What AI revolution?
After doing research on some signifigant issues that food companies are having with integrating with 3rd party delivery servies like Uber Eats and DoorDash, it seems like the problem is much bigger that it initially seemed.
Original thread: https://www.wallstreetoasis.com/forum/investment-banking/are-restaurant-pos-systems-are-a-potential-buyout-target-for-companies
Basically, the vast majority of food companies (Papa Johns, Burger King, etc) are not accurately estimating prep time for their food, and this is causing a huge issue with integrating with 3rd party delivery services since they can force a delivery driver to sit and wait in a pizza shop for 30-45 minutes instead of being out delivering. In extreme cases, I've seen a single McDonalds shut down delivery for half a mid sized city because they were chronically understaffed and ended up queing 10+ 3rd party delivery drivers waiting 30+ minutes for food in their loby.
After talking to managers at several major chains etc, its actually a much biggere issue than simply integrating with 3rd party deilvery platofrms. People have been ordering online for probably close to a decade now, but as a whole, virtually every major company is completely failing their customers regularly by not being able to give proper time estimates for online orders. If you call in an order, the employee can look at their order que and give a rough estimate of how long the wait will be, but as far as I can tell, Online ordering AT BEST just uses afterage wait times for "thursday at 5pm". If someone ahead of you ordered 50 pizzas, you're going to be sitting in the lobby waiting an hour since the online ordering system doesnt take into account any actual factors in estimating your wait time. I regularly see lobies full of pissed off customers waiting 30+ minutes for their take out food that was ordered online. Some companies like Little Caesars have taken steps to slightly aleviate this issue by tracking where your pizza is in its production process, aka "your pizza just went into the oven", etc, but thats still a long way off from just telling someone an accurate wait time online.
A.I. that can factor in order ques and staffing levels can easily give a very accurate order estimation. The thing is, this level of AI has been around forever and can run on any old PC. The following is more than sophisticated enough and its an intro course: (google "kaggle intro to machine learning").
It seems like these companies have no clue what they are doing when it comes to AI since they haven't even implemented a decade+ old solution to a HUGE problem they are all facing.
Just based on my research into one single segment of the US economy, it seems like comapnies aren't actually doing anything real in regards to AI implementation outside of possibly big tech companies. I see a ton of companies laying off programmers and not hiring so they can claim that AI is making them more efficient, while in reality they are completely failing their customers and not even implementing the most basic AI solutions. Maybe CTO positions at more ops focused companies are still just there to serve as a head to roll if and when a data breach occurs?
I understand that many of these companies may want to adapt, but are unable to because their software providers aren't adapting. Someone told me that most restaurants use a Micros platform which is owned by Oracle, and it cannot do what I outlined.
Isn't Oracle supposed to be an AI focused company?
Just one example:
At one of the largest pizza chains in the country, in the lobby were two doordash drivers. Both were told after arriving the wait would be over 30 minutes. One of them already had a previous hot food order, that would be ruined by a wait that long, but they still got the pizza order added onto their delivery. Either critical order wait time is not being shared or doordash has the worst algorithms in the world. The pizza restaurant manager confirmed that their software cannon estimate wait times.
Of note: the driver with a previous order removed themselves from the new pizza order so the first customers food wasn't ruined. Doordash then lowered that drivers rating as punishment.
Its not always just as simple as "implementing AI", like in this case, you must have software platforms covering online ordering, scheduling & clocking in, and order processing/production that all talk and communicate with eachother. For modern software with functional API's this isnt an issue, but many companies have old legacy software that may even be running on COBOL that "just works".
Well, no matter how bug free it is, if you arent taking care of your customers that order online with a good estimated wait time (as pizza companies have been providing their customers for decades who called in...) then your software absolutely doesnt "just work" anymore. It is now a complete failure.
Also, this isnt my analysis, but when I spoke with the manager of a popular nationwide donut chain, he said if he ever runs out of one type of donut, he has to log into 5+ different systems just to X it so customers dont keep ordering it and get dissapointed. When he has more in stock again, the process repeats itself. This isnt AI, but it seems like an easy fix... if the applicable software platforms have modern APIs built in.
The AI revolution in the food industry, particularly in addressing operational inefficiencies like those you’ve outlined, is indeed lagging behind its potential. Based on the most helpful WSO content and insights, here’s a breakdown of the situation:
1. The Problem with Integration and Estimation
2. AI as a Solution
3. Why Companies Are Failing to Implement AI
4. Examples of AI in Food Tech
5. Opportunities for Improvement
6. The Bigger Picture
In summary, the AI revolution has the potential to transform the food industry, but many companies are failing to capitalize on it due to outdated systems, lack of expertise, and misaligned priorities. There’s a clear opportunity for innovation, particularly in developing AI-driven POS systems and fostering collaboration between restaurants and delivery platforms.
Sources: https://www.wallstreetoasis.com/forum/investing/leisure-and-hospitality-take-the-lead-the-daily-peel-4102023?customgpt=1, Google partners with Goldman Sachs in automating Investment Banking, From Hoenig to Hamlet | The Daily Peel | 4/12/2023, Amazon: King of Commerce, Will robots replace your consulting or financial career?
This is a good writeup, and it generally hits on a lot of the failures of these models to actually predict things.
I have a bit of a technical background, so I generally have an idea of how these predictive models work. At their core, they all run on the basis of K nearest neighbors. That's simply a predictive model based on past data.
Let's imagine we're Zillow, trying to predict a housing price. We can try a regression(basically just putting traits in with coefficients to create a line of best fit as a predictor), but this oftentimes will miss on certain elements. Instead, what we might do is take the nearest K houses, say 3, and use them to predict our house price. So, if the nearest houses sold for 200k, 300k, and 350k, we can simply say the average house in the area is 283k. We can use that to predict our house price. We can add more components, like sq footage and bedroom count. Soon, we have a model that can kinda predict our housing prices. We can backtest this model on previous houses to see how accurate it is, and based on that adjust our model.
At its core, this is sort of what LLMs are doing. They're taking the K nearest sentences, and predicting the next word based on a history of what other sentences have said. Of course, this is a VAST oversimplification, but at its core the predictive element is the same, just with a fuck ton of parameters and other elements such as attention.
There lies a core problem in this. It is predictive, but oftentimes our predictions are wrong. In your example, a model might predict that the last 10 Wednesdays have been pretty quiet, so it's unlikely this next one will be too loud. What the model fails to predict is that a party could be happening, or a concert in the area, or some other event, that could massively increase traffic to the restaurant. This is something that a person might be able to intuitively figure out based on reasoning, but these models cannot reason. They cannot think. They can simply use the data they have to predict things.
These models are generally very good at predicting things 98% of the time. That's just the case. But those 2% of the time is when things get majorly fucked. A self driving car that is able to predict things 98% of the time will still hit someone 2% of the time, or break traffic laws 2% of the time, or cause some other disaster 2% of the time. A model using past data might say that since the S&P 500 has never declined over a 20 year period, that it is probabilistically impossible for it to happen, and could shovel money into the S&P. But lets imagine a world where, say, the S&P did have a major multi decade decline. Well, then this model is broken, and has lost a lot of money. Predictive and risk models have done this plenty of times before. In fact, in almost every major crash, predictive models broke, because they can only predict based on past data. It's not the models fault. And that's how you get events like 2008, where many models predicted that a crash was impossible because housing had never declined nationally, despite all the evidence to suggest otherwise.
The models are really good. Really really good. I work with these models in trading every day, and they are really good models. But they're still just mathematical models doing their best with the data they're given, and so having people who can look at the models outputs and say "Well that looks really wrong" are incredibly important. Otherwise, we get bit shitstorms.
Agree except that safe driving cars are actually much safer than human drivers. No lapse in attention, no using phone, no old/young driver stupidity.
While this might be the case, it's not necessarily the point. The point is that human drivers can think, while self driving cars cannot. Most of the time, this isn't an issue. But take the example of the cruise car in SF that dragged the woman under the car, because it couldn't properly detect that it was a person. A human driver reacts very differently to that because the human can think, while the self driving car can only do its best based on the data it was trained on.
False. Most data we have is very good and high quality data. You can feed a model amazing data, and it will work really well with that data, but everyone in finance knows the saying: past performance does not guarantee future results.
Also, data can be great, but a lot of data is naturally very noisy. You cannot remove that noise. In predictive models there is a Y = I + ε. ε is irreducible error, and with some datasets or problems that irreducible error is impossibly high.
That's interesting that you brought up real estate pricing since that's what the kaggle lesson that was referenced was about.
For fast food, I would argue that these models (ones that can't use order queus or staffing levels and can only use historica busyness) are wrong the majority of the time. For a sit down restaurant, the user interactions are roughly 45 minutes (guess), so the flow to the restaurant is smoothed out or averaged, so busyness predictions can be much more accurate outside of things like concerts etc like you mentioned. However, for fast food where the total interaction time should be 5 minutes or even less, the volatility is exponentially higher. In addition both sides of the prediction being wrong are bad. If it says 20 minutes but the store is empty, someone's food is going to sit out 20 minutes before they arrive. Imagine looking at your charts showing 1 minute ticks with a 3 minute moving average vs a 1 hour moving average. Long story short, past busyness based predictors for fast food or other low interaction time situations, at least for a single fast food restaurant, are likely very wrong more often than they are close to right. Further, any peaks tend to carry down the line like traffic backups on the interstate. Obviously if you increase total volume, say the world's biggest and busiest McDonald's etc, you get more accurate, but with typical fast food or pizza joints, the volatility is way too high to use models that just use past busynes and can't factor in order queus and staffing levels.
Anothe issue that makes this even worse for online orders is that there are virtuall no natural regulators for demand, while supply tends to stay capped. If someone pulls up to a McDonald's and sees a line around the block they might avoid it, decreasing demand, but that doesn't happen with online orders, so things can get worse than normal. Kinda like how a stock being up 20% might make some traders sell (and some buy...).
Edit: just to be clear, I'm not saying the models in general are bad, my point is that those models only using past busyness are flying blind in an environment with the volatility of a fast food joint and are thus fairly worthless. It would be like you trying to do what you do without knowing the actual stock price or something like that. With staffing levels, current order queus, and possibly some other variables, the models can likely be very accurate for fast food wait time projections.
Yes, I think I generally made this point clear. These models are very good most of the time, but we don't care about most of the time. We care about the few times when the models get things very very wrong. And it is generally the case that something like fast food has extremely high variance in terms of customers from location to location, day to day, hour to hour, etc. that make it really quite hard to predict. You end up with models that are overbiased or have too much variance due to an overfit. Such is the variance bias tradeoff. With many other things, like stock prices or predicting shows on Netflix, this is usually not too hard to work with. You can usually find a model that works for the general case than minimizes the least squared errors while also not overfitting. For something like fast food, however, the variance is so high that the risk of overfitting becomes so high the moment you add more than even a few elements to a common regression(or in a KNN case, the less nearest neighbors you have, the higher the variance and lower the bias). This is a fundamental issue with using a predictive model that cannot actively reason. A predictive model can predict something, but a human might reason that the models figures are totally off.
Yeah, you definitely did a good job explaining that. I've actually never heard someone explain the bridge between data based AI and large language models so well or simply.
I more wanted to hone in on why those models that only use past busyness as the only factor work well for the food industry as a whole, like sit down restaurants, but completely fall apart when applied to something like fast food. Like you were saying, for fast food, the variability in demand can vary drastically even minute by minute during peak times. It's been forever since I worked retail, but I remember it being more marked by surges and boredom vs a more steady flow, even during peak times, and that seems to align with my general observations as a customer these days.
Ex: at 5:15pm on Wednesday there is a line of 6 people in the lobby, but from 5:21 to 5:27, there is nobody in line. By 5:40 there are 8 people in line again. That's just the normal ebb and flow, not even looking at things like a tour bus showing up.
Also, please excuse all my typos, I'm usually trying to type on my phone with one hand while eating or something.
Of course AI doesn't have a major commercial use case. It's a bubble. A fraud, essentially, being shoved down the public's throat to give the impression that big tech companies still have a path forward for major growth.
Otherwise you have to start valuing these businesses based on their actual profits, and in the same manner as any other mature business. Which means you're talking a 25-50% write down in valuations, across the board, for tech firms.
Impedit facilis velit vitae temporibus expedita. Consectetur eligendi sed eligendi ut. Explicabo labore adipisci quibusdam vel velit. Est reprehenderit laboriosam sit dolorem. Et facere voluptas qui velit.
See All Comments - 100% Free
WSO depends on everyone being able to pitch in when they know something. Unlock with your email and get bonus: 6 financial modeling lessons free ($199 value)
or Unlock with your social account...