Every organization is a mix of early adopters and technology curmudgeons, and artificial intelligence (AI) will likely be no different. However, the sheer number of ways that businesses are using the most recent iterations of computer systems that can help perform human tasks, most notably the large language models (LLMs) that have proliferated since the introduction of ChatGPT, a chatbot developed by the San Francisco-based Open AI, are an indication that the business value is real.
Granted, there might be some premature overhyping at play for technology’s shiny new toy. Microsoft founder Bill Gates called generative AI the most revolutionary technology in years, one that is on par with personal computers, the internet, and cellphones. While that remains to be seen, observations like that are among the reasons that business operators are paying attention.
According to a recent AI Trends Report from Enterprise WordPress, 41% of company AI teams are using the technology for design, illustration, and image creation, 39% are using it for social media content generation, and 38% are using it for writing copy and content creation. They are taking the technology out for a spin for both internal uses and for the benefit of clients.
For the AI industry segment, there was some unsettling drama. For some mysterious reason, OpenAI’s board of directors ousted company founder Sam Altman, only to reconsider after an employee revolt and reports that Microsoft, which is developing its own suite of products, began to recruit Altman.
Then there is the sword of Damocles that is constantly hanging overhead — the fear that the potential for harm will far exceed AI’s benefits. No other new technological innovation has produced as many doomsday scenarios, but so far the purveyors of doom are outnumbered by the people who believe we can derive benefits without the downsides.
As is the case with most new technology, the potential for mischief has not prevented early adoption. To learn more about local AI test drives, we spoke to the following Madison-area professionals: Ben Hirby, vice president of brand experience, and Elizabeth Jones, creative director, KW2 Marketing; Phil Lindemann, VP of data and analytics informatics at Epic; and attorneys and shareholders David Hanson and Michael Gentry, both of the Reinhart Boerner Van Deuren law firm.
Their advice is reflected in the way each organization has approached this evolving technology.
Tip 1: Let it bubble up
When we approached KW2 Marketing, we assumed that CEO Jennifer Savino had directed employees to explore the AI possibilities, but this is a case not of top-down development, but percolate-up exploration by staffers. Hirby and Jones serve as the primary percolators on a team of early adopters, with Savino’s blessing and her direction to draw ethical lines around the agency’s AI practices.
Hirby, who has been with KW2 for less than a year, brought with him an interest and passion for the technology — particular the generative AI produced by LLMs such as ChatGPT. He and Jones have championed its internal and external use, but their efforts are more focused on tools and practices that can actually move the needle.
“We’ve certainly brought Jennifer specific ideas, but it’s less of a ‘Hey, we need to have an initiative and we need to put our shoulders into this and make sure this is part of our future’ and a little bit more of selecting great tools to meet our clients’ needs or strategically use them to our advantage,” Hirby explains. “It’s also been done on more of a case-by-case basis and where we can best leverage this new technology.”
As Jones notes, the organic approach best suits the team. “I don’t think that you can keep creative minds away from new toys and new technology,” she states. “We’re just eager to try things as a general rule … From the top, that kind of forward-thinking is always encouraged, and from the bottom it just naturally bubbles up.”
What did come from the top was a desire for ethics and transparency. “It became from them a conversation of, ‘OK, then how do we implement this responsibly in ways that our clients can understand and embrace?’” Jones notes.
Tip 2: Illustrate ideal ideation
There are three principal ways that KW2 is using AI — for internal and client-focused projects — beginning with ideation. Hirby and Jones are part of the agency’s creative team that uses generative AI tools available to everyone or, for a small fee, uses text-based and text image-based tools to concept, to think, and to build a list. According to Hirby, it’s about trying to understand the character of a particular audience and generate appropriate ideas as though you’re using a super-powered thesaurus. In essence, he’s trying to create a proxy of the audience he intends to serve, whether it’s part of a campaign to destigmatize mental health or to provide the latest information about managing a chronic health condition.
“A writer might say, ‘How can I say this differently? How could I express this in a way that’s more colloquial?’” Hirby explains. “It’s those kinds of ideas — getting and using generative AI in that ideation process.”
According to Jones, AI has been an incredibly helpful tool for KW2 and it can be for anyone whose business is in the realm of ideas because of the heightened potential to generate unconventional solutions to problems and help partners imagine where you’re going, especially in advertising. “In those early phases of an idea, you’re relying oftentimes on close-but-not-quite-there tools such as stock photography that gives the client an idea of the kind of imagery you want to use,” she explains, “and tools like mood boards that kind of get you in the space but don’t take you all the way. AI has really given us an ability to create and help clients imagine what we want to create for them in a much more complete way.
“That journey takes a lot of trust on the client side,” she adds. “It’s a little better because now they get a clear idea of what’s in our head and we don’t have to give them ‘sort-of-likes’ anymore. We can give them a much closer and truer idea of the solutions we want to create for them.”
Tip 3: Try timeless innovation
Ideation is followed closely by experimentation to test the value of new ideas. Tools can be used to learn and experiment on your own organization before taking ideas to market. “We have a product or project that we could talk about where we’re doing exactly that — the kinds of things that we might do for clients in the future, we’re trying them on ourselves first, and we’re really excited about what’s come out of it,” Hirby says.
For an internal project, one in which entertainment media was used to recognize staff members, KW2 relied on an image-generation tool called Midjourney. It enabled members of the agency’s design team to merge existing photography with new imagery and create astonishing likenesses of employees on [trading] cards. Another tool was used for face swapping, enabling the team to put another face — a staffer’s face — on an image of a superhero. Due to potential copyright issues, it was only used internally, but it was a creative way to celebrate employees and recognize them for their contributions.
It was an example of KW2 using a tool on itself before using it to benefit clients. “This is definitely a nerd-based project that we put together,” Jones says. “We have a few members of our team who, when they were younger, were deep into Dungeons & Dragons and Magic the Gathering trading cards, and we started talking about trading cards as a classic way of presenting superheroes … We wanted all of our team members to feel like they were superheroes, and so we did use Midjourney as a starting point for capturing and creating these amazing portraits of our team members and giving them an epic tribute with a trading card.”
An old-school approach to innovation was to see that employers were dedicated a certain percentage of their working time to experimentation, but that’s not the approach taken at KW2. “That is not regulated,” Hirby says. “We’re not saying that you need to spend 10% of your day doing these kinds of things. Everyone has a certain tolerance for what they’re interested in. There are some employees who are not yet using them on a regular basis. There are others like myself who use them every day.”
Tip 4: Develop and deploy
KW2 is developing a chatbot for a Midwestern client in the municipal public health space, so it’s actively deploying its own tools for use in the market. The chatbot has yet to be released, but it’s an example of an agency developing a tool for prospective clients rather than an existing ones.
“Actually, we brought this idea to them [clients],” Hirby notes. “As we said, we have an underserved population of people who are susceptible, at-risk, for HIV-AIDS, and they are in need of further education and having resources provided to them regarding how to protect themselves, how to prevent the illness from taking place with prep and regimens or drug regimens like that. So, we’re deploying a chatbot that is going to be an expert on all things HIV-AIDS, and we’re going to serve that to them through an ad campaign that reaches them where they are on the internet and offer to answer some questions right then and there.
“This is essentially building a product that is specifically fine-tuned and trained on the data that we want to be able to provide,” he adds, “and then it’s a sort of a narrowly guided chatbot for a specific purpose.”
As Jones notes, public health is an area of focus for KW2, and it’s an area where guidance changes as new discoveries change the dynamic. Consumers will be able to ask the chatbot questions — such as how has treatment of HIV changed? — and get a clear, understandable, and conversational answer that can change people’s lives, their treatment paths, and the way they understand their treatment options.
“It’s one of the products that I’m most excited about,” Jones says. “For KW2 especially, we’ve focused on marketing in areas that fundamentally change people’s lives. We focus on health and public health education, and these are spaces where your need for accurate, transparent information that you can understand is higher than if you’re buying a car or looking for the best nail polish.”
Tip 5: Tool around
In 2023, ChatGPT emerged as the most popular AI tool for creating copy, with 39% of survey respondents using it. That was nearly twice the percentage of Chatfuel (20%), the next most popular text generator.
Obviously, the LLMs such as ChatGPT are most commonly used, but tools also have emerged for image generation (the aforementioned Midjourney), video development, search engine optimization, and other uses. Hirby describes ChatGPT as big and powerful, but he prefers Claude by Anthropic because of its constitutional AI aspects. Claude is centered on building responsible AI that’s more carefully curated with an eye toward respect for rights and responsibilities and alignment with human needs for a technology that could have an enormous impact on society.
“They call it constitutional AI, and essentially they set forth their beliefs and essentially a charter of rights and responsibilities that they believe in,” Hirby explains. “The people at Anthropic came out of OpenAI early because they felt that OpenAI wasn’t doing a good enough job with alignment with that idea of the human needs and the machine … how do we make sure that our machine is adhering to the things that we want it to and not getting away from us?”
Tip 6: Watch the bias
At KW2, the internal trading card project not only was a great bonding experience for the creative team, as members began to use these different tools and blend talents, it also was a great learning experience. Development of the trading cards gave the team a chance to learn about the inherent biases that are built into AI tools. Once they recognized bias, they also learned how to adjust and use the tools to counter it. “That’s something that anyone who is using these tools, especially on behalf of clients, needs
to understand,” Jones states.
The team encountered a lot of bias, Hirby recalls, which is another instance where the human element comes into play. “Part of our job with our clients is to recognize and hold bias as a part of the parameters of our work to make sure that we’re overcoming it, to make sure that we’re moving around it, to make sure that we’re very aware of it,” he states. “So, when you type in ‘a queen and a forest with deer,’ the first image that’s going to come back is going to be a low-cut dress, a very idealized object of a woman, not an individual, not a real person. So, we worked hard to honor the people around us and to not over sexualize them, to not make them idealized creations, but see the individual in them.”
Tip 7: Draw lines
Bias checking is part of building a set of principles and rules, finely tuned to your own organization and to clients. Another part, especially when serving public health entities, is using evidence-based information and from trusted sources such as the Centers for Disease Control and Prevention (CDC), which is what KW2 is doing with the HIV-AIDS chatbot. AI is not immune from disinformation — hallucinations, or bad AI-generated outputs, are one example — and that is one of the fears people have when they talk about how badly this technological advance could unfold.
“For our clients especially, evidence-based information is crucial,” Jones states. “I have to be incredibly accurate, as accurate as a journalist, with the things I write. When I’m providing information for our chatbot on HIV, it’s incredibly important that the information
I provide is accurate and up to date, and a lot of these AI tools now in the future, as we start to get more proprietary tools, I’m sure that they will do a much better job of scraping the internet for information. But right now, I can’t rely on ChatGPT to give me a summary of HIV treatment options that I can guarantee for my client is 100% accurate.”
Therefore, each organization that embarks on an AI journey must create a set of rules about transparency, accuracy, and how they use the tools. “Through it all, that experimentation is something that you must keep having conversations about,” Jones advises. “That knowledge sharing has been an essential part of how we’re moving forward as a company, a company that’s full of Bens and Elizabeths.”
Epic AI
For Epic, the Verona-based electronic medical records maker that is Greater Madison’s largest employer, AI is really nothing new and is almost entirely client driven, according to Phil Lindemann. The time and energy Epic spends with any technology is almost always based on the problems that its health care providing customers must solve.
“The interesting thing is that our customers have been doing AI, some of them, for over a decade now,” Lindemann explains. “It’s the new generative AI that you’re hearing about in this last year, ChatGPT for short, that sort of reinvigorated a lot of this creative work, but if you look across the Epic community, hundreds and hundreds of organizations, you see AI on a daily basis.”
However, it’s often not thought of as AI because for Epic customers, it’s used for very simple things like confirming patient appointments and determining whether a patient needs transportation to the appointment, and for more complicated administrative tasks such as optimizing how patients pass through a facility. “It’s a bit of an orchestra here,” Lindemann says. “You’re playing musical chairs, or musical beds, when trying to get patients to the intensive care unit [ICU], to surgery, to the emergency department [ED], and ultimately out the door, and there’s a lot of predictions that can go on to that.
“When a patient comes into the ED,” he adds, “you can start predicting whether they are going to be staying overnight … It’s very administrative in nature; it’s not doing any clinical prediction, but it’s just basically saying, ‘Hey, Phil’s probably going to stay overnight tonight, we should probably start figuring out which bed he’s going to go to, and make sure it has the right services.’”
When it comes to AI, that’s what Epic has been working on for the better part of eight years, but it’s only in the past year that LLMs, the generative AI concept, have evolved to the point where other business organizations can find value in them. “When you use ChatGPT, you’re now on the third or fourth version of that technology,” Lindemann explains. “We had the earlier versions in-house to try and play with them, and they just didn’t have it. It wasn’t something that would be consumer grade that we could give to clinicians.”
Epic does not use ChatGPT, but its AI repertoire includes different model types such as regressions and neural networks, depending on the use case. But as LLMs began generating business interest, Epic decided to explore its value to physicians and clinicians. Thus far, Epic has identified several existing and potential ways in which the technology can help them solve problems:
Solution 1: Save time
The most valuable thing LLMs can do for physicians is help reduce time spent on administrative functions that are not the reason they pursued medical degrees. Generative products are well suited to capture the diagnostic conversations that providers have with patients. The doctor is basically playing detective, asking questions, and trying to figure out what’s wrong so that the best course of treatment can be set in motion.
During that process, there is a new functionality called ambient listening that has been in development for several years, Lindemann explains. What ambient listening does is detect when the patient’s talking and when the caregiver is talking, so it can take all the voices in the exam room, ignore any distractions, and compile it into a synopsis of what happened in clinical terms. “That’s where people are thinking, ‘Wow, this is next level,’” Lindemann states. “This is one of the biggest innovations that’s happened in a decade — not only the ability to detect who’s talking in the room, but also how relevant it is to the actual problem.”
That summarized note gets synthesized in a form clinicians can edit and save. It’s not perfect, but it does save valuable time that can be devoted to patient interactions.
Solution 2: Save for future use
As Lindemann explains, the generative note can not only be saved for current treatment, but also for future use. “The next thing we’re looking at is taking that same concept of large blocks of text, or a large clinical chart, and summarizing it into something. Let’s say you go to the doctor and unfortunately haven’t been there in a few years, or you’re a specialist and you want to know what’s going on with the patient. It can do a ‘since I last saw you’ summary, essentially a summary of what we know and what is unique about this patient as sort of the cover sheet.”
Primary care physicians can’t be expected to memorize the hundreds or thousands of people they treat, but if they can spend a little time reviewing a summary of patient history before they go into the exam room, the treatment process can be expedited. “The ‘since I last saw you’ summary is one thing we’re working on, and then we’re focused on how we can do the same thing that we’re doing for doctors, and do it for nurses,” Lindemann says. “So, when you’re admitted to the hospital, and you’re staying overnight, at some point the nurse has to go home and a new nurse has to take over, and they do a handoff. They basically write a little narrative of what’s going on with the patient, and it could be qualitative things like they’re little grumpy today, or they’re not having the best day, and that handoff is usually synthesized from the patient’s vitals, the meds they’ve gotten that day, whether they had to go to physical therapy [PT], and that’s used for the handoff to another nurse.”
“The neat thing here is the nurse was spending time formulating that summary, and then they would go and explain it to the other nurse … Well now, we’re minimizing the amount of time they spend compiling the information and maximizing more time when the two nurses are having an opportunity to go back and forth and talk, which is ideal.”
Solution 3: Build better billing
Epic also is looking at ways it can help, from an administrative standpoint, in the billing office. There are a number of behind-the-scenes issues in the administration of insurance that Epic believes it can help with. “We’re looking at ways that AI can help look at a chart and say, ‘These are the types of things you should bill insurance for appropriately,’” Lindemann explains. “Those are in the early stages, but we are looking at those too. It’s all the different places where we can reduce administration in the delivery of health care and let them focus on the delivery of care. That’s the goal.”
Legal liabilities
The Reinhart Boerner Van Deuren law firm recently announced the formation of an Artificial Intelligence Group. It’s a new legal service area consisting of a cross-disciplinary team of attorneys that counsels client companies on the integration of AI-powered solutions and services into their businesses. The group is led by attorney David Hanson, and served by team member and attorney Michael Gentry, but it also includes people from every practice group in the firm.
In addition to counseling clients on how they can best use the technology, the group is charged with exploring AI for internal use and how to use it in serving clients because one of the core competencies of attorneys in any firm is competence in the technical tools that impact the business world and an understanding of the risks they pose. The firm has issued guidance to clients, helping them refine their AI-use policy and teach it to their employees.
The challenge with AI is that it’s different than establishing, for example, a Family and Medical Leave Act [FMLA] policy, which complies with the state and federal law and is relatively straight forward. “This is a different beast entirely,” Gentry states. “What we’re dealing with in the AI space is what technology are you trying to use? What technology are you trying to prevent your employees from using? And how do you think, given the vast array of AI products on the market right now, that you might be able to realize some efficiencies or opportunities by using these AI products that you’re not currently using?”
According to Hanson and Gentry, there are a number of ways organizations can expose themselves to more legal risk if they do not establish ethical boundaries around the technology, and they urge the following precautions:
Precaution 1: Don’t fail to police
Whether your staff uses ChatGPT, or Google’s Bard, or another generative product, where people put information into a publicly available, generative AI software that generates something new, the failure to establish specific prohibitions can result in your employees integrating confidential information, or customer information, or their own personal information into those products on company time.
Notes Gentry: “If employers are allowing that, or really not policing how their employees are going to use the product, then you could be violating confidentiality obligations. You could be compromising your trade secrets because in an intellectual property space, you have an obligation to take reasonable measures to protect those in order for them to be respected as trade secrets or recognized as trade secrets by the courts.”
In Hanson’s view, the primary areas of concern are quality control, hallucinations (inaccurate outputs), and confidentiality protection. Trade secrets, including things that could be patentable, could be at risk. “You would never let that out of out of the bag if you were using reasonable measures to protect not only trade secrets, but the proprietary, eventually patentable ideas … to me, that all goes to the public generative AI. The public generative AI tools are the ones that are most susceptible to hallucination.”
One question Hanson likes to ask is: where are your employees feeling the temptation? They have inboxes, and they’re getting messages from vendors trying to sell to them the latest and greatest AI tools. “Trust your workforce because they’re being pitched with AI tools and they may feel that they’re falling behind the competition, and they may be right,” Hanson says. “You want to know where the pressure is coming from outside the organization so you can get that up to the decision-makers to see whether that’s something you need to consider, rather than hiding from it.”
Precaution 2: Keep up
One of the things we’re all learning is that everything is accelerated in this technological space, Hanson notes. It burst onto the scene in the fall of 2022 and by halfway through 2023, a measure of AI fatigue had set in. Now, the pendulum is swinging back, the fatigue is subsiding, and a second growth spurt is occurring as “the smarter stuff is coming out,” Hanson says, “and it’s compressed. All of this has been within a year.”
One thing that happens when events occur fast and furious is that the law can be slow to respond. That’s not always a bad thing, Hanson notes, especially as attorneys and their clients discover that some of the existing laws and processes fit a new, rapidly exploding, and disruptive technology. That’s reassuring to business operators who want to see just how transformative AI can be because it means there is a legal framework to operate in.
It also means there is one less excuse to begin adopting the technology, especially when your competitors have not placed themselves in AI chains.
“We can be reassuring to folks and say, ‘Look, we know you want to compete. You’re out there in a private, for-profit business, and you want to compete and do your jobs most efficiently and realize efficiencies and drive profits,’” Hanson notes. “But we can use our model of counseling that we’ve been using for decades, and that is listen to you, ask the right questions, go with it, and have a ‘don’t fear it, it’s coming, it’s already here mentality.’
“The more you know about how folks either are already using it without your knowledge, or want to use it, the more you can enact a policy that doesn’t just throw up its hands, and I think you can safely proceed. That’s what we found to be really reassuring to clients.”
Precaution 3: Experiment safely
Reinhart’s larger clients have found that if they can build or adopt a private chatbot, they can allow their employees to use generative AI prompts, much like ChatGPT, in a “safer sandbox,” Hanson says. “Sandbox is a phrase you’re probably hearing a lot about. It’s something that’s safer to explore. We find that most clients get the message about the dangers of wide open, generative public AI.”
What’s next? From an attorney’s standpoint, the counseling answers come from knowing their client’s business, knowing their objectives, and knowing their current structure, including their level of information technology sophistication. Some of the lessons that have been learned from data privacy and cybersecurity efforts, including the need for cyber insurance, also fit the AI counseling model. “We don’t want to be a wet blanket on our clients here,” Gentry states. “We’re really trying to partner with them and realize how they can utilize these products to better their businesses.”
Human touch
Everyone we interviewed emphasized that AI is not some magic dust that you sprinkle on something. Using it effectively takes hard work because it’s a process that is not devoid of the human touch, especially when editing and shaping what’s generated from AI. “You’re trying to solve a problem,” notes Epic’s Lindemann, “and AI just happens to be one component of how to solve that problem.”
