We were able to generate this analytics and reporting in ChatGPT 4o model, the lower versions of the ChatGPT were not helpful at all for the analytics, not only that even Google Gemini's 2.5 Flash version failed to perform the analytics. The takeaway from this is that such high level of accuracy and capability of analytics is not freely available to the users as the higher versions of both of the above referred LLMs are chargeable.

For experimenting purpose, we prompted these very details to the "ICAI Direct Taxes GPT", it served the answers, but all the details served were unrelated to the data set furnished by us, it felt that the results furnished by the GPT was based on the data set it was trained on.

5. Conclusion

The biggest takeaway from this research is that LLMs have the capability to not only reduce the Cost but also to increase the Revenue. This is a double-edged Sword which will play a very crucial role in the growth and relevance of an Organization in the coming days.

Generally, an organization has three major level of Management groups, Top Level Management comprises of Board of Directors or Owners who are responsible for Strategic Decision making, Middle Level Management comprises of Team Managers who oversee the Operations on day-to-day basis and report to the Top Level Management, at last the Lower Level Management, it comprises of the personnel who are responsible of doing the job. In terms of Accounting and MIS Reporting leg of any Organization, the use of Artificial Intelligence can be a tremendous game changer, if implemented correctly it can blur the demarcation between Lower and Middle Level Management.

OpenAI's CEO Sam Altman was recently interviewed with Bloomberg, he was questionedregarding the Riks of investing such hugely in this technology, to which he replied that, "I mean, maybe, you know, people stop wanting to like pay for AI Services".

But, the results of AI adoption are very much tangible for any organization, recently, Paytm, One 97 Communications Limited's Chairman, MD & CEO, Mr. Vijay Shekhar Sharma stated that "it used AI automation internally, leading to a 10-15 per cent reduction in employee costs."Recently Zomato, Eternal Limited, laid of 600 Employees under the Customer Support Roles due to AI adoption.

Microsoft Inc., has already integrated it's AI Platform Copilot, into its ERP System. Satya Nadella on the Future of SaaS, during a conversation with Varun Mayya on 13th Jan, 2025, stated that, "right now in my own use case, I go to co-pilot, I say at sales, which is actually touching Dynamic CRM brings back whatever the account information. Then it brings back information from OƯice 365. I put it into Pages, I share it with people. The entire workflow. Now, I'm everyday querying my CRM database because it's so much eaiser" Considering the benefits, and the huge investments made in development of this technology internationally, we must assume that it is going to stay, hence, the sooner we adopt, upskill and internally reorganize the higher would be the rewards.

Overlooking all the positives and benefits from using this technology, there are vast concerns regarding Data Privacy and the Confidentiality. Conversations with these models are used to train them; also, these models expressly caution the user regarding sharing Sensitive Personal Information with the Model. Though the users are empowered to manage data control preferences, their conversations may be used for training the model in the free versions.

References:

1: What is an LLM - Definition from Chat GPT

2: Future of Jobs Report 2025 - Published by World Economic Forum (WEF) on 7th of January 2025, can be accessed at, https://www.weforum.org/publications/the-future-ofjobs-report-2025/

3: Interview of OpenAI CEO Sam Altman with Bloomberg can be accessed at,https://www.youtube.com/watch?v=GhIJs4zbH0o

4: Statement of Mr. Vijay Shekhar Sharma can be accessed at, https://www.businessstandard.com/companies/news/paytm-joins-hands-with-ai-startup-perplexity-to-offer-in-app-search-125022700262_1.html

5: Article regarding LayoƯs by Eternal Limited can be accessed at, https://www.businesstoday.in/latest/corporate/story/zomato-layoƯs-600-employeeslet-go-as-ai-customer-support-takes-over-470197-2025-04-01

6: Satya Nadella on the Future of SaaS conversation with Varun Mayya can be accessedat, https://www.youtube.com/watch?v=GuqAUv4UKXo