Marcus is made up, but his week is pretty typical for someone in his job. He’s an AI implementation consultant, and on Monday he’s at a bottling plant in Ohio looking at motor sensor data. By Thursday, he’s on a call with a regional bank whose fraud model keeps blocking customers on holiday. Friday is spent arguing with a grocery chain about tomatoes.
Spend a week in his shoes, and the chatbot headlines start to look like a sideshow. The bigger shift is in judgment calls that used to live in somebody’s head, or in a spreadsheet only one person understood. Factories now spot a failing bearing days ahead. A hospital gets a second look at a scan before the radiologist’s coffee goes cold, and a supermarket can order closer to what people will actually buy.
Marcus would be the first to say it isn’t magic. Most of his invoices go on cleaning up data and figuring out who’s responsible when the model is wrong. Here’s what that looks like in a few different places.
The skills gap Marcus runs into everywhere
Almost every client asks him the same thing sooner or later: “Can we just train someone in-house?”
Usually, yes, but it takes more than a weekend course on prompting. Anyone building a system a bank or hospital can trust has to understand how a model learns from examples and why it might fail on data it hasn’t seen. They also need to be comfortable with statistics, Python, and messy databases.
Marcus came into the field sideways, from supply chain work. A few of the junior analysts he’s mentored went a more structured route while keeping their day jobs. An online bachelor’s degree in artificial intelligence, like Champlain College Online’s Data Science and Artificial Intelligence BS, is one example. It’s a 120-credit program taught entirely online, starting with linear algebra and calculus and moving on to Python, R, and SQL. Later courses cover machine learning, deep learning, computer vision and language processing.
He still tells every new hire that the code is the easy half. “I’ve seen a model with 98% accuracy answer a question nobody at the company was asking.” Knowing the business is what stops that from happening.
Monday: a vibration nobody could hear
The bottling plant used to service its conveyor motors on a calendar, every twelve weeks whether a motor needed it or not. Last year one failed between checks during a holiday rush, and the line sat idle for most of a shift.
So the plant fitted sensors that track vibration and temperature, and Marcus’s team trained a model on months of readings. It learned what “normal” sounds like for each motor. Now, when a reading starts creeping upward, the night supervisor gets an alert, and the technician swaps the bearing during a planned break.
The plant also runs cameras over the bottles to catch crooked labels and missing caps. Teams using AI vision for quality control tend to get steadier results on long runs, when people’s eyes get tired.
Things still go wrong, of course. In March, someone recalibrated a sensor without telling anyone, and the model spent two days crying wolf. The technicians had the final say, which was the right call.
Tuesday: the hospital that didn’t want a robot doctor
The hospital’s chief medical officer opened the meeting by saying, more or less, “Nobody here is letting software make treatment decisions.” Marcus agreed with her, and that made the rest of the day easier.
What the hospital wanted was help sorting through information. A model that flags suspicious areas on a chest scan means the radiologist checks those spots twice. A forecasting tool predicts how busy the emergency department will be on a Saturday night so staffing can be planned. Doctors had also been trying ambient AI scribes in healthcare, which draft visit notes while the doctor talks to the patient. Several said they finally looked patients in the eye again.
The worry in the room was about bias, and it was well founded. A model trained mostly on one group of patients can quietly perform worse on others. A confident wrong answer is risky when a tired clinician is on hour eleven of a shift.
They left with a simple rule. Every AI suggestion gets checked by a qualified person, patient data stays protected, and a named clinician owns every decision.
Thursday: the bank and the tourists
Nobody can check every card payment by hand, so banks lean on machine learning that learns each customer’s habits, like where they shop, how much they usually spend, and how often. When five charges hit a card in ten minutes from a city the customer has never visited, the system pauses and asks.
This bank’s problem was that its model was too jumpy. Customers travelling abroad kept getting their cards declined at hotel check-in, and the complaints piled up. A strange purchase isn’t automatically fraud, and the old model hadn’t learned that.
Marcus’s fix wasn’t very exciting. His team retrained the model on more recent travel data, tested it for fair treatment across customer groups, and set up a review queue so a human handles any borderline decision affecting someone’s access to credit. Regulators expect that kind of oversight anyway, and the bank’s compliance team was visibly relieved.
Friday: tomatoes, mostly
Retail forecasting sounds glamorous until you’re arguing about salad ingredients. The grocery chain wanted to stop throwing away fresh produce every week while also avoiding empty shelves on Saturdays.
The forecasting model looks at sales history, the weather, promotions, and the local school calendar (that one surprised them) to predict demand per store. For produce that spoils in days, even a small improvement makes a real dent in waste.
Then a heat wave came along that the model hadn’t seen coming, and salad sales doubled overnight. The store managers who ignored the forecast and phoned in extra stock came out ahead. Marcus now builds an “override and explain why” button into every retail dashboard he delivers.
A side project: a farm in the Central Valley
A friend of a friend grows almonds and tomatoes, and Marcus spent a weekend helping set up a drone imaging trial. The model looks for early signs of disease or water stress in the leaves, then lines that up with soil moisture readings and the weather forecast.
The appeal is being able to treat the field patch by patch. Instead of watering 200 acres the same way, the farmer targets the dry corner by the road.
It also misread a new almond variety as stressed for most of June. The farmer had been looking at those trees for thirty years and knew they were fine. He was right.
Logistics, briefly
Marcus spent part of the spring with a regional delivery company. Route optimization saves drivers miles and reroutes them when a bridge closes, and vision-based sorting speeds things up in the warehouse. Demand forecasting helps place stock closer to where orders will come from.
The biggest problem turned out to be bad inventory records rather than the AI. If the system thinks forty pallets are in aisle six, and there are actually twelve, the smartest route in the world won’t help.
What about schools?
His sister teaches eighth-grade math, and she has opinions. Her adaptive practice platform gives a struggling student more fractions questions and lets a confident one move ahead. She uses AI to draft quiz variations too, which saves her a Sunday afternoon now and then.
Her students got there before the school did. Pew’s survey on teens using AI for school shows how quickly these tools have become part of homework, and her classroom matches that.
But the software has no idea a student’s parents are divorcing, or that the quiet kid in the back understands far more than his scores show. She also spends real time teaching students to question answers that sound confident and are wrong.
Marcus says almost every project ends with a version of the same conversation. The data was messier than anyone admitted, and someone had to agree to own the outcome. Somewhere along the way, a person who knew the work spotted what the model missed. Next month he’s back at the bottling plant because they want to add a second line.
Related: 2028 Global Intelligence Crisis: Is the AI Crash Already Starting?
