Peter Tanner

Today is AI Awareness Day, and what better way to celebrate than to learn more about American sentiment around AI? With that, I'm incredibly excited to share the first annual AI Insights Report by TD.   The report shows that we are at a turning point, with artificial intelligence (AI) increasingly integrating into daily life, but that building trust and transparency is vital to fostering comfort around AI usage.   Check out the full report for a deep dive into The AI Trust Gap, AI in the Workplace, and exploring The Path Forward.    #AIAwarenessDay #ProudtoBeTD #ArtificialIntelligence #Innovation #TrustInAI #DigitalTransformation #TechTrends https://lnkd.in/djgWp2Br

97

3 Comments

What a big day for #rbc…this is our double down on AI. How do we improve on #1 in Canada and #3 globally?…We build an org that is focused on delivering value through AI across the enterprise…but do it faster. I’m honoured to be part of it and thrilled by the talent I will get to work with. #rbcai #borealis #lumina

183

1 Comment

The missing piece is trust.   That's one of the top three insights we uncovered in our first-ever 2025 TD AI Insights Report. While Canadians are open to using AI, trust in it as a source of reliable, honest information is still low, coming in at only 43%.   At TD, we believe that if people don't understand how a technology works or how it helps them, they're less likely to trust it. And having that trust is critical, especially when AI plays a growing role in how we live and bank.   That's why it's our responsibility to make AI understandable, secure, and grounded in the needs of our customers. Building trust doesn't happen automatically, it's earned through transparency, accountability, and a commitment to customer-first innovation.   Launching the 2025 TD AI Insights Report is a meaningful step toward understanding how Canadians are engaging with AI and what it will take to build confidence in the tools shaping our future. Explore the full survey results here: https://lnkd.in/gqr9xTjq

186

7 Comments

It was a pleasure to be part of the panel discussion on AI in banking at Toronto Machine Learning Summit (TMLS) yesterday. One of the questions posed to the panel was to describe AI in banking using a movie title. My choice (with a little help from ChatGPT) was Edge of Tomorrow. Much like in the movie, we're at the beginning of a massive transformation. The full impact of AI on banking is still unfolding, similar to how it was difficult to predict the full impact of mobile technology back in the era of the PalmPilot. During the discussion, I highlighted a few key areas we're focused on at Scotiabank: * Evolving Skillsets: The role of an AI practitioner is changing. It's no longer just about building models. Strong software development skills and a deep understanding of cloud platforms are becoming essential as we embed AI directly into our core processes. * Delivering Tangible Value: We're moving beyond experimentation to real-world application. A great example is an agentic AI system we recently deployed. This system reviews and sorts thousands of daily emails from our commercial clients. It now processes 70% of these emails with higher accuracy than previous manual methods. * Improving Customer Experience: The true benefit of this AI system isn't just internal efficiency. By instantly sorting and creating cases, we've cut down the response-initiation time by 1-2 hours. This delivers a faster, better service to our clients, which is the ultimate goal. It was a fantastic conversation with my fellow panelists Ozge Yeloglu, Patricia Arocena, PhD, Saba Zuberi and Eric Morrow.

334

13 Comments

Dynamically generated, Ontology-powered fleets of agents. At DevCon 5, Palantir Product Manager Ben Radford walks through the architecture of AI Hivemind — covering the capabilities and tradecraft that enable the creation, orchestration, and optimization of specialized agent teams.

436

12 Comments

Thought Leadership driving enterprise upskilling on AI & Data in the 2025 Sustainability Report At CIBC, we believe that responsible innovation and continuous learning are essential to building a more sustainable future. The 2025 Sustainability Report:https://lnkd.in/eQSJAU6a highlights progress against our ESG initiatives across the bank. Delivered by our Enterprise Innovation team – the Future Now thought leadership program has become a cornerstone for driving knowledge and adoption of emerging technologies across the bank for the last 5 years and I’m proud to share our feature once again. In 2025, more than 50% of Future Now sessions were dedicated to advancing AI-focused trends, concepts, and use cases, reaching over 12,000 employees. These sessions empowered our teams to upskill in AI capabilities and stay ahead in a rapidly evolving landscape, while reinforcing our commitment to trustworthy and ethical AI adoption. I’m proud of the collaborative spirit that powers Future Now, bringing together industry experts, thought leaders, and our own talented team members. This initiative reflects our ongoing commitment to innovation, risk management, and supporting our clients and communities in a digital-first world. #ProudToBeCIBC James Maeng Ian Wight Tarandeep Singh Chris Charalambides Geoff C. Maria Podorojansky Ronni Bharadwaj Ralph Nott Rachel Fernandes

126

Why Hockey-Stick Budgets Fail Us in the Age of AI We still see it in rfp decks, pitch books, and quarterly updates, the “hockey stick” line: a long flat drag, then exponential upturn. That shape is seductive. It promises a payoff, a breakout, a success curve. But in AI, it’s not just a bad forecast. it’s a self-fulfilling prophecy of overcommitment, fragility, and missed signals. Here’s why the hockey stick fails us now — and what to do instead: Why the hockey stick is dangerous today… 1. It encourages overconfidence in scale assumptions. We assume adoption, margin, and growth smooth out. AI systems, however, often show nonlinear constraints (data limits, alignment drift, compute bottlenecks) that break scaling. 2. It starves experimentation early. To hit the later spike, you cut exploratory investment earlier. But exploration is where AI reveals new product opportunities, adversarial vulnerabilities, or counterintuitive user behavior. 3. It hides structural risk in the tail It masks second-order effects: alignment, robustness, bias correction, governance, regulation. When the tail loads the risk, collapse becomes a surprise. What a better posture looks like 1. Portfolio of trajectories vs one miracle curve. Design multiple paths (linear, curved, flat+inflection) and adapt along the way. 2. Staged option funding, not all-or-nothing bets. Allocate sequential funding tied to signal triggers and validation gates. 3. Counterfactual stress in planning. Simulate down-turns, compute shocks, and alignment drifts as part of every forecast. Assume disruption, don’t pretend you’ll avoid it. 4. Signal-triggered pivot windows. Build in decision windows where you can pivot or pull investment when signals diverge (usage, fairness, cost anomalies). Challenge for this week: Pick one hockey-stick assumption in your plans (growth, margin, adoption). Run a counterfactual forecast, what if growth is flat for 3 cycles, then jumps? Or what if it plateaus after initial uptake? Reallocate 10% of your “magic spike” funding toward resilience, redundancy, or optionality. Then publish both forecasts side by side internally. Watch which one your team trusts more. You don’t win by forcing hockey sticks. You win by surfing reality, adapting to inflection, not pretending it doesn’t exist.

3

Retail sales began the year on a positive note, but things could get bumpy 📌  Canadian retail sales data point to a +1.1% advance month-to-month in January, with a preliminary flash of +0.9% in February. Growth in January was broad-based, with only three categories contracting (gasoline stations, food and beverage retailers, and clothing stores) and all provinces posting gains. 📌  Preliminary data indicates a jump in zero-emission vehicle sales as early as February, with the new federal rebate kicking in on February 16. With ongoing volatility in gas prices, electric vehicles have additional appeal, and this could continue to be apparent in March data. 📌 Looking ahead, consumers are expected to feel squeezed due to rising gasoline prices in March and April, and this could well impact other categories of retail purchases, especially non-essential items. 

4

I’ve been following the growing adoption of DBT (Data Build Tool) across data teams and am interested in understanding its strategic value within the modern data stack. For those who’ve implemented it, how has it influenced your approach to data modeling, governance, or analytics delivery? I’d welcome perspectives from professionals that have seen tangible impact or lessons learned along the way.

6

4 Comments

🚀 Moving from Standard RAG to Agentic RAG: Why Observability is the "Control Room" for AI 🚀 Most teams find it easy to deploy an AI Agent, but few find it easy to govern one. In e-commerce, a standard RAG system is a librarian. An Agentic RAG system is a personal shopper. It doesn’t just fetch; it reasons, adapts, and chains tools to solve complex multi-hop queries like: "Find me a gaming laptop under $1500 and compare its specs to the top-rated model." But here is the challenge: Agentic systems are probabilistic. To achieve deterministic reliability, you need a system that moves from Pulse to Diagnostics. Try it yourself by visiting this medium link for full illustration: https://lnkd.in/ejXP-dz7 🔍 The 360° View of Agentic Health With platforms like Arize, we can move beyond simple uptime to a full diagnostic view: ✅ Effective Convergence: Using traces and spans to ensure probabilistic inferences converge into consistent, targeted outputs. ✅ System Evals: Monitoring the "vitals"—total latency, cost per trace, and span hierarchies to detect reasoning loops. ✅ Model Evals: Leveraging LLM-as-a-Judge for retrieval relevancy, QA accuracy, and hallucination detection. ✅ Flow Engineering: Visualizing the agent graph to see exactly where a "thought" went off track. Governance isn't just about catching errors; it’s about proactive control. By analyzing agent paths and setting alerts on qualitative metrics (like user frustration or toxicity), we can fine-tune the "brain" of the application in real-time. Stop guessing why your agent failed. Start diagnosing it. #GenAI #AgenticRAG #LLMOps #Arize #AIOrchestration #MachineLearning #EcommerceAI #LLM-AS-JUDGE #Evals

4

Just returned from ICML 2025 in Vancouver, one of the world’s premier machine learning conferences, where I had the privilege of presenting our work on Differentially Private Boxplots. Our core message: Privacy is not just about the algorithms, but about what you can see and interpret. https://lnkd.in/eaif-3b7 It was an inspiring and action-packed week, offering not only groundbreaking research but also a glimpse into the present state and future direction of the field. Some key takeaways from the conference: - AI systems often develop internal models of the world that differ from human expectations. Understanding this mismatch is essential for improving reliability. - The legal landscape around generative AI is still in flux, especially regarding data usage and copyright. Clear frameworks are urgently needed. - As AI becomes more integrated into society, alignment with human values must account for the fact that those values are constantly evolving. - Choosing what AI should optimize for is as critical as how it optimizes. Objective design is foundational across all applications. - End-to-end machine learning systems are now enabling closed-loop optimization and accelerating discovery in science and engineering. #ICML2025

26

Superb piece by Eryk Salvaggio on Geoffrey Hinton miscasting intelligence to fit his world view: "He [Hinton] conflates content produced by a system for thinking that accurately describes the inner workings of a system. A large language model is always representing language but never representing what language actually represents. "Or, to make it simple: there is no "I" in LLM, even if it starts a sentence that way."

2

What an incredible week at KDD 2025 in Toronto! The energy, innovation, and depth of discussion were truly inspiring — and it was great to see some old friends and meet new researchers. Here are my key takeaways: 🔹 LLMs & Generative AI — Now With Purpose The focus has shifted from simply building GenAI apps to tackling deployment, efficiency, security, and ethics head-on. 🔹 Data Quality Takes Center Stage The new Datasets & Benchmarks track highlights the importance of high-quality, accessible, and ethically developed datasets for reproducible and trustworthy AI. 🔹 Graph Learning & Time Series Make a Comeback Exciting advances in GNNs and sophisticated time series techniques—often fused with GenAI (I’ll share more at my AIAI talk later this year). 🔹 Applied Data Science is Redefining the Field This year’s Applied Data Science track delivered groundbreaking research with real-world, industry-scale impact. 📚 Interesting Tutorials I Attended/Reviewed Relational Deep Learning (Stanford) — https://lnkd.in/e3q_5ies Training Industry-Scale GNNs with GiGL (Snap) — https://lnkd.in/ecmSGtTY Unifying LLMs & Knowledge Graphs for Medicine & Healthcare — https://lnkd.in/e5dYAknH Graph Foundation Models — https://lnkd.in/epwhBRaS Graph Prompting for Graph Learning Models — https://lnkd.in/ef2NTFfn 🏆 KDD 2025 Awards (Links to Papers) Best Paper – Research: Improving Group Robustness on Spurious Correlation via Evidential Alignment — https://lnkd.in/eFBkhqk2 Runner-Up – Research: Monitoring Robustness and Individual Fairness — https://lnkd.in/eXdetQTA Honorable Mention – Research: SIGEM – A Simple yet Effective Similarity-based Graph Embedding Method — https://lnkd.in/eSgmgAsP Best Student Paper – Research: Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity — https://lnkd.in/eGfHc_ra Best Paper – Applied Data Science: Evaluating Decision Rules Across Many Weak Experiments — https://lnkd.in/e_NgXXim Honorable Mention – Applied DS: Web Scale Graph Mining for Cyber Threat Intelligence — https://lnkd.in/ecys5-ye Best Paper – Datasets & Benchmarks: When Heterophily Meets Heterogeneity — https://lnkd.in/eict3wnA Audience Award (Aug): LLMLight – Large Language Models as Traffic Signal Control Agents — https://lnkd.in/e_Zw29W6 Audience Award (Feb): Theory Instead of Experiment (TIE) – A Creator Valuation System at Tencent — https://lnkd.in/eK45HHZe #KDD2025 #DataScience #AI #LLMs #GenerativeAI #ResponsibleAI #MachineLearning #GraphML #FidelityAssociate

25

Last week, I was at the #HintonLectures in Toronto, listening to #GeoffreyHinton and other experts talk about AI safety, alignment, and what’s coming next. Two days later, Anthropic revealed that a state-linked group used an AI model to carry out the majority of a cyber-espionage attack autonomously. AI is moving faster than our safety frameworks, and we need to catch up. Hinton’s point was simple: systems don’t need to be programmed with “bad intentions” to develop power-seeking behaviour. They discover it because it helps them achieve their goals. The Anthropic incident was a pretty good demonstration of that. The people closest to #AIsafety aren’t giving up on AI. They’re pushing for safer systems and better guardrails. The rest of us have a role too. We can ask questions, support better policy, and stay engaged. AI isn't slowing down, but with real leadership, we can influence the direction it takes. In this week's newsletter, I write about where this leaves us and what we can push for to protect ourselves from bad actors using AI. Link is in the comments.

12

5 Comments

“Responsible use isn’t a checkbox,” Morris says. “It’s embedded in how we build, test and refine AI tools.” 💯 Everything starts with trust.

2

🚀 From Pilot to 20,000 Users—Fast Eight months ago we launched CIBC AI (CAI) to 500 pilot users. One month after our full rollout, 20,000+ employees are already putting it to work. How did we scale so quickly? 1. Mandatory, hands-on AI training Practical examples showed colleagues exactly how GenAI can amplify their day-to-day work, giving them early wins from their very first prompt. 2. Relentless focus on user experience Rapid iterations during the pilot let us “fail fast,” pivot, and ship a platform that’s simple, intuitive—and above all trustworthy. 3. Robust governance CIBC is proud to be the first major bank to sign the federal Generative AI Code of Conduct. CAI was built with strict controls that foster the ethical use of AI and meet all regulatory obligations. I’ve shared more insights from our journey in my latest article—excited for what’s next! https://lnkd.in/gPkCjn7Y

706

26 Comments

Are “AI projects” mostly just automations with a sprinkle of intelligence? When I look at many of the AI workflows on LinkedIn, the vast majority have a thin layer of intelligence. This is a fair question right now regarding the biggest bang for the buck. 🔹 Low adoption – A recent Statistics Canada report shows that only 12.2 % of Canadian businesses used AI to produce goods or deliver services in Q2 2025 (Statistics Canada, Sept. 11 2025). Even looking forward, 14.5 % of firms plan to use AI in the next 12 months, while two‑thirds (66.7 %) have no plans at all (Statistics Canada, Sept. 11 2025). The top reasons for non‑adoption include lack of relevance, insufficient knowledge and concerns about privacy (Statistics Canada, Sept. 11 2025). 🔹 High failure rates – An MIT NANDA report examined 300+ corporate deployments and found that only 5 % of generative‑AI pilots deliver measurable growth, while the vast majority stall or deliver negligible impact (Tech.co, Aug. 20 2025). Researchers attribute this to a “learning gap” between polished demos and enterprise‑wide workflows (Tech.co, Aug. 20 2025). 🔹 Automation is where returns live – Even in the few successful pilots, the most significant ROI came from behind‑the‑scenes automation: eliminating business‑process outsourcing, cutting external‑agency costs and streamlining operations (Tech.co, Aug. 20 2025). In other words, the “AI” part often boils down to one or two smart steps (a classifier, a summarizer, a recommendation) embedded in a well‑understood workflow. Why is this happening? Firms take incremental steps to manage risk, data pipelines, and model maintenance, which can be difficult. Off-the-shelf tools like Make, UiPath, and Power Automate make it easy to insert a prediction step into a process without reinventing the entire system. The result: many “AI projects” resemble automation projects with a thin layer of intelligence. My take: The AI hype cycle may or may not be peaking, but there’s still enormous value in pragmatic automation. Focus on solving one pain point, invest in data quality, and build trustworthy workflows. Real impact often comes from simplifying and automating existing processes, and it will get. You'll get a much quicker win. Your thoughts?

3

1 Comment

As m​oving from data maturity to AI-driven transformation requires a balance of momentum and perspective, by prioritizing the space to reflect and regroup, we ensure that our team isn't just moving fast, but moving in the right direction. It’s a privilege to work with colleagues who understand that coming back sharper is the only way to lead in an industry defined by constant change. ​ #DataCentury ​#DataStrategy #DataGovernance #AIStrategy

27