21 Ways People Are Using A.I. at Work

Mr. McNulty, who owns restaurants, brewpubs and dance clubs in Cleveland, uses ChatGPT to analyze sales reports and brainstorm how to grow sales. He’s also used it to help pick wines. He sent a “voluminous” wine portfolio from a distributor to the chatbot and gave it some instructions — specific pricing and particular regions among them — and got back a list, including:

Herdade do Esporão Monte Velho Branco

Region: Vinho Regional Alentejano, Portugal

Grapes: Antão Vaz, Roupeiro & Perrum

Wholesale Est.: $7-9 per 750 ml

Why It’s Great: A crowd‐pleasing white that combines citrus, stone fruit and saline notes with bright acidity — an ideal food‐friendly pour for small plates or seafood.

“The results were astonishingly good and saved me and my team countless hours of meetings with wine reps, tastings and debate,” he said. “The only part of the wine-program building process I missed was the tastings … so far the A.I. can’t recreate the joy of taking that sip.”

There are eight million dried plant specimens at the Missouri Botanical Garden herbarium in St. Louis. Now A.I. is helping identify them.

Experienced taxonomists can quickly recognize most specimens, said Mr. Teisher, but that requires years of training.

So the garden is building an A.I. model using spectral data — the pattern of light reflected by the plant. Leaves from many different kinds of plants are scanned, labeled and put into the model as training data. Then new plants can go through the same process, and the model will identify them. If the model is quite certain that the spectral data look the same as they do for other plants, it’ll say so. If not, the plant can go to an expert.

“We can cut down on the time expert taxonomists are spending on common species,” Mr. Teisher said. “Rather than them getting five boxes of plants that come in, they can get a small box that says, ‘Here are the ones the model isn’t sure about.’”

Those could be species so rare or infrequently seen in the herbarium that the model simply couldn’t match it, or a new species entirely.

And this kind of project is only possible, the garden staff said, because of advances in cheap computing power. The GPUs necessary to train A.I. models quickly are easier to get than before. And the garden has enough funding to process several hundred thousand specimens. Identified specimens can be used for a variety of research, including on biodiversity and climate change.

“We want these data to be compatible with other institutions, so we’re collaborating closely,” Mr. Teisher said. “We have the money to do a big chunk of our herbarium, but ultimately we want this to be a tool usable by everyone.”

Mr. Frazier designs and sells things like bumper stickers and magnetic signs. To help with the graphic design, he uses Adobe Photoshop’s Generative Fill, a two-year-old A.I. feature that adjusts images automatically.

“If I take a picture of a product, and don’t like the glare or reflection I see on some shiny surface, I can use generative fill to ‘imagine’ that part of the photo, and usually one of the resulting images will be acceptable to me,” he said. “Or if I want to use a head shot of a politician on a bumper sticker, and I want to show a little more of the coat or shirt than appeared in the photo I am using, I can use generative fill to imagine that additional clothing.”

A problem that might have taken 20 minutes to address now takes 20 seconds, he estimated.

In one recent case, he wanted to post an image of a bike helmet he’d built himself, but he didn’t think he was the best model for it. So he used Photoshop’s A.I. to generate a woman’s face.

But A.I. can only fill in the gaps for Mr. Frazier. “I have found generative fill to be less useful at creating images from scratch,” he said. “I once wanted to create an image of Joe Biden looking like one of the founding fathers, maybe like George Washington. But I was not happy with any of the results I was getting. I ended up melding a photo of Biden with a lifelike painting of Washington using traditional Photoshop techniques.”

Mr. Soto, an E.S.L. teacher in Puerto Rico, said the administrative part of his job can be time consuming: writing lesson plans, following curriculum sent forth by the Puerto Rico Department of Education, making sure it all aligns with standards and expectations. Prompts like this to ChatGPT help cut his prep time in half:

Create a 5 day lesson plan based on unit 9.1 based off Puerto Rico Core standards. Include lesson objectives, standards and expectations for each day. I need an opening, development with differentiated instruction, closing and exit ticket.

After integrating the A.I. results, his detailed lesson plans for the week looked like this:

But he’s noticing more students using A.I. and not relying “on their inner voice.”

Instead of fighting it, he’s planning to incorporate A.I. into his curriculum next year. “So they realize it can be used practically with fundamental reading and writing skills they should possess,” he said.

Anyone who has ever assembled a “works cited” section knows the dizzying array of styles, formats and specific punctuation rules required for a bibliography. What’s the Chicago Style rule on how to cite books? Do you use quotation marks or underline? Or is it italic? What about in A.P.A. style? A.I. has freed Ms. de Bruin from the most annoying parts of the task. “No more consulting handbooks, guidebooks, cheat sheets, Purdue Owl, fretting about the right punctuation, whether guidelines have changed, and how to cite a three-volume work written in the 18th century, translated by God knows who, edited by Jesus only knows, and originally published where?”

All of this has been replaced, in her words, by “peace, serenity and Claude” (the large language model).

She uses prompts like:

Please cite in MLA format the book University Finances by Dean O. Smith.

or

Give the mla citation for this article: www.chronicle.com/article/higher-eds-financial-roller-coaster

Occasionally Claude cites “Doe, Jane,” and Ms. de Bruin challenges the answer.

“Then, only then, does it respond that it took its best guess at the author because the article was behind a paywall,” she said.

Ms. Swank uses A.I. to take unstructured notes from a visit and turn them into SOAP notes, which is a structured documentation format for health care providers. (It stands for Subjective, Objective, Assessment and Plan — a way of summarizing the visit and the next steps.) It saves her a couple of hours each week, she estimates, “but more so it helps me complete the task that is so easy to put off.”

Ms. Triandafellos uses A.I. as inspiration for her art practice. She uploads dozens of images of her artwork to get the A.I. model to understand her style, then guides the model with prompts to generate new works based on her style. What she gets back are hundreds of abstract images in a grid:

She studies them the way a psychiatric patient interprets an inkblot test.

“I looked at each image and wondered what it reminded me of, reaching my subconscious,” she said.

From there, she sorts them into themes and uses them as a base for a more fully finished work. She also asks the model to be her critic:

Please act as an art critic and evaluate this piece based on its subject, themes, how it makes you feel, and historic connections. Consider how it may be connected to science or math. Then, provide me with an appropriate title.

“It may not be as nuanced as a human art critic,” she said, “but it does decipher key aspects of the work which I refine further.”

She doesn’t use A.I. to create final pieces, though: “I tried — and was bored and frustrated.”

When a water system springs a leak, you might not notice until it becomes a big problem. Mr. Sutherns’s company, Digital Water Solutions, is trying to catch leaks early by placing small sensors inside fire hydrants that record the noise water makes as it flows through the pipes. That data is fed to a machine learning model that looks for certain patterns suggesting a leak.

It’s a relatively simple concept, but it’s hard to reproduce quickly, said Mr. Sutherns, in large part because every system is different: different pipe material, sizes and pressures.

“If we had to build individual machine learning models for every one of these unique systems, it would take us months, a whole bunch of data scientists,” he said.

Instead, the team uses “autonomous machine learning.” A.I. figures out, on the fly, what the parameters of the model should be for a specific system, meaning the company doesn’t need to know anything about the system ahead of time — it just has to start collecting data. Within a couple of weeks, typically, the models can provide some information on possible existing leaks.

Mr. Sutherns started the company in 2018, but recent advancements in machine learning, cheaper computing power and data storage have made the business far more feasible.

Small water systems, serving fewer than 10,000 people, make up the vast majority of water systems in the U.S., and have small budgets. Offering the technology to those systems at a reasonable price? That wouldn’t have been possible a few years ago, he said.

It’s one of A.I.’s simplest and most common use cases — one that even the A.I. engineers are leaning on: writing code. Mr. O’Sullivan is one of them: As the C.T.O. of DraftPilot, a legal A.I. company that helps lawyers with contract review, he frequently uses Anthropic’s Claude Code.

“I can give it tasks and just walk away,” he said. “It writes the code itself.”

At Mr. Valenti’s hospital, an A.I. tool, Abridge, is built into the electronic medical record system to take notes when he meets with patients. The tool listens to his conversation with the patient, then creates an organized record of the visit — the kind he would otherwise have to produce manually.

It saves him about an hour each day, he estimates, “but the real benefit is that it captures details I would have otherwise forgotten.” If a patient comes in for diabetes, but briefly mentions back pain, that aside makes it into the record whether or not he remembers it. And he’s able to focus on having a real conversation with patients, without transcribing every word.

He worries that the tool may replace human scribes. But for providers on tight budgets, it makes a difference. “For those of us in primary care who are drowning in paperwork,” he said, “this will be a plus.”

For his research in cognitive neuroscience, Mr. Morgan works with neurosurgery patients. While their brains are exposed, he runs experiments that attempt to examine how the brain encodes things like language and meaning — often by asking them questions while directly measuring their neural activity.

Because there’s usually limited time and subjects on whom he can run experiments, he has to prioritize research topics. That’s where A.I. comes in.

Like a human brain, artificial neural networks take some kind of input (words, say) and produce outputs (other words). For the human brain, what happens in the middle is something of a black box, but we know that words we hear are translated into neural activity that represents meaning, then decoded into other words. Mr. Morgan says artificial neural networks do something similar, only using numbers.

“There’s good, and growing, evidence that L.L.M.s encode syntax and words in a similar way as the brain,” Mr. Morgan said.

But unlike with a brain, you can directly examine these encoding processes in a large language model just by looking at the code. So the A.I. can act as a pseudo brain to test hypotheses about language that are hard to test in real brains.

“In my work, I figure that if I find that the middle layers of a computer model are sensitive to a particular property that I’m interested in in the brain, it’s a decent indication that the brain might care about that,” he said.