Several clients have inquired about our stance on the use of Artificial Intelligence and Large Language Models. At first, our understanding of AI was limited to its use in language and writing – not in the areas of deep analytics and decision-making. 

Although AI tools can process vast amounts of data and understand the natural language in command, they still lack the strategic judgement needed to shift paradigms or think outside of the box. 

Sure, they can help us answer the “what” and even the “how,” but not the “why” or the “what next.“, especially where there is a lot of external data or unknowns that the AI needs to process. For example, AI doesn’t know when a KPI is flawed, when a business rule is outdated, or  whether the real issue lies in governance and not operations. 

Acknowledging these limitations, we set out to explore how AI could strengthen our data analysis work across client projects. We focused on building and testing a reliable process—one that uses AI to speed up delivery, improve insight, and support decision-making, all while keeping our expertise and judgement at the core. It’s a practical approach that worked for us, and one we believe is worth sharing.

The Trigger: A Real Client Challenge

Our exploration of AI solutions for projects started with a client that engaged us for a Rapid Inventory Reduction Work Package, where we analyse their data and use our best practices to optimise and propose new stock levels and policies to reduce working capital. 

This client was using SAP. Our usual process would start with pulling the data from the relevant SAP tables and loading them into our Excel template and SQL queries. These tools help us clean up the data, carry out our own data validation, and enhance the data. From here, we then build the data model, run our analytics simulations, and apply our best practices and knowledge. 

But this time, we proposed something new: what if we ran the entire analysis using ChatGPT? The client was open to experimenting, and that was the spark.

From Concept to Execution: Our AI-Driven Process

To explore the real-world potential of AI in supply chain optimisation, we developed a clear, auditable framework—this was never a black-box exercise. Every step was grounded in business logic and guided by our expert oversight. From data extraction to simulation, validation, and optimisation, our process ensured control, accuracy, and repeatability. 

What follows is a practical, step-by-step account of how we used ChatGPT—not just to analyse data, but to embed logic, run scenarios, and drive decision-making. Domain expertise shaped each phase, reinforced through testing and refinement.

1. Data Preparation: Extracting SAP Tables

As always, we would pull out the data from the standard tables, and it represents about 3 months of operational history

We exported the following core SAP tables as CSVs:

• Sales orders – Header and item data (VBAK, VBAP)
• Stock levels – Plant and storage location stocks (MARD)
• Material master data – Basic and planning views (MARA, MARC)
• Purchasing data – Purchase order header and items (EKKO, EKPO)
• Material movement history – Transaction-level movement data (MKPF, MSEG)

We also pulled some non-SAP data from the third-party logistics provider (3PL): Transport cost on the delivery level.

From here, we would usually clean, filter, validate and rejoin the tables in our database back into Excel to start the analytic work. 

2. Joining the Data: Guided by SAP Expertise

However, this time, we uploaded all the data into ChatGPT and asked it to join the tables using its knowledge of SAP tables. Success, almost! While ChatGPT knows how to join most tables, we still need to check and correct them. 

The integration of the non-SAP data into the data model was quite simple as well. 

To speed things up, we imported our current data model from our database in SQL and imported it into ChatGPT for validation. 

Things look good so far.

3. Cleaning & Validation 

Next, we started to clean and filter the dataset using direct prompts. For example, we told ChatGPT to 

• Only consider Sales order with the transactions category = C
• Show us any materials that were not using the reorder point
• Find and fix any issues where the reorder points did not have realistic value. 
• Detect and fix material master issues that we usually know about. 

All this was done through natural language—so SQL or VBA was required. This made the task feel very natural to the user. It also helps that GPT handled all the complexity of the query building so we could make the validation and data cleanup much faster. 

4. Contextualising & business model

Next, we had to build our data model. While we do have tools (SQL and Excel) that comb through the data, convert or correct missing data, or adjust data using ChatGPT, it was pretty simple, as it knew most of the data in SAP. However, we did have to correct quite a few details, such as:  

• Telling ChatGPT to treat a custom movement type (e.g. 901) as equivalent to standard (101)
• Overriding zero reorder points with realistic values
• Filtering out transactions irrelevant to planning logic

Easy peasy, but we were careful to always double-check the work back in Excel.

5. Embedding Business Logic

So now it was time to teach ChatGPT to identify certain scenarios, such as: 

• Identify backorders 
• Detect late supplier deliveries
• Calculate demand spikes, or spikiness of materials 

This involved both clear rule-setting and illustrative examples, as without it, ChatGPT would’ve made basic errors. But once trained, it could reason like a junior planner—with guidance and training wheels, of course.

Crucially, we also brought our SAP consultants and business advisors into the process to explicitly add these rules into ChatGPT. 

We didn’t rely on guesses—we embedded proven logic from real-world planning environments. Only with this embedded expertise did ChatGPT become a valuable tool for structured simulation and diagnostic work.

This can’t be overstated enough: ChatGPT hallucinates, makes incorrect links and makes guesses, so it alone is not the solution. The results we achieved came directly from years of domain experience—knowing what to ask, how to test it, how to correct errors, and how to interpret outcomes. Without that expertise, the system either drifts or delivers misleading output. 

What made this work wasn’t just AI—it was guided AI, driven by specialists who know SAP inside and out.

6. KPI Validation Against SAP BI

Once the data model was built clean and validated, it was time to test ChatGPT’s outputs against the client’s BI dashboards:

So we compared ChatGPT’s results against the client’s data on:

• Backorder ratios
• Stock turnover
• Supplier OTIF (On-Time In-Full)

If, at any point, the numbers were off or incorrect, we always went back to the raw data from ChatGPT and the client’s KPI calculations to see where it went wrong. In some cases, we had to fix assumptions or tell GPT how to carry out some of the calculations, hence why it is essential to always check GPT’s work. 

Only once they matched did we proceed with the simulations and exploratory analysis.

7. Exploratory Analysis

So, now it was time to carry out our exploratory analysis. This is where we will make changes to the data to simulate key policy changes and their effect on the business. 

Some examples are:

• Top 10 Stock Keeping Units (SKUs) by backorder frequency
• Materials causing the most revenue loss
• Suppliers or customers with consistent delay patterns

Here, we noticed that asking a generic question like “show me all vendors that are consistently late” returns unexpected results. Queries should be more explicit. In this case, the question was, “Using the rules for vendor lastness, show us the top 15 vendors that are late by more than 15%, and rank them by the material ranking.” 

This means that GPT was using our logic that we taught it before:

• Determine if a Purchase Order (PO) was late by looking at the SAP tables. 
• Elimination of certain PO types
• An internal scoring of lateness based on material, vendor, etc
• Rank that in a list

The key accelerator here was that we were able to change these rules quickly, say ranking by profit margin, and re-run the simulations. This is the advantage of ChatGPT: to quickly and easily change the rules once all the data is in place.

8. What-If Simulations

So, the next step is to test the  “WHAT IF” scenarios, in which we change the parameters of the data to see how the model reacts. Examples were 

• What if customer demand increased by 10%? 
• What if suppliers delivered everything on time?
• What if we changed the replenishment frequency?

While we do have Excel tools for this kind of modelling, it took somewhat more complex data manipulation, in-house logic, code, and time to apply this logic. 

With ChatGPT, changing parameters was simpler as it happened in natural language rather than changing some parameters in the spreadsheet. Again, validations were required, but once stable, we could change the scenarios using natural language logic, usually using our internal expert’s formulas.

9. Daily Stock Simulation

Finally, simulation models. This is where, based on the dataset, we build a simulation engine to execute these transactions in sequence and inject our “WHAT IF” into the simulation. 

We taught ChatGPT to simulate daily stock flow by processing every transaction in sequence like so:

• Goods Receipts → increase stock
• Inbound Sales order → reduce stock; if there is not enough stock, place on backorder. 

Of course, we have to add quite a few assumptions to simplify the model, but we did create a working model of the daily stock levels so we knew that stock level at any point in time. 

This allowed us to inject what-if scenarios directly into the simulation and immediately observe the impact. For example, we could tell it to increase Goods Receipt (GR) quantities by 10%, shift shipping frequencies, or introduce artificial demand spikes—and it would recalculate the stock position and service performance automatically for each day. This gave us an interactive, flexible sandbox built on real SAP transactions.

We even tried a number of basic optimisation heuristics where we asked GPT to run the simulation x number of times, each time changing a value or factor for each run and ranking the results. 

10. Embedded Best Practices

With the simulation working, we then layered in ABA’s optimisation rules and best practices (again in natural language) and re-ran the simulation, trying out scenarios such as

• Reduce safety stock for low movers
• Convert daily shipping to milk runs
• Increase buffers for high-value SKUs

Of course, the definition of slow movers, milk runs, etc., had to be taught to GPT, but we should quickly stress-test the system to measure items such as the cost of capital, service impacts, etc.

Again, lots of testing and lots of validations.

Challenges, Opportunities, and Considerations

As we’ve established, AI-powered tools like ChatGPT do present a powerful opportunity. But likewise, using ChatGPT requires an understanding of its limitations, risks, and user knowledge on how to decipher the root of the problem and rectify those errors. You need both a driver and train tracks for the train to arrive at its destination safely.  

Limitations and Learnings

 During testing, we discovered a number of key risks when using AI:

• Guardrail drift: While we have put in guardrails, it is pretty easy for an untrained user to accidentally go off course and start to misalign the data or model, making it very difficult to back-track afterwards. You still need a trained, knowledgeable user guiding the ChatGPT. 
• Poisoning the well: Once a bad rule is fed, it contaminates downstream results unless reset. Rules are needed for the user as well as the ChatGPT. 
• Hallucinations: ChatGPT can completely make things up or make wildily wrong assumptions either because of incomplete data or rules that aren’t correct or explicit enough. Another key factor here is ensuring the user has the knowledge to identify incorrect information and where to make adjustments to instructions.  
• Dilution: Across analytics (or writing), ChatGPT has a tendency to dilute information as prompting continues in a Chat. Chat GPT can often “forget” to do a task such as updating a table or field etc. The lesson learned here is never assume; consistently and explicitly ask ChatGPT to perform each task.  

One example came early in our testing. We asked ChatGPT to “read the total gross weight from the sales order and distribute the transportation cost using kilograms” in order to analyse product profitability on a per-kilogram basis.

The output looked plausible—but during validation, we noticed the numbers didn’t add up. The issue? ChatGPT had tried to extract gross weight from the sales order header table (VBAK), which doesn’t contain that information. In SAP, weight is only recorded at the item level (VBAP). ChatGPT had effectively invented header-level values.

Once we corrected the prompt and told it to sum VBAP-NTGEW (net weight per item), the results were accurate.

These risks stem from an inherent limitation of ChatGPT: you cannot immediately see how it works step-by-step. ChatGPT is functioning more like a black box rather than a transparent program code that we could trace. 

In this case, because ChatGPT had written Python code to perform the calculation, we were able to inspect and debug the code to pinpoint the error. 

Despite being an extremely accessible tool presenting information in a chatty and familiar way, AI cannot be used for complex, value-driving work without an experienced user both in ChatGPT’s underlying data structure and the subject matter.

Reusability

A core advantage of using ChatGPT is the ability to consolidate and reuse the rule-based knowledge – especially those implemented during testing – across different contexts. We can ask it to “download” all the rules that it was taught, such as:

• Data joins
• Cleanup rules
• KPI formulas
• Simulation logic

These can be stored and reapplied consistently to streamline work for future client engagements. Unlike complex SQL or macros in Excel, inputs are written in plain English – transferring knowledge, making changes, and reviewing or adjusting is significantly easier with streamlined processes using ChatGPT. 

 Broader Considerations

It is important to note that this article details a specific, controlled use case of ChatGPT Enterprise with anonymised data. Appropriate security safeguards are in place so that no data is used to train the LLM or be accessed outside the ABA environment. 

While our experience focused on ChatGPT, we acknowledge that it is not the only tool in the market. We are aware of alternatives like Google Gemini and SAP Joule, and are open to exploring them should the opportunity arise.

AI Didn’t Replace Our Expertise—It Amplified It

This wasn’t just a tech demo. It was a working proof that AI – when guided by deep domain knowledge – can help us deliver supply chain insights and simulation, and support faster decision-making.

We now see AI as a tool to extend our capabilities—not replace them.

But let’s be clear: without the right expertise, none of this would have worked. It had data and a simplified model, but not the human intelligence to make the best choice. It took real SAP and business consulting experience to structure the data, build valid scenarios, and detect flawed assumptions or incorrect data. ChatGPT was powerful, but only because we made it so.

AI can help answer the “what” and even the “how,” but not the “why” or the “what next”.  This depth of insight still requires experienced professionals who understand the broader context of the data. They can decide when to break the current business model or way of working, when to refine and improve it, and how to do so. Human intelligence allows us to make the best possible decisions with limited information and balance moving parts in the real world. 

This approach isn’t about automation. It’s about leveraging AI as an extension of expert judgement to accelerate delivery, raise quality, and scale insight across clients.

So, moving forward, we’ve codified our ruleset and internal guardrails—not just for reuse, but to continuously evolve how we deliver insight at scale to help our next client.