AI Assistants In The Workplace

In our new paper we ran an experiment at Procter and Gamble with 776 experienced professionals solving real business problems. We found that individuals randomly assiged to use AI did as well as a team of two without AI. And AI-augmented teams produced more exceptional solutions. The teams using AI were happier as well. Even more interesting: AI broke down professional silos. R&D people with AI produced more commercial work and commercial people with AI had more technical solutions. The standard model of "AI as productivity tool" may be too limiting. Today’s AI can function as a kind of teammate, offering better performance, expertise sharing, and even positive emotional experiences. This was a massive team effort with work led by Fabrizio Dell'Acqua, Charles Ayoubi, and Karim Lakhani along with Hila Lifshitz, Raffaella Sadun, Lilach M., me and our partners at P&G: Yi Han, Jeff Goldman, Hari Nair and Stewart Taub Subatack about the work here: https://lnkd.in/ehJr8CxM Paper: https://lnkd.in/e-ZGZmW9

The Voice Stack is improving rapidly. Systems that interact with users via speaking and listening will drive many new applications. Over the past year, I’ve been working closely with DeepLearning.AI, AI Fund, and several collaborators on voice-based applications, and I will share best practices I’ve learned in this and future posts. Foundation models that are trained to directly input, and often also directly generate, audio have contributed to this growth, but they are only part of the story. OpenAI’s RealTime API makes it easy for developers to write prompts to develop systems that deliver voice-in, voice-out experiences. This is great for building quick-and-dirty prototypes, and it also works well for low-stakes conversations where making an occasional mistake is okay. I encourage you to try it! However, compared to text-based generation, it is still hard to control the output of voice-in voice-out models. In contrast to directly generating audio, when we use an LLM to generate text, we have many tools for building guardrails, and we can double-check the output before showing it to users. We can also use sophisticated agentic reasoning workflows to compute high-quality outputs. Before a customer-service agent shows a user the message, “Sure, I’m happy to issue a refund,” we can make sure that (i) issuing the refund is consistent with our business policy and (ii) we will call the API to issue the refund (and not just promise a refund without issuing it). In contrast, the tools to prevent a voice-in, voice-out model from making such mistakes are much less mature. In my experience, the reasoning capability of voice models also seems inferior to text-based models, and they give less sophisticated answers. (Perhaps this is because voice responses have to be more brief, leaving less room for chain-of-thought reasoning to get to a more thoughtful answer.) When building applications where I need a more control over the output, I use agentic workflows to reason at length about the user’s input. In voice applications, this means I end up using a pipeline that includes speech-to-text (STT) to transcribe the user’s words, then processes the text using one or more LLM calls, and finally returns an audio response to the user via TTS (text-to-speech). This, where the reasoning is done in text, allows for more accurate responses. However, this process introduces latency, and users of voice applications are very sensitive to latency. When DeepLearning.AI worked with RealAvatar (an AI Fund portfolio company led by Jeff Daniel) to build an avatar of me, we found that getting TTS to generate a voice that sounded like me was not very hard, but getting it to respond to questions using words similar to those I would choose was. Even after much tuning, it remains a work in progress. You can play with it at https://lnkd.in/gcZ66yGM [At length limit. Full text, including latency reduction technique: https://lnkd.in/gjzjiVwx ]

As AI drives massive productivity gains, businesses may consider cutting back on their human workforce to boost efficiency—but I don’t believe that’s the right choice. The real promise of AI isn’t in replacing humans—it’s in amplifying their potential. For centuries, humanity’s greatest leaps forward were not achieved by replacing humans completely with new tools we’ve built but instead by using the tools to accelerate human agency and potential. The automobile amplified human movement. The computer amplified human creation.  And today, AI is amplifying human intelligence. There are certain jobs—especially those involving repetitive, robot-like tasks—that AI will transform. After all, robots will always be better robots than humans. But humans thrive when they’re empowered to be better humans. With AI, people will unlock new skills and deepen natural talents –– achieving what I call “superagency.” Ultimately, the surge in productivity will guide business leaders to a realization: the right move isn’t to do the same work with fewer people but to create even greater value by leveraging more employees with new AI-driven superpowers. Our goal should be clear. Build technology that works with us, not for us—tools that extend what it means to be human, not replace it. Because when we amplify human ingenuity, the possibilities are infinite. 

I tried EVERY major AI Coding tool so you don’t have to. Here’s what I learned about each one - and which one’s the best for your particular use case 👇 After an entire weekend of hands-on testing 15+ AI coding assistants, building the same real-life application (tax comparison calculator), and documenting every step - here's the comprehensive breakdown to separate the signal from the noise: 🏆 Best Overall: Cline - 100% open source and free version of Cursor + Windsurf that’s a simple VS Code extension - Truly thoughtful agentic coding with extensive tool use (terminal, computer use, websites, etc) - Wrote the best code with fewer mistakes, better self-healing, but no inline chat 🎨 Best for Non-Technical Users: Vercel V0 - Fast, Easy, intuitive UX - Strong community and templates - Component-specific editing via AI is magical ⚡Best for Quick Prototypes: Anthropic Claude 3.5 Sonnet - Fast & clean responses - Great reasoning & logic clarity - Artifact is great for prototyping, with ability to publish and share Replit: Good for full-stack cloud development, but sits in an awkward spot—too complex for beginners, too constrained for advanced users. StackBlitz Bolt.new: A standard cloud IDE with AI codegen, but nothing special. Lovable: Similar to Bolt, but unreliable AI-generated code, hard to toggle/see code. Cursor: Great Copilot alternative, but lacks extensive agentic capabilities like Cline. Codeium Windsurf: Strong agent mode but agent was sometimes lazy and incomplete. GitHub Copilot: Good for simple inline edits, but lacks full agentic workflow (though an agent mode was recently released). Aider: Terminal & keyboard only. Feels like Vim/Emacs on steroids. Too hardcore. OpenHands: Open-source and free Cognition Devin with strong agentic coding, but SaaS version is unstable. OpenAI (o3-mini-high): Good logic depth but lacks a coding canvas. Anthropic (Claude 3.5 Sonnet): Fast + clean. Artifact is great for prototypes, but can’t edit code directly inside it. Google Gemini 2: Poor experience—lazy, incomplete code. Generated separate files that I had to manually combine. DeepSeek AI R1: Strong long reasoning chains, but gets a lot of logic wrong. Tempo (YC S23): Promising PRD → Design → Code → Deploy workflow, but still in early stages. Onlook: Strong for design-first workflows but inconvenient for direct code editing. Reweb: Generates only UI components, not code with logic. My Final Recommendations: - For non-technical users: Vercel V0 is the best no-code/low-code option. - For cloud-based development: Try Bolt. - For local AI-powered coding: Cline is free and outperforms Cursor/Codeium. - For rapid prototyping: Claude 3.5 Sonnet is fast and effective. - For designers: Tempo or Onlook provide a strong UI-first workflow. Do you want to see a full write up of my AI coding experiences? Let me know if I should make a full post comparing AI Coding tools in detail by sharing this post and commenting below.

Excellent new paper "Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce" from Stanford University researchers including Erik Brynjolfsson. The paper is based around a Human Agency Scale (HAS) to quantify automation vs augmentation in work. Some really powerful insights in there: 🧹 Workers want AI to handle the boring stuff. 46.1% of tasks received positive ratings for automation from workers, who mostly want AI to take over low-value, repetitive, or tedious duties. The top reason, cited in 69.4% of these cases, was to free up time for high-value work, not to replace jobs entirely. 🚦 Most investment misses the mark on what workers want. Despite high demand for AI automation in certain tasks, 41.0% of current AI startup investments (e.g., from Y Combinator) focus on tasks that workers don’t want automated. Meanwhile, many highly desired tasks with feasible AI capability remain underfunded, showing a clear mismatch between R&D priorities and workforce needs. 🤝 Human-agent collaboration is the sweet spot. The most common preference across occupations was HAS Level 3—equal partnership—selected by 45.2% of workers. This highlights the value workers place on retaining involvement, with AI as a collaborator rather than a replacement. 🧠 Workers fear job loss but trust issues loom larger. Among the 28% of workers who voiced concerns about AI, the top issue (45%) was lack of trust in AI’s accuracy and reliability—greater than fear of job loss (23%). Qualitative data suggest workers want control, creativity, and decision-making to remain human-led. 🔍 AI doesn’t align with actual user needs—yet. Occupations with the highest desire for automation make up just 1.26% of LLM usage data from tools like Claude.ai. This points to a major gap between what workers actually want help with and what AI is currently being used for. 🎯 The Human Agency Scale reveals capability gaps. Only 26.9% of tasks had matching human-desired and AI-assessed levels of required human involvement. In 47.5% of cases, workers preferred more involvement than AI experts deemed necessary, signaling friction and a need for more user-aligned AI design. 📉 Data tasks are losing steam, people skills are rising. Tasks like “analyzing data” and “updating knowledge” ranked high in wages but low in required human agency. In contrast, interpersonal and organizational skills, such as “guiding others” and “monitoring resources,” scored high in human agency, hinting at a shift in what skills will matter most. 🛠️ Workers want customizable AI agents. A significant 23.1% of workers envision AI systems tailored to specific roles or routines, while 23.0% want general-purpose assistants. Only 16.5% advocate full automation, emphasizing the demand for adaptable and user-defined AI collaboration tools. More of the best in Humans + AI research coming...

Microsoft just released a 35-page report on medical AI - and it’s a reality check for healthcare. The paper, “The Illusion of Readiness”, tested six of the most popular models (OpenAI, Gemini, etc)… across six multimodal medical benchmarks. And the verdict? The models scored high on medical exams. But they’re not even close to being real-world ready. Here’s what the stress tests revealed: ▶ 1. Shortcut learning Models often answered correctly even when key information, like medical images, was removed. They weren’t reasoning - they were exploiting statistical shortcuts. That means benchmark wins may hide shallow understanding. ▶ 2. Fragile under small changes Making small tweaks caused big swings in predictions. This fragility shows how unreliable model reasoning becomes under stress. In visual substitution tests, accuracy dropped from 83% to 52% when images were swapped - exposing shallow visual–answer pairings. ▶ 3. Fabricated reasoning Models produced confident, step-by-step medical explanations - but many were medically unsound… or entirely fabricated. Convincing to the eye, dangerous in practice. And more importantly, healthcare isn’t a multiple-choice exam. It’s uncertainty, incomplete data, and high stakes. So Microsoft’s team calls for new standards: - Stress tests that expose fragility - Clinician-guided guidelines that profile benchmarks - Evaluation of robustness and trustworthiness - not just leaderboard scores The takeaway is simple: Medical AI may ace tests today. But until it proves reliable under stress, it’s not ready for the clinic. When do you think popular LLMs will be clinic-ready? #entrepreneurship #healthtech #AI

Oracle rolled out a major agentic AI release. It introduced 𝐎𝐫𝐚𝐜𝐥𝐞 𝐅𝐮𝐬𝐢𝐨𝐧 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐀𝐩𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬  — a new class of enterprise applications designed to help AI agents work inside real business workflows, not around them. A few details worth adding beyond what I covered in the video: 𝐅𝐨𝐫 𝐅𝐢𝐧𝐚𝐧𝐜𝐞 𝐚𝐧𝐝 𝐒𝐮𝐩𝐩𝐥𝐲 𝐂𝐡𝐚𝐢𝐧 — helping teams move from manual follow-ups and fragmented handoffs to more proactive execution across collections, claims settlement, cost accounting close, sourcing, logistics, warehouse operations, and sales order exceptions. 𝐅𝐨𝐫 𝐇𝐑 — focused on workflows like hiring, workforce operations, manager support, employee help, team learning, career advancement, and talent review. 𝐅𝐨𝐫 𝐂𝐮𝐬𝐭𝐨𝐦𝐞𝐫 𝐄𝐱𝐩𝐞𝐫𝐢𝐞𝐧𝐜𝐞 — built for sales, service, and marketing processes, including sales command centers, service management, cross-sell programs, marketing workflows, and contract compliance. 𝐎𝐫𝐚𝐜𝐥𝐞 𝐀𝐈 𝐀𝐠𝐞𝐧𝐭 𝐒𝐭𝐮𝐝𝐢𝐨 + 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐀𝐩𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬 𝐁𝐮𝐢𝐥𝐝𝐞𝐫 — gives companies a way to build, customize, connect, and run their own agentic workflows using Oracle, partner, and external agents. The broader point is simple: Enterprise AI becomes much more useful when it understands the systems, rules, approvals, and workflows it is expected to operate inside. 📍Full release here: https://lnkd.in/gwD7kvJd Oracle AI Database #oraclepartner

🤖 Should AI Agents Be Named, Trained, and Onboarded Like Employees? We’re at a tipping point. AI agents aren’t just tools anymore—they’re performing work that entire teams once handled. They: → Draft reports, process transactions, and even manage projects → Are integrated into company workflows, CRMs, and knowledge bases → Can work 24/7 without breaks, pay, or vacations So here’s the question: If we trust them with the responsibilities of teammates, shouldn’t we treat them like teammates? 📌 The case for treating AI agents like employees 1️⃣ Naming creates connection & accountability Calling your AI “Ava” or “Max” changes the dynamic. Studies in human-computer interaction show that people trust, remember, and collaborate better with systems that have human-like identifiers (Stanford HAI). It’s the difference between “run the bot” and “ask Ava to draft the proposal.” 2️⃣ Training ensures alignment Unlike humans, AI doesn’t absorb company culture or mission unless explicitly taught. Feeding agents your style guides, policies, and datasets during onboarding keeps outputs consistent. (TechRadar). 3️⃣ Onboarding speeds agent integration Just like new hires, agents need staged exposure to systems, tasks, and rules. Phased onboarding—starting with shadowing, then supervised execution, then autonomy—reduces early mistakes and speeds time-to-value. 4️⃣ Scaling without model drift Structured onboarding isn’t just for day one—it’s ongoing. Refreshing training data and periodically re-running onboarding workflows ensures your agents stay accurate and compliant as your business changes. ⚠️ But here’s the danger → They’re not human. AI can’t pick up unspoken cues, office politics, or emotional context. A perfectly “logical” answer may be a cultural disaster. → Over-familiarity can breed over-trust. When an AI has a name, we’re more likely to give it more autonomy than it should have. → Guardrails are non-negotiable. Without clear boundaries and monitoring, AI agents can misinterpret instructions, expose sensitive data, or act on flawed logic (RelevanceAI). 💬 My take: If we’re inviting AI agents into our workflows like colleagues, we should: → Name them — to foster collaboration → Train them — to align with our vision → Onboard them — to understand our objectives → Audit them — to maintain trust and safety But let’s never forget: They don’t feel loyalty. They don’t share values. They execute. And execution without human judgment is just… automation. Question for you: If we treat AI agents like employees, will we make them more effective… or just more dangerous?

𝗡𝗮𝗶𝘃𝗲 𝗥𝗔𝗚 𝘄𝗼𝗿𝗸𝘀 𝗶𝗻 𝗮 𝗱𝗲𝗺𝗼. 𝗜𝘁 𝗳𝗮𝗶𝗹𝘀 𝘁𝗵𝗲 𝗺𝗼𝗺𝗲𝗻𝘁 𝗿𝗲𝗮𝗹 𝘂𝘀𝗲𝗿𝘀 𝘀𝗵𝗼𝘄 𝘂𝗽. Embed → retrieve → generate looks clean in a notebook. Real requirements break it: → Questions whose answer is spread across many documents → Industry terms that embeddings get wrong → Bad chunks the pipeline never catches → Answers that live in how things connect, not in any single chunk → PDFs full of tables and images a text-only index cannot read These 5 architectures are how serious teams stay ahead in the agentic AI era: 𝟬𝟭 𝗛𝘆𝗯𝗿𝗶𝗱 𝗥𝗔𝗚 → Dense vectors find meaning. BM25 finds exact words. → Reciprocal Rank Fusion combines both ranked lists. → A safe baseline for almost every team. 𝟬𝟮 𝗚𝗿𝗮𝗽𝗵𝗥𝗔𝗚 → Pull entities and their relationships into a knowledge graph. → Retrieve subgraphs and community summaries, not chunks. → Best when the answer lives in how things connect. 𝟬𝟯 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗔𝗚 → A planner agent picks the right tool: vector, web, or SQL. → A reasoner agent keeps trying until the answer is solid. → Retrieval becomes a plan, not a single step. 𝟬𝟰 𝗖𝗼𝗿𝗿𝗲𝗰𝘁𝗶𝘃𝗲 𝗥𝗔𝗚 (𝗖𝗥𝗔𝗚) → Grade every retrieval before you trust it. → Correct → answer. Unclear → rewrite the query. Wrong → search the web. → This is what production RAG actually looks like. 𝟬𝟱 𝗠𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗥𝗔𝗚 → One embedding model (CLIP, ColPali) for text, images, and tables. → One vector index. One multimodal LLM. → No more separate pipelines for PDFs with charts. I built a runnable example for each of the five patterns. GitHub link in the first comment. The best teams in 2026 do not pick one. They combine them — hybrid retrieval inside an agentic loop, with a corrective grader, over a multimodal index. Naive RAG is a starting point, not a finish line. That is why most enterprise GenAI projects stall at the demo. Which of these five becomes the default RAG stack in the next 18 months — and which stays a specialized tool?

The anatomy of a sales call has changed dramatically. Last week, I shadowed some of HubSpot’s top reps and what struck me was how differently the best sellers work today. They’re using AI at every stage: before, during, and after the call. And the results are real. The brain: before the call. AI does the heavy research — scanning 10Ks, news, emails, and past calls to surface the insights that matter most. Tools like Breeze Assistant can prep a full company overview in seconds. According to our State of Sales Report, 74% of sellers say buyers are showing up to calls more informed than ever before. Salespeople need to be just as ready. The heart: during the call. AI notetakers capture everything: next steps, budget mentions, open questions, so reps can focus on listening, not typing or scribbling notes on the side.  Also, AI assistants surface the right case study or testimonial in real time, making every answer sharper and every example more relevant. That means as a sales rep you are more engaged and relevant. The muscle: after the call. AI follows through fast. It drafts personalized follow-up emails in your own voice, outlines next steps, and flags what needs attention. More time with customers and less time writing emails. The result: sellers who prepare better, connect deeper, and close faster. The anatomy of a great sales call used to be manual effort and hustle. Now, it’s human connection powered by intelligence.