Productivity has always been a central concern of business leaders, operations researchers, and innovators. But what it means to be productive—and how we measure and achieve it—is rapidly changing in the age of artificial intelligence. The Mack Institute’s Spring 2025 Conference, Productivity in the Age of Intelligent Automation, brought together leading academics and business practitioners to explore a critical question: how can AI enhance not just efficiency, but also creativity, decision-making, and customer experience?
Productivity 2.0: What is Productivity in the Age of AI?
In his opening remarks, Mack Institute faculty co-director Prof. Christian Terwiesch drew a sharp contrast between the modest gains traditionally pursued by operations teams and the “step change” promised by AI. He noted that operational improvements in healthcare, for instance, typically target incremental optimizations of 5 to 10 percent, whereas generative AI opens the door to gains of 30 to 40 percent.
Operations professionals typically target improvements of 5-10 percent, whereas generative AI opens the door to gains of 30-40 percent.
The gains are much higher because they are driven not by cost reduction but, as Terwiesch pointed out, by a less commonly discussed benefit: value creation. Because AI can enhance quality of care and customer experience, it can ultimately increase what he calls “willingness to pay.” In other words, by making services more responsive, personalized, and “emotionally intelligent,” AI can elevate a firm’s perceived value in the eyes of its customers.
“You can use AI to increase willingness to pay while holding productivity constant, or you can hold the quality-of-care constant and focus on cutting costs and pocketing the savings,” he said. “Most companies we’ve spoken with take the latter approach of cost reduction. That’s a valid rationale, but it doesn’t have to be the only one.”
Automating Customer Service
Terwiesch’s discussion of value creation set the scene for the day’s first panel discussion on automating customer service. Chatbot technology has advanced rapidly in recent years and is now capable of handling everything from rebooking canceled flights to answering medical questions. But chatbot implementation remains difficult, especially in regulated industries like financial services or healthcare.
Vibs Abhishek, co-founder and CEO of the startup Alltius, kicked off the panel by arguing that customer service is on the verge of a radical shift, predicting that, in the next five years, most customer interactions will be “almost entirely AI-led.” The reasons are twofold: rapidly advancing technology and rising customer expectations.
In five years, customer service interactions will be “almost entirely AI-led.”
“Customer service reps are inundated with information,” he said, “and that information keeps changing, especially in the financial services sector. It’s a very difficult problem for firms to solve. To add fuel to the fire, the typical call center has a high churn rate, averaging a 30-60% turnover rate. There is simply no way these reps can become experts.”
At the same time, customers are demanding more. Abhishek noted that people see the “cool interface” of tools like ChatGPT and wonder why they’re still “stuck on the phone for 30 minutes when they call their bank.” This mismatch between modern digital experiences and outdated systems is a growing liability, but also an opportunity.
To illustrate this, he shared a case study of a large Indian brokerage that grew from 15 to 33 million customers in one year. Rather than expand its 1,000-person call center, the company turned to AI.
“Now, they’ve reduced their call center to 600 agents while handling more than twice as many customers. Seventy-three percent of customer queries are resolved automatically, and average response time has dropped by 98%,” said Abhishek. “These are true 10x improvements.”
Abhishek also explained that “making AI as empathetic as possible” can drive value. One of his clients used a bot that asked customers how they were feeling that day, a friendly gesture human reps typically avoid because they are pressured to answer as many calls as possible in a given shift. And in some cases, users even prefer AI: “There is social friction in a lot of use cases where you’re talking to another human being,” he said. “You might be more comfortable talking to an AI bot in a situation where you could feel judged.”
Danielle Corey, Head of Digital Enablement at Vanguard, shared her experience implementing generative AI in a highly regulated environment. She explained that Vanguard takes a “portfolio approach” to innovation, experimenting across multiple teams and tools with the understanding that not all will succeed.
Vanguard also tests different approaches within individual pilots. For example, one version of its call summary tool generates results immediately after the call, while another version delivers them 10–15 minutes at a lower cost.
“Balancing these AI initiatives across multiple programs is challenging,” she acknowledged. “We know that if we only try one or two projects, we’re unlikely to succeed. A portfolio approach gives us better odds, though it means accepting that many efforts may fail. The key is knowing when to cut losses and when to double down on what’s working.”
Designing Workflows with Humans in the Loop
Successful use of chatbots is about more than just building new technology: it requires designing workflows that integrate AI and human operators effectively. Our second panel focused on how “human-in-the-loop” systems can be implemented across industries, exploring how the proper balance between human and bot can be calibrated.
Parth Thaker, who leads AI Strategy and Business Development at Comcast, spoke about how the telecommunications giant embeds AI into its complex service environment. Comcast Cable serves 30 million households and millions of businesses. Parth outlined several internal use cases where human-in-the-loop design is essential: locating answers in vast procedural documents and databases to respond to customer questions; arming employees with AI tools so they can focus on more complex tasks and insights; and delivering high customer satisfaction in all interactions.
By focusing on use cases where AI augments decision-making, Comcast is designing systems that scale knowledge, reduce friction, and keep both employees and customers at the center of the experience. As Parth put it, “We want to make people’s jobs easier so they can focus on higher-value tasks and enabling best-in-class products and experiences for our customers.”
While Thaker zoomed in on large-scale service environments, Dr. Martin Bittner, Co-Founder and CEO of Redouble AI, zoomed out to examine the foundations of human-in-the-loop design. Bittner proposed three core theses: first, that AI is fundamentally different from traditional software; second, that a deep understanding of workflows must precede any meaningful AI deployment; and third, that human expert feedback is essential, not optional.
“AI is inherently different from software,” he said. “The data you train it on is only a subset of the real world, and once deployed, the AI will inevitably encounter edge cases it’s never seen before.” These edge cases lead to an element of inherent unpredictability which makes it essential to keep human feedback in the loop in order to maintain the dependability and quality of the AI process and its outputs.
AI strategy should never be designed in isolation from the business procedures it aims to improve.
Bittner’s second thesis focused on the importance of workflow understanding informing any automation efforts. AI, he argued, should never be designed in isolation from the business processes it aims to improve. “The business needs have to come first,” he said. Technical (IT/AI) teams and functional (output-focused) teams are often siloed, but both sides are critically important to successfully automate workflows in a way that creates value for the enterprise. Much of the functional teams’ knowledge and expertise is tacit and undocumented; thus, capturing and incorporating this implicit knowledge into an intelligent automation workflow is one of the most powerful steps an organization can take toward building successful AI systems.
On the third thesis—on the importance of human experts for successful AI deployment—Bittner emphasized that human-in-the-loop design does not mean humans review everything. “We shouldn’t use human creativity indiscriminately,” he said. Instead, the goal should be to identify which AI outputs are most likely to deviate from the optimal path and focus human review there. In other words, he said, we need to triage AI outputs, and involve experts where human expertise, creativity, and ingenuity can actually add value. Together, the combination of humans and AI can drive value creation.
Olympia Brikis, Head of AI Research at Siemens USA, brought the human-in-the-loop conversation into the industrial world, discussing the unique challenges of bringing Generative AI technologies to technical tasks performed by engineers and operators. In these environments, GenAI faces different challenges than in digital settings. Some of these challenges are technical—for example, the data is often proprietary, and typically not the kind that makes it into large language models. But others are more fundamental.
“Engineers think in different modalities than just language,” she explained. “Engineers think in electrification diagrams, in P&ID diagrams, in BOMs and BOBs and CAD models. These modalities cannot be put into an LLM because LLMs are language models. Yes, we have some tricks to get around this, but, fundamentally, LLMs do not natively understand these modalities.”
Despite these barriers, Brikis made the case that GenAI can play a critical role as a bridge between machines and people. At Siemens, her team is developing AI-powered “copilots” that work alongside factory operators and maintenance engineers. These tools support workers by pulling relevant system data, suggesting error resolutions, and helping them adapt to shifting production needs.
Brikis described Siemens’ roadmap as unfolding in three stages. The first is deploying simple, contextual copilots that surface relevant system data in real time. The second is building more connected systems that integrate ERP, quality, and operational data to support more holistic decision-making.
“You don’t only want to know when your machine is down,” she said. “You might also want to know, how can I maybe reprioritize my orders for that given day?”
The third stage is about identifying parts of the workflow that could be automated end-to-end. These tasks might range from triggering maintenance requests to more advanced closed-loop control systems. This stage will be powered by a new generation of Industrial Foundation Models that natively understand the language of engineers and operators—a technology Siemens is actively incubating.
Throughout her presentation, Brikis made clear that while industrial settings are more complex and less flexible than digital ones, they also present opportunities for human-in-the-loop design. Because much of the work is unstructured, variable, and grounded in physical processes, AI can’t simply replace human operators, but it can empower them with better tools, faster context, and smarter support.
Automating R&D
Innovators, whether in corporate R&D labs or startup teams, have often assumed their work was safe from automation. But recent advances in generative AI, and in particular large language models, are beginning to challenge that assumption. Our final panel of the conference asked: how good is AI at generating and implementing innovative ideas? Can it support, or even automate, parts of the R&D process?
Jared Coyle, Chief AI Officer, SAP Americas, approached the question of automating R&D with a dose of realism. “The best AI,” he said, “is boring.” Rather than showcasing grand, futuristic applications, Coyle argued that true value lies in grounding AI in specific, well-scoped business needs.
“The best AI is boring.”
“There are a lot of people spending a lot of money,” he said. “But sometimes we’ve got to step back and ask ourselves, are we making money?”
Accordingly, SAP’s approach to AI is built around embedding intelligence into workflows that already exist, especially in R&D-heavy industries like pharma, chemicals, and industrial manufacturing. One example of how a simple AI tool can assist R&D is a feature that helps scientists locate the latest version of a batch formulation in the pharmaceutical process. In another example, SAP used historical engineering data from a major agricultural equipment manufacturer to generate new design recommendations based on past tractor iterations. The goal, Coyle emphasized, isn’t to invent out of thin air, but to speed up decision-making using what companies already know.
In the end, Coyle called for a shift away from speculative uses of AI and toward practical ROI.
“One of the most common AI use cases across 34,000 organizations?” he asked. “Receipt scanning on a business trip.”
He wasn’t discouraged by that result, only insistent that AI is most effective when it solves real pain points.
“You don’t want to look at it in terms of, I’m just going to throw AI at a problem and fix it,” he said. “You have to think about the workflows people already rely on and build from there.”
Jeremy Greenberg, founder of the AI-driven research startup Crowdwave, also emphasized that successful innovation depends on identifying real use cases, not just building “flashy” tech. He told the story of a research group at MIT that spent years trying to build “Jetsons-style” robots that could do household tasks. After years of work, they created the prototype for the Roomba robot vacuum, which ultimately led to the real moneymaker: working with the U.S. military on robots that could detonate landmines on a battlefield.
“In short, cool stuff is not enough,” he said.
Greenberg’s company, Crowdwave, offers a unique way of overcoming the limits of human imagination. Crowdwave simulates human audiences to test messaging, branding, product concepts, and content in minutes instead of weeks, an approach Greenberg called “simulated primary research.” Unlike traditional surveys or focus groups, which can be slow, expensive, or logistically difficult, Crowdwave uses ensemble models and real-time data to create audience personas and gather feedback at scale. This approach has proven especially valuable in situations where conventional research would be difficult or ethically complex, such as testing messaging for late-stage cancer patients. But it also works for marketers, journalists, and product teams.
At its core, Crowdwave’s system is designed to accelerate early-stage decision-making, helping teams test faster, iterate messaging, and de-risk ideas before going to market. For Greenberg, this isn’t a replacement for human insight, but a powerful new tool to augment it.
“I’m optimistic that research is shifting toward more simulated methods,” he said. “It’s faster, cheaper, and sometimes more honest than the real thing.”
Braj Thakur, Senior Director at Ricoh USA, closed the panel with a grounded perspective on what it takes to implement AI in operational settings. Like the other speakers, Braj Thakur highlighted the gap between flashy AI narratives and day-to-day business needs.
“Everyone’s focused on AI for drug discovery or clinical trials, but those are full of variables outside your control,” he said. ”You can discover a drug early, but if the FDA still requires you to file a PDF and waits two years, you’re not shortening the development cycle.”
Braj Thakur encouraged organizations to return to the basics: clarify which business functions can be handed off to AI, establish outcome metrics, and understand where human judgment still adds the most value.
“Unless you have that clarity,” he said, “you can use all the technology and hire all the consultants you want—you’re not going to achieve the business results.”
For Braj Thakur, the most impactful applications of AI in R&D are the least glamorous: automating documentation, reducing cognitive load, and surfacing insights in real time.
“You don’t need millions of dollars,” he concluded. “You need to start with problems you can actually solve and build from there.”
Wrapping Up
Dr. Valery Yakubovich, Executive Director of the Mack Institute, closed out the conference:
“As this conference made clear, AI is redefining productivity—not merely through cost savings, but by creating genuine value through enhanced customer experiences and personalization. The most effective AI applications are those thoughtfully integrated into existing workflows, guided by clear business objectives, and supported by robust human-in-the-loop design. When organizations prioritize practical value over hype and focus on empowering human expertise rather than replacing it, AI becomes a powerful engine for sustainable innovation and long-term impact. The Mack Institute is proud to foster the collaborative environment that brings these critical insights to the forefront.”