Editor’s note: This is AI Impact, Newsweek’s weekly newsletter where each week, we will explore how business leaders are unlocking real value through artificial intelligence.
Read More on AI
Tap or click here to get this newsletter delivered to your inbox.
Signal Capture
Signals from the frontlines of AI adoption
Qualcomm’s Durga Malladi Says AI Agents Will Decide Where AI Runs
By Adam Mills
AI agents could make one of computing’s biggest infrastructure questions harder for users to see and more important for companies to solve: Where should the work run?
Durga Malladi, executive vice president and general manager of technology planning, edge solutions and data center at Qualcomm Technologies, said AI’s next phase will require coordination across device and cloud rather than a simple choice between them. A user may ask one question, point a camera at something or give an agent a broader task. Behind that interface, several models could be running at the same time. Some may run locally; others may run in the cloud.
“Tomorrow it’s the agent that’s going to be doing the work and the agent will figure out what runs on the cloud and what runs on the device,” Malladi told Newsweek. “That’s a very big change.”
Qualcomm’s recent AI push follows that logic. At Investor Day 2026, the company outlined a broader data center strategy, an agreement to acquire AI software company Modular, an expanded relationship with Hugging Face, an AI platform and developer community, and a larger push to support AI across devices, edge systems and data centers.
Malladi framed the issue as more than a data center expansion. Agentic AI, in his view, changes the way computing is organized. Instead of a user opening one app, then another, then another, an agent could become the layer that decides which models, tools and systems are needed to complete the task.
A person may never know where the work happened.
Qualcomm calls that approach hybrid AI. Malladi said the decision depends on the task, the latency required, the local context available, the cost of running the model and the sensitivity of the data rather than on a fixed preference for device or cloud.
“We’re not religious about, ‘Everything has to run on the device’ or ‘Everything has to go on the cloud,’” he said.
Robotics shows why the answer can vary. A robot that has to perceive its environment, plan its next move and act cannot always wait for a round trip to the cloud. Advanced driver-assistance systems face a similar constraint. Other work may be better suited for larger models running elsewhere, especially when the task is less sensitive to delay or requires more compute than a local device can handle.
Agents add another layer because they are not just returning answers. They can coordinate multiple steps, pull in context and route work across different systems. Malladi said that changes what the underlying hardware and software have to support.
“That agent is this master orchestrator,” he said.
Behind the scenes, the processor mix changes too. Agentic AI may depend on several kinds of processors working together, including CPUs, GPUs and NPUs. Malladi said CPUs become more important in that environment because agents have to handle the coordination layer around the models.
Moving deeper into data centers extends Qualcomm’s role from devices toward the cloud infrastructure running larger AI workloads. The company built its reputation in mobile and edge devices, where power consumption, battery life and performance limits are constant constraints. Malladi said that history matters as inference becomes a larger share of AI computing and companies look for ways to run models more efficiently.
Software is becoming just as important. Qualcomm announced in June that it had reached an agreement to acquire Modular, whose software is designed to help AI run across different hardware architectures. Such tools are central to Qualcomm’s push to make its AI platforms easier for developers to use, from edge devices to cloud infrastructure.
To Malladi, portability matters because developers and customers should not have to rebuild applications for each processor, accelerator or deployment environment.
“As a developer, I don’t have to do five different things,” he said. “I just do one thing and it works on everything.”
Qualcomm’s relationship with Hugging Face points in the same direction. The platform gives developers access to more than 3 million open models. Qualcomm and Hugging Face plan to work on agentic AI model onboarding and hybrid orchestration across Qualcomm-powered devices and data center systems.
Malladi described a workflow in which a developer or agent could choose a model, decide where it should run and send different parts of the task across devices or larger systems without turning every deployment into a separate integration project.
The experience Malladi described feels less app-based and more task-based.
Once agents mediate more of that work, the infrastructure decision may become invisible to the user. Qualcomm is trying to build across the layers that would make that possible, from the device in a person’s hand to the data center infrastructure running heavier AI workloads.
Asked what would show that idea moving from roadmap to everyday reality, Malladi pointed to adoption across ordinary devices.
“The proof point that you will start seeing is … more agentic workflow in all kinds of devices around you,” he said.
Core Intelligence
AI Benchmarks Need a Real-World Test
By Adam Mills
AI benchmarks become less useful the closer a buyer gets to the work a model is supposed to do.
Pearl Enterprise, which builds AI systems for professional services, tested leading models against expert-authored answers across business, health, law, pets and technology. The top model, OpenAI’s GPT-5.5, reached 72.7 percent expert alignment overall, according to Pearl’s leaderboard.
A buyer choosing a system should read that number with caution. Pearl’s results showed the same model performing differently by domain, with GPT-5.5 reaching 80.9 percent expert alignment in business, 68.8 percent in health and 62.1 percent in pets. A single score can make a model look broadly capable while hiding the place where it may fail a company’s actual use case.
“A C- student who knows they’re a C- student is manageable,” Andy Kurtzig, CEO of Pearl, told Newsweek. “A C- student who’s convinced they’re an A+ student, and answers every question with total confidence, is the one who gets a company in trouble.”
A leaderboard can help companies build a shortlist. It cannot show how a system will perform once it is used for claims review, contract analysis, security triage, customer support, internal search or coding. Those workflows come with company data, permissions, audit requirements and business consequences that do not appear in an abstract benchmark.
Rob Clark, president of Seekr, an AI company focused on model evaluation, said rankings can help compare systems, but they cannot decide whether a tool is ready for a specific environment.
“A leaderboard ranking is not a trust decision,” Clark told Newsweek. “Confidence scores measure how sure the model is, not how right it is.”
Enterprise buyers need more than proof that a model can perform a task under clean test conditions. Varun Badhwar, founder and CEO of Endor Labs, a software security company, said the real risk appears once a model is connected to company data, embedded in a product, pointed at a workflow or allowed to use tools.
“You are not deploying the model,” he said. “You’re deploying the model, inside a harness, pointed at your data, under your constraints.”
Review has to become more specific once AI moves into real work. Human oversight should rise with the cost of failure. An internal draft memo does not need the same control as an AI system affecting a payment, security decision, legal right, health recommendation or compliance record.
Patrick Sullivan, vice president of strategy and innovation at A-LIGN, a cybersecurity compliance company, said professional settings require evidence that can survive scrutiny later.
“The right question isn’t ‘How well does this model perform?’” Sullivan said. “It’s ‘Can I produce evidence that it performed correctly, on my work, for the use case I deployed it against?’”
The fastest companies may be the ones that know exactly where a system works, where a reviewer has to stay close and where the model should never act alone.
Clark framed testing as a way to make deployment faster, not slower.
“Speed is the payoff for testing the model against the job it’ll actually do, in the environment it’ll actually run, before you ship it—not after,” he said.
You can read the full article here: Top AI Models Might Be Confident—Doesn’t Mean They’re Right.
Upcoming Webinars
Humanizing AI: Personalization, Engagement and the Customer Experience
AI is giving companies new ways to personalize customer interactions, respond across more channels and make digital experiences more relevant. The risk is that faster, more automated engagement can also feel less personal if companies lose sight of trust, transparency and the human judgment customers still expect.
In an upcoming Newsweek webinar presented by Cognizant, Gabriel Snyder, Newsweek’s executive editor, enterprise, will lead a discussion on how companies are using AI across service, support, product experience and customer engagement. Nick Mehta, entrepreneur in residence at Bessemer Venture Partners and former CEO of Gainsight, will join the conversation, along with an executive from Zendesk and additional guests to be announced.
Join the live discussion on Thursday, July 30, at 2 p.m. Eastern. Register for free.
Is India on the Right Side of the AI Trade?
India’s role in the global technology economy has long been shaped by scale: a massive IT services sector, fast-growing capital markets, digital public infrastructure and a young workforce. AI could raise the stakes even further by changing where value is created, who captures it and how countries compete.
In an upcoming “AI Impact Forum” session, Dr. Ranjit Tinaikar speaks with Shri Ashishkumar Chauhan, CEO of the National Stock Exchange of India, about whether India is positioned to benefit from the AI revolution—and what the technology could mean for the country’s economy, financial markets, market infrastructure and next stage of growth.
Join the live discussion on Thursday, July 23, at 10 a.m. Eastern. Register for free.
Prompt Injection
What’s one recent insight you’ve learned about AI?
“The realization that has stuck with me is that AI in healthcare is fundamentally an access story.
Too often, the conversation focuses on how much AI we’re deploying. I think the better question is whether we’re making it easier for patients to get the care they need, when they need it.
Most of the friction patients experience happens long before they reach the exam room. It’s the scheduling call they must make, the question that turns into 20 minutes on hold, or the reminder that arrives too late to be useful. That’s where access breaks down.
AI’s biggest near-term value is helping remove the coordination burden that sits between patients and their care. When you can connect outreach, responses, and service interactions into a single experience, you reduce friction for patients and free up staff to focus on more meaningful work.
The future of healthcare won’t be defined by how much AI we deploy. It will be defined by whether patients can get answers faster, access care more easily, and spend less time navigating the system. If AI helps accomplish that, healthcare will feel more human, not less.”
Have your own lesson to share? Email us at: a.mills@newsweek.com
Run Log
AI use case of the week
Machinify Uses AI to Speed Up Claims Review
By Adam Mills
A hospital stay can end with a record large enough to become its own administrative problem.
A patient treated for several serious conditions may leave behind thousands of pages of notes, orders and documentation. Before a payer can decide whether the codes on a complex claim match the care that was delivered, someone has to read through the record and determine what actually happened.
Dr. Darshak Sanghavi, chief medical officer at Machinify, a company that uses AI to automate health care claims administration, said that review has traditionally required hours of work from highly trained coders, physicians or other experts.
Machinify used AI to read the full medical record and extract the codes needed for claims review. By turning a large, messy chart into a structured starting point, the system gives reviewers a faster way to check the claim against the underlying medical record.
According to Sanghavi, work that once required four to six hours of chart review can now take four or five minutes. He described that work as administrative waste because it happens after the patient’s care is already complete.
“It’s like performing an autopsy on a clinical record,” Sanghavi said.
Machinify is not making clinical decisions or changing what care was provided. It is targeting the work that comes later: reconstructing care from a chart so reviewers can decide whether the claim is accurate without searching through thousands of pages themselves.
Have an interesting AI use case to share with us? Email us at: a.mills@newsweek.com
Context Window
■ AI is putting new pressure on middle management as companies flatten organizations, but workplace experts told Newsweek that businesses still need leaders who can build trust, explain change and connect teams to technology decisions. [Newsweek]
■ Allianz Commercial says confirmed AI-related claims remain limited, but AI has risen to No. 2 in its global risk ranking and could amplify losses across cyber, liability, D&O and fraud as adoption scales. [Allianz]
■ Illinois Gov. JB Pritzker signed a law requiring large AI developers to publish catastrophic-risk plans, report critical safety incidents and undergo annual third-party audits, making Illinois the first state to mandate independent AI safety reviews. [Chicago Tribune]
■ Anthropic plans to lease a 16-story building in Lower Manhattan and double its New York City workforce to 1,000 this year, as AI firms expand closer to enterprise customers in finance, health care, law, consulting and media. [The New York Times]
■ OpenAI launched GPT-5.6 across ChatGPT, Codex and its API, positioning the new model family around stronger performance per dollar, multi-agent workflows, coding, knowledge work, cybersecurity and science. [OpenAI]
Transfer Protocol
Tracking executive moves across the AI landscape
Claude Alexandre, previously vice president of digital media, B2B product and campaign marketing at Adobe, has joined airSlate as president, where he will help scale operations as the company expands its AI-powered document and workflow automation products.
Alan Stukalsky, most recently chief digital and information officer at LHH, has been appointed chief product and technology officer at Kelly, overseeing product strategy, technology capabilities and digital innovation as the workforce company scales AI across its operations.
Bhavesh Dayalji has moved into a new role as head of Kensho Data & Intelligence at S&P Global Market Intelligence, leading a client-facing data product and AI delivery layer as the company reorganizes around AI-powered tools, workflows and experiences.
Tres Seippel, previously an executive at Arbor Realty Trust and Genesis Capital, has been named chief innovation officer at TrustPoint AI, where he will lead data intelligence products and AI-driven workflow automation across construction, renovation, bridge and land development lending.
Keith Soura, previously an engineering leader at Better.com and chief technology officer at VERO, has been appointed chief technology officer at Arch, overseeing architecture, engineering and product delivery for the company’s AI-powered private markets platform.
Know someone on the move in AI? Send job change info to a.mills@newsweek.com
Magic Moment
What’s the most fun or unexpected way you’ve used AI lately?
“The Bedford team has used an AI-generated podcast experience to help students develop their voice and key artifacts to save for the future. When the learning process is thoughtful it can enrich the context AI employs to individualize support.”
Experience some AI magic? Tell us about it at a.mills@newsweek.com
Tap or click here to get this newsletter delivered to your inbox.