ChatGPT goes to Wharton
Can ChatGPT run a business? Color me very skeptical.
This is a shockingly easy question, but I would not have predicted that ChatGPT would get it right, given its limited ability to do math and think logically. So, I am updating my assessment of ChatGPT (slightly) in a favorable direction.
Still, before we get too excited, it helps to contextualize this a bit.
Suppose you are a newly hired executive or strategy consultant. You are trying to figure out how to add value. Identifying the bottleneck in a production process might be a useful thing to do. But to get to the point of doing this analysis, you first need to figure out that identifying production bottlenecks is worth doing, by asking questions that would allow you to home in on this issue. You would start with very general questions, and gradually get more specific: Who are your customers? Who are your main competitors? How does the firm access capital? What explains recent changes in financial performance? How does your pricing work? Do you know your unit costs? The list goes on, and on, and on.
It would likely take a highly seasoned executive/consultant several weeks to get to the point where they could determine that production bottlenecks are even potentially a serious issue (unless the company already knows this and you are hired specifically to address a well-defined question).
Even if you identify production bottlenecks as a potential issue, you cannot make a sensible recommendation without knowing the company’s cost structure and the shape of its demand curve. This means talking to the pricing people (whoever they happen to be) and to the accountants who make cost allocation decisions (regarding joint and fixed costs) and to the engineers who can tell you the capacity and prices of new machinery. If the second reactor – the bottleneck – is very expensive and the 50 ton unit is the smallest size available, it may be that the current arrangement is optimal (unless you face a flat demand curve and have unlimited access to capital, in which case you could calculate the scale that would enable you to achieve 100% capacity utilization throughout the process, but this is obviously highly unlikely).
Collecting data to do your analysis is hard and takes judgment. For example, many companies do not have a good handle on prices and costs. Will you make your own estimates? To do this you will need to ask many questions. Or you may probe and discover that there is a process change that can expand the capacity of the second reactor without the need to buy an additional unit.
The point is, you would need to ask lots and lots of questions to identify capacity as a potential issue that is worth spending your time on, and then you would have to ask lots more questions to get the information stipulated in the problem. You would have to ask questions to interpret the information you get. Along the way you need to constantly reassess the likelihood that this particular line of inquiry will bear fruit and is worth a commitment of your valuable time, and the time of the people who will help you get and analyze data.
I suspect that the capacity to do general management and think strategically in a complex, unstructured decision environment with poor data will require a type of AI that has a model of the world in its “head” and that knows how to ask questions and identify problems and pitfalls. ChatGPT doesn’t do this (at least as I understand it), and my guess is we are years if not decades away from this.
In the meantime, the real benefits from AI will come from smaller and more structured applications.
Eric:
There are systems such as ERP and MRPII which have the ability to consolidate all the information you discussed. Inventory, Demand, Capacity, Throughput, Labor Hours, Production Plan, Business Plan, etc.
This is what I have been doing since the seventies. Computer systems can gather data and supply it to decisionmakers who then decide. Much of this was done at middle management level if there was no capital expense involved.
Similar questions could be answered by using the same data, developing the overall Production Plan for 1-2 years out with a sales forecast. Then weighing the accuracy of variability of the forecast against capital-equipment purchasing. Does the purchase make sense in terms of demand. If it is Labor, can we use some OT. If machinery, etc. related, maybe older less-efficient equipment will suffice in supplying more capacity to throughput.
From that should evolve your overall Business Plan.
Goldratt’s “The Goal” was a nice nontechnical recital of the process.
When consulting, it was fun when the company in question listened to the plan even when they disagreed with capital purchases. Showing them how to accumulate the information and ascertain what could be done was often an improvement in abilities. It went even as far as teaching people who did not know how to manually schedule how to do so.
Thank you for the lead-in to Supply Chain from a manufacturer viewpoint.