How to build AI Analysts that work in production with Atlan AI Labs

A really good high level description of how small AI projects can augment workflows. Two things really stand out to us: 1) AI will fail if it obscures key information elements of the work units to the people managing the effort. 2) In order to determine what "key" elements are you have to collaborate closely with the domain experts and you have to understand the workflow properly and deeply. It's notable how the people who are using AI the most (and Jorge is one of the most practiced users of AI in all of Information Architecture) return, again and again, to the importance of deeply understanding what the work is in the first place. https://buff.ly/Yh9qKm2

In many conversations about AI in the public sector, one truth stands out: 𝘀𝘂𝗰𝗰𝗲𝘀𝘀 𝗱𝗲𝗽𝗲𝗻𝗱𝘀 𝗹𝗲𝘀𝘀 𝗼𝗻 𝗮𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺𝘀 𝗮𝗻𝗱 𝗺𝗼𝗿𝗲 𝗼𝗻 𝗮𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁. Too often, projects stall between ambition and delivery. The technology works, but the systems around it don’t move fast enough to turn pilots into public outcomes. At our recent Qlik analyst meeting, I shared how we can change that by shifting from massive, multi-year programs to small, accountable wins that show real value fast. Imagine one automation agent that cuts fraud claim processing from months to hours or gives citizens instant access to critical information. These quick wins rebuild trust in data, in systems, and in what AI can truly deliver. I believe accountable automation is the way forward for the public sector — grounded in trust, transparency, and strong governance. It’s not about chasing innovation for its own sake, but about delivering results people can see and rely on. At Qlik, we’re helping make that shift happen from assessing data readiness with our AI Trust Score to empowering agencies to build AI systems that are transparent, auditable, and effective. Real transformation doesn’t come from the biggest projects. It comes from the ones that work — and keep working, for everyone. Thank you GovInsider for this insightful read! (link in comments)  Many thanks to Kelly Forbes and Khor-Ping Quek 郭可鑌 for joining us for this discussion. 

How do you future-proof your C# projects with AI in 2025? What if the key to boosting your productivity lies in choosing the right AI tools and frameworks? Dive into customizing your workflows with AI-driven code reviews and automation while strategically navigating cost and provider challenges. Are you ready to rethink your development strategy and embrace AI's potential for flexibility and efficiency? Where do you see this taking your projects? 🤔 🔗 Full article: https://lnkd.in/evQ3XQxX

AI ≠ magic. It’s workflow + prompts + data. Here’s how we think about it 1️⃣ Capture → Reply → Route → Report 2️⃣ Refine prompts to guide AI behavior 3️⃣ Track the right metrics weekly Better workflows > bigger models. Start here. DM PROMPTS for the pack. #AIForBusiness #Ops #ProcessDesign #Flacron #BuildInPublic #Automation #WorkflowDesign #AIIntegration #FlacronConnect

Metadata-Driven Pipelines with AI Logic Manual pipeline logic doesn’t scale. AI-powered metadata pipelines do. By combining metadata-driven frameworks with AI: You can create reusable pipeline templates that auto-adjust to schema changes Use AI to recommend transformations or detect schema drift Build adaptive orchestration that changes based on data characteristics This creates pipelines that don’t just run — they evolve. #MetadataDrivenPipelines #AIinETL #ADF #Synapse #DynamicPipelines #ScalableDataEngineering #MillenniumAnalytica

The difference between messy AI output and clean, reusable output is markdown. When you embed markdown syntax inside prompts, you gain deterministic control over structure. Tables, bullets, bold terms—each tokenized element becomes a formatting command the model follows. That control compounds when paired with template patterns. The result: perfectly formatted data, reports, or documentation directly from the model, no post-processing needed. Markdown is the bridge between text generation and structured automation.

From Models to Systems: the unglamorous work that makes AI useful Most AI programs don’t stall on modeling. They stall on systems thinking. The hard part isn’t getting an LLM to answer; it’s making that answer reliable, auditable, and repeatable inside business workflows. A practical frame I use: 3 layers, 6 agreements. 1) Context Layer (data + events) -Data contracts: schemas + SLAs for the few sources that matter (not all of them). -Freshness policy: how “up to date” is “good enough” for each use case (90m for CX, 24h for finance, etc.). 2) Decision Layer (models + orchestration) -Decision boundary: what the model is allowed to decide vs. what still needs rules or human review. -Feedback loop: how signals (accept/reject/override) flow back to retraining, on a schedule you can sustain. 3) Control Layer (governance + risk) -Traceability: log prompts, inputs, outputs, and feature versions so you can explain a decision six months later. -Guardrails: PII handling, role-based access, denial lists implemented in code, not policy docs. If those six agreements are explicit, most “AI failures” become integration tickets, not existential debates. What to measure (so this doesn’t drift into theatre): - Cycle time to decision (pre vs. post). - The % of decisions auto-approved within guardrails. - Override rate and top override reasons (drives next iteration). - Time to root-cause a bad outcome (should fall as traceability improves). This isn’t about clever prompts or bigger budgets. It’s boring excellence: clear contracts, bounded decisions, observable systems. Do that consistently, and AI stops being a demo and starts compounding value, quietly. #AI #EnterpriseArchitecture #Governance #SolutioningLeadership #IntelligentTransformation

In a recent Snowflake whitepaper titled “Deploying AI Agents at Scale: 3 Patterns to Move Beyond POCs”, I came across a well-defined framework for thinking about how organizations can operationalize AI beyond the pilot stage. Here are the three core agent types the paper highlights: 1️⃣ Data Agents - Designed to combine data and tools efficiently, delivering data-grounded insights with a strong emphasis on accuracy and trust. 2️⃣ Conversational Agents - Focused on interacting with humans naturally, providing informed, context-aware responses to queries or tasks. 3️⃣ Multi-Agent Systems - Built to orchestrate multiple specialized agents, enabling complex workflows where each step may require different expertise or data retrieval. What stood out to me is how these patterns align with real-world enterprise needs - from data reliability to collaboration between intelligent systems, this seems to be the direction AI architecture is evolving toward. 💡 As organizations move from experimentation to implementation, understanding these agent patterns could be key to building scalable, production-ready AI systems. #AI #Snowflake #AIAgents #DataEngineering #EnterpriseAI #MachineLearning #GenAI

AI success starts with data that’s AI-ready 🚀 Too often, AI initiatives stall not because of models, but because the data stack isn’t built for activation. Our Four Elements Framework — Trust, Context, Productization, and Actionability — turns any data stack into an AI-ready foundation that powers faster, more reliable outcomes. 🤔 Curious where your organization stands? 📊 Take the Enterprise AI Readiness Assessment to see your current state and get a clear roadmap to activate your data for AI success. 👉 https://lnkd.in/gDRcpvf5

The Great MLOps Pivot: Why Data is the real AI moat? For years, AI teams have focused too much on building fancy models - new architectures, better fine-tuning & endless benchmarks. But once these models go live, they often fail because real-world data keeps changing. The real game-changer isn't a new model - it's shifting from model-centric to data-centric AI. In a model-centric world, the model is everything and data is just fuel. In a data-centric world, data becomes the main asset - it's cleaned, improved, and updated to make models work better over time. Data-centric MLOps means building systems that keep learning and improving. It's powered by three key ideas: -> Feedback Loops: Learn from how users interact - what they like, ignore or correct. Detect when data or predictions start drifting off. -> Data Curation: Treat data like code. Version, track and validate it using tools like DVC or Pachyderm. -> Continuous Retraining: Automate retraining when performance drops or new data arrives - with safe rollouts. This turns static, fragile models into living systems that get smarter every day. As Andrew Ng says - improving data quality gives a far higher return than endlessly tweaking models. Start small: pick one feedback signal and start logging it. That's the first turn of your data flywheel. Stop building fancy models. Start building systems that learn. #MLOps #DataCentricAI #AI #MachineLearning #DataScience #AIML