Building Trust in AI Applications

Most people think having a human approve an AI decision means the decision is safe. It does not. 👀 There is a term for what actually happens when humans rubber stamp AI outputs under time pressure. Automation bias. It is one of the most documented and underreported risks in enterprise AI right now. After 13 years and 200+ deployments, here is what I have learned about building genuine oversight into AI systems. The human reviewing an output needs three things to actually be in the loop. They need to understand what they are reviewing. They need the context to catch what the model gets wrong. And they need to be genuinely empowered to say no without institutional pressure to simply keep moving. Most organisations have none of those three in place. They have a signature process. That is not the same thing. Before any high-stakes AI output reaches a decision point in your organisation, ask these questions. ➡️ Does the person approving this understand the underlying data well enough to catch an error? ➡️ Is there time built in for genuine review or just enough time to click approve? ➡️ What happens if someone says no? Is that genuinely supported? If the answer to any of those is no… you do not have human oversight. You have automation bias with a human signature attached. What does genuine human oversight look like in your organisation right now? #ai #leadership #futureofwork #artificialintelligence #aistrategy #teamhuman #intellectualatrophy #criticalthinking

🚨 Zero Trust for AI Agents Anthropic just released "Zero Trust for AI Agents." As we're thinking about agentic permissions, applying a Zero Trust discipline is critical to secure adoption. AI agents interpret goals, call tools, chain actions, delegate to other agents, and maintain context across sessions. The trust surface is different. The paper introduces "least agency" — a concept that OWASP has been promoting — and the distinction from least privilege is worth sitting with. 👉 "Least privilege" asks what an identity can access. 👉 "Least agency" asks what an agent can do, under what conditions, with which tools, and with what level of oversight. Autonomous agents introduce action risk alongside access risk — and the boundaries around behavior need to be architecturally enforced, not assumed. The paper includes a design test worth writing down: 🔥 Does the control make the attack impossible, or merely tedious?🔥 The practical controls follow directly from Zero Trust fundamentals — cryptographic agent identity, short-lived credentials, tool allow-listing, sandboxed execution, and full traceability from prompt to action to outcome. None of this is new doctrine. It's existing architecture applied to a harder problem. Full disclosure: the paper cites NIST SP 800-207 on Zero Trust Architecture and the CISA Zero Trust Maturity Model, both of which I co-authored during my time supporting Federal Zero Trust efforts at CISA. Zero Trust is built for a world where we have to remove implicit trust. Agentic AI is the next version of that same problem — valid identities, valid credentials, legitimate-looking actions, and still no basis for assumed trust. Access is earned. Actions are constrained. Agency must be governed. 👉🏼 Link to Anthropic's paper in the comments.

𝗡𝗮𝗶𝘃𝗲 𝗥��𝗚 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?

Trust at scale has always been the hardest thing to build in business. Word of mouth was the original mechanism. One person tells another, credibility transfers, trust builds slowly. It worked, but it was a limited mechanism you couldn't control. What's changed today is the infrastructure. Reach, repeated visibility to a large audience, is now one of the most powerful trust-building tools available to any founder or business. I am not saying being seen is the same as being trusted, but trust requires repeated exposure before it forms. The people and businesses that maintain high engagement at scale on their social media are the ones that showed up repeatedly, with a clear point of view, long before the numbers got impressive. Trust is a perception built over time through repeated signals: what you say, what you stand for, what you consistently show up for. Reach accelerates that process. Every post is another data point for your audience to evaluate whether your judgment is worth following. Enough of those data points, delivered consistently, and reach becomes evidence that you are someone worth trusting. The people and businesses who understand this aren't just building audiences. They're building credibility that makes everything else, fundraising, hiring, selling, structurally easier. #rajshamani #figuringout

🌳 Design Patterns For Building Trust. With practical guidelines for designers on how to make products — AI and non-AI — more trustworthy, reliable and honest. In the noisy and polluted world today, trust doesn’t come for free. It doesn’t emerge by default. It must be earned and meticulously preserved — by being reliable, accountable and treating customers with respect. This holds true for people but it also for software. According to Anyi Sun, there are 5 psychological foundations of user trust: 1. Reliability 🔰 The degree to which the product consistently behaves as expected. It's a sense that that the product is dependable — based on a track record of past actions. Reliability comes from promising what you do, and doing what you promised. 2. Technical competence ⚡ Perceived intelligence, sophistication and capability of the product. It's user's belief that the product can successfully perform what they are being trusted to do. It's about trusting product's capability. 3. Understandability 🧠 The extent to which users feel they can understand how the system works or why it made a certain decision. The product must be able to articulate how a decision came along, with references to fragments that underpin a decision. 4. Faith and Care 🌱 Emotional, almost "blind trust" in the product, especially when users don't understand the underlying logic. It's a belief that the trusted party actually cares about the positive outcome for you, and intends to do good. 5. Personal attachment 🌳 A sense of rapport, connection or emotional engagement with the product. Typically it emerges when a user feels that they get meaningful value from the product, and from interactions with people supporting it. Personally, I would also add the value of repeated positive experiences that build confidence in the quality of the product, and hence its reliability. --- With AI products, hitting all these psychological foundations is extremely hard. Surely some people trust AI almost instinctively, others are more critical. But people's attitude often changes dramatically once they realized that they've made severe mistakes because of AI. Recovering from it is very hard. We can help with some design patterns: 1. Avoid "Ask me anything" → push for scoping and constraints 2. Slow down users in prompting → request specific details 3. Present multiple viewpoints, explain that experts disagree 4. Allow users to manage “memory”, profiles personalization 5. Highlight what is AI-generated and what isn't (AI disclosure) 6. Allow users to override AI-generated suggestions manually 7. Allow users to tweak AI output and refine it for their needs 8. Adapt AI's tone depending on the severity of user's task Trust is why people stay or leave. It builds long-term loyalty and helps users overcome hesitation. But it must be designed and retained — across all psychological foundations and with thoughtful UX work. I think designers will be quite busy for years to come. #ux #design

Anyone can ship a chart. Trusted analysts aim for influence. Trust isn’t a vibe. It’s observable. Here are 20 signs of a data analyst you can trust 👇 1. They document their methodology transparently ↳ Every stakeholder can follow their analytical journey 2. They admit when they don’t know something ↳ “I need to investigate this further” builds more trust than guessing 3. They validate data quality before sharing insights ↳ Trust starts with clean, verified information 4. They communicate uncertainty honestly ↳ Express confidence levels and margin of error upfront 5. They follow up on previous recommendations ↳ Track whether their insights actually drove results 6. They explain their assumptions clearly ↳ Make their thinking process completely visible 7. They anticipate data limitations ↳ Proactively address what the analysis cannot prove 8. They use consistent definitions across reports ↳ Ensure metrics mean the same thing every time 9. They provide multiple scenarios when forecasting ↳ Present best case, worst case, and most likely outcomes 10. They cite their data sources religiously ↳ Full transparency on where every number originates 11. They avoid cherry-picking favorable results ↳ Present complete findings, even when inconvenient 12. They explain complex concepts in simple terms ↳ Technical accuracy doesn’t require technical jargon 13. They provide actionable next steps ↳ Never leave stakeholders wondering “what do we do now?” 14. They seek feedback and incorporate it genuinely ↳ Show they value others’ perspectives and domain expertise 15. They standardize their reporting formats ↳ Consistency reduces cognitive load for decision-makers 16. They proactively flag potential data issues ↳ Alert stakeholders to collection problems or anomalies 17. They maintain the confidentiality of sensitive data ↳ Respect data privacy and security protocols religiously 18. They provide training on how to interpret their outputs ↳ Empower others to use insights correctly 19. They collaborate with domain experts ↳ Combine analytical skills with business knowledge 20. They respond promptly to questions about their work ↳ Accessibility builds confidence in their expertise Trust isn’t about being perfect. It’s about being transparent, reliable, and genuinely committed to accuracy. Which trust-building practice do you prioritize most as a data analyst? ♻️ Repost to help your network build trusted analytics practices 🔔 Follow for daily insights on building credibility through data

New study finds legal AI’s efficiency gains erased by verification burden. A new law review paper by University of Auckland scholar Joshua Yuvaraj (https://lnkd.in/gzdEQs-k) makes a brutal point: Every minute AI saves a lawyer may be lost checking whether it got the law right. He calls this the verification-value paradox. The faster AI gets, the more time lawyers must spend verifying its work. He argues that the core rule of lawyering, which is the duty to verify, cancels out most of AI’s promised efficiency. You can’t delegate accuracy. You can only audit it. The study warns that hallucinations, hidden reasoning, and opaque logic make every AI tool an untrusted intern with infinite confidence. The takeaway: Until AI can show its work, human verification will remain the most expensive part of automation. For firms chasing efficiency through AI, that’s the paradox worth understanding and budgeting for. What do you think? Is this because it's early AI days or will this be the case going forward?

My experience of working at Urban Company taught me one key lesson about Indian consumers: Convenience may be tempting, but control wins every single time! You see, I was a product guy at UC then. - And Abhiraj (the founder and CEO) had tasked me and my colleague Sripad (now heads product at Dezerv) to improve new user conversions - In our shoes, most people would have thought of shortening the flow by selecting a few options by default, so the user would have to make fewer decisions or taps - But, we went by the approach of adding more options that the user could choose. This was because we wanted to make the user feel way more in control and in charge - Driving the calls And this worked wonders for our conversion rates. Why? Because customer trust went up massively. Thus, when launching the feature to schedule weekly bookings for our Dubai business, we actually added a step, elongating the flow. And again, we saw conversions go up! This taught me: - You can cut a user journey by two clicks, nail a sleek UI, and still see drop-offs. Why? The user didn’t feel in charge - In a country where ration shops and bank queues have taught people to expect friction, control is power. Control is trust And anything that makes the user feel that they hold the decision making, the control - IT WINS! And this shouldn’t be surprising. I’ve seen users manually enter OTPs over auto-read because typing feels safer. They skip recommendations to re-search, ensuring they’re not tricked. That’s not inefficiency - it’s defence. Ignore this, and your retention tanks. A good example is that of a fintech (I won’t name) which launched an “auto-invest” feature - It ended up driving away 20% of users who felt sidelined. But apps like PhonePe thrived with the same product with “confirm payment” prompts. - It’s pretty simple and logical. Every flow should ask: “Where does the user say ‘I’m in charge’?” - A “you can change this later” label, a manual toggle, or a “review before submit” step builds comfort Zomato’s customisable delivery instructions are one more example. These trust signals scale because they align with India’s psyche, where almost every user prefers double-checking. Thus, I now always recommend to founders in my circles, if you are building for Indian audiences, audit for control points. Add confirmations, transparent labels like “No hidden fees.” Don’t force automation - offer manual options. Test retention, not just conversion. Study PhonePe or Paytm’s “over-communicative” designs. Those extra prompts aren’t accidents - they’re trust engines. They’re not hurdles - they’re well planned and well placed handshakes. What do you think? Do share below. Best, Shiv

Question your assumptions—or reality will do it for you Imagine you’re confidently walking into a meeting, armed with your data, your vision, & your gut instincts. You’re certain you know what’s going on. Then, 5 minutes in, someone throws out a fact that obliterates your entire argument. You blink. You sweat. You consider faking a WiFi issue. But there’s no escaping it—reality just punched you in the face. That’s the problem with #assumptions. They feel true right up until the moment they spectacularly fail. & in #leadership, every assumption left unchallenged is a potential explosion waiting to happen. Your brain is an efficiency machine. It builds mental shortcuts—heuristics—that help you process information quickly. But fast isn’t always accurate. Psychologists call it confirmation bias: your brain loves to find evidence that supports what you already believe & conveniently ignores the rest. This is why companies that were once giants—think Nokia, MySpace, & Blockbuster—crashed & burned. They assumed they had it all figured out, right up until they didn’t. How to beat assumption-induced disasters? 1. Run an assumption audit Make a list of the assumptions you’re currently operating under—about your market, your customers, your competition, & your team. Then, go full detective mode: What if each of them was false? What data supports them? If you can’t prove them, they’re risks, not truths. 2. Use the ‘Reality Check’ rule For every major decision, ask: What would make me completely wrong? If you can’t answer that, you’re not questioning deeply enough. Great #leaders don’t just prepare for success—they anticipate the ways they might fail. 3. Don’t trust “We’ve always done it this way” There are 6 words that should send shivers down your spine: Because that’s how we’ve always done it. If those words show up in your strategy, your problem isn’t execution—it’s a lack of curiosity. Business graveyards are littered with companies that assumed they were safe because their old playbook used to work. 4. Hire the skeptics Yes-people are comfortable. They also keep you blind. The best leaders surround themselves with people who will challenge their thinking. Hire people who force you to back up your beliefs with evidence—not just your instincts. 5. Prototype, don’t preach Instead of assuming a new idea will work, test it. Build a small, low-risk version & let the data decide. Don’t bet the house on theories—let reality be the judge. The only way to avoid being blindsided by reality is to put your own thinking under the microscope. If you don’t make a habit of questioning your assumptions, reality will happily do it for you—& reality doesn’t do performance reviews, it does public executions. & remember, the real cost of being wrong isn’t a bad decision—because bad decisions can be undone. It’s the irreversible toll of lost time, squandered money, & shattered credibility. & as every #leader knows, those are the 3 things you can’t afford to lose.

Breast cancer can now be detected 5 years before it develops thanks to AI: Recent advances in artificial intelligence have shown remarkable potential for early breast cancer detection. AI systems are being developed that can analyze mammograms and identify potential cancer risks up to five years before clinical manifestation. These systems operate through sophisticated deep learning models trained on extensive mammogram databases, enabling them to detect subtle imaging patterns that might escape human notice. Different research teams have taken varied approaches to this challenge. For instance, scientists at MIT and Massachusetts General Hospital created a comprehensive model that examines entire mammogram images for cancer-predictive patterns. Meanwhile, Duke University researchers developed AsymMirai, which takes a more focused approach by analyzing breast tissue asymmetry between left and right breasts, achieving similar accuracy through a more streamlined and transparent method. AI is also proving valuable as a complementary tool for radiologists. The Mia system, currently being tested by Britain's National Health Service, serves as an additional layer of scrutiny, helping identify minute cancerous formations that human reviewers might miss. This capability for earlier detection can lead to more timely interventions and less aggressive treatment options.