Dustin Tucker

Everyone wants the penthouse. Nobody budgets for the basement. What usually gets funded: → AI strategy → GenAI use cases → Dashboards and insights What quietly gets skipped: → Definitions → Data quality → Metadata → Lineage → Ownership Until the building starts to lean. AI doesn’t fail at the top. It fails because the bottom was never finished. Foundations aren’t boring work. They’re the only reason anything above them stays standing. Credit: John

2

2 Comments

Great clarification from my colleague Emily Pfeiffer for ARs about scheduling time with Forrester analysts. Both clients and non clients can request briefings. Additionally, if you're not aware, you don't have to be a client to be considered for a Wave. What you do need is to be on our radar regardless of client status. Check out Joseph Blankenship 's post about how to crush an analyst briefing. ( https://lnkd.in/e7zBdwuf)

13

6 Comments

Great to see Christoph Molnar's thoughts on tabpfn v3. I just updated the ScoringBench results under https://https://scoringbench.com/ Check out the results! Main things I noticed: v3 improved a lot for point estimates and scoring rules from the beta energy family (including CRPS) as well at various interval scores. I also added some more diagnostics like PIT calibration and coverage at various levels in this update of ScoringBench. Enjoy exploring the results!

2

In 2026 we would experience a profound transformation💡in how Agentic systems would be designed: From: “How do I craft the perfect prompt for Agent?” ➡️ To: “How do I design the entire information ecosystem for Agent?” From: “Just add more examples to the prompt” ➡️ To: “Build strategic retrieval mechanisms, caching layers, and memory hierarchies” From: “Let’s hope the model figures it out (Magic 😃)” ➡️ To: “Continuously measure, instrument, and optimise context at every level”

57

1 Comment

Your AI vendor is not the problem. Your data quality is not the problem. Your team is not the problem. The problem is that your organization has no memory. Every pilot learns how decisions actually get made here. Not the org chart version, the real one. Which precedents are important?. What worked last time?. Why certain approaches are non-starters?. Then the pilot ends. The team moves on. The knowledge vanishes. Next pilot, same archaeology. You're not failing at AI. You're succeeding at forgetting. #EnterpriseAI #AIStrategy #DigitalTransformation #AIAdoption 

12

One thing this year is already showing us: Model capabilities aren't the only dimension of AI lab competition right now. Model capabilities are just the beginning. People want models that remember things, connect to relevant data sources, use appropriate tools, and can be taught via skills. And they want this all wrapped up in a great, intuitive user experience. The year has kicked off, for example, with a battle between OpenAI and Anthropic around healthcare. But this battle didn't involve new models. It's entirely around complementary things like memory (sequestered health memories in ChatGPT Health), data connections (e.g. Apple Health), agent skills (e.g. prior auths), tools (e.g. analyzing data and generating health charts), and user experience (e.g. ChatGPT Health app). Model capabilities enable all of this to work, of course. But, increasingly, that's just one component of differentiation.

19

We tested Snowflake's Gen2 warehouses on 2 billion query profiles across 14K warehouses here at Capital One. Here is where we saw improvement: 🍎 Approximately 30% of queries fell into categories that could benefit from Gen2—specifically those involving DELETE, UPDATE or MERGE. 🍎 Long-running DML operations: A significant subset of queries running longer than 30 minutes were performing complex DML tasks. These are prime candidates for Gen2 migration. 🍎 Large scans: Numerous queries scanned over 1 TB of data, where Gen 2’s improved scan performance can provide material runtime reduction. 🍎 Combined heavy patterns: A smaller but critical set of queries both scanned over 1 TB and ran for 30 minutes or more, often involving DML operations—these represent the highest potential gain from Gen2. These findings support a targeted migration strategy: focus on warehouses with a significant concentration of eligible workloads rather than migrating all warehouses indiscriminately. Check out the full blow written by our resident expert and all around awesome guy - Ganesh Bharathan https://lnkd.in/g8XufZaV

292

14 Comments

The litmus test for a great self-service data tool is simple: do data staff use it themselves, or not? This may seem counter-intuitive. Self-service data tools are for people who CAN'T write SQL, right? Who cares what the data staff use? They can use more complex tools. This test works for two reasons. - Data staff are discerning. They will reject a tool that has bad visualizations, bad data, or poor performance. Because they CAN opt for something else, winning their trust is a sign you have built a great tool. - If you do NOT win the trust of data staff, you will struggle to maintain consistency between data sets. The data exposed in the self-service tool will tend to be the "basic" data for other people, and be less complete and less well-maintained. It's intuitive to think that the point of a self-service tool is to meet the needs of everyone BUT the data staff, and focus on them as your potential customers. This is backwards. Start with data staff as your target, and if you meet their needs you can expand to include a broader set of users over time.

55

2 Comments

Data infrastructure is hot again. My friends in this space have seen demand pick up sharply just since January. Why? 1) Bad data foundations prevent rapid AI adoption. The tooling doesn't matter if your foundations aren't set up to enable speed. 2) "Democratizing" data access via AI tools puts new and often heavy load on data infrastructure that previously only had to deal with experts using it. 3) Agents have different patterns of interaction than people. They stress-test data foundations in a way people just do not. 4) As AI use picks up, cost becomes a real problem. Hitting small tables isn't a big deal. Hitting tables with billions of rows many times is an issue. 5) Uncertainty. Companies know this is becoming problematic, but don't have the in-house expertise to deal with it yet. I imagine this is only going to continue as agents (and analysts) become more data-intensive.

93

12 Comments

Building agentic workflows? There are actually only a few key concepts to be aware of. StateGraph orchestration: A StateGraph is the agent’s blueprint. It consists of nodes that handle tasks with edges (directions) to other tasks, defining the order. Together, they provide the agent with a flow to follow. Conditional routing: At decision points, the agent inspects results and follows “if-else” statements that skip failures, handle edge cases, etc. Memory and threads: Thread memory holds recent conversation turns. Essentially, it stores the short-term memory that you are John Doe and were interested in the weather. Long-term memory stores user facts, letting the agent stay context-aware, reduce repetition, and personalize replies, such as that you have been a customer since 2018. Human-in-the-loop: Strategic hand-offs let people review low-confidence outputs, catch critical paths or errors, and feed corrections back to the model, keeping quality and alignment tight. If this still feels too abstract, I’ve included a more detailed explanation in the article below. https://lnkd.in/eAr-6YAQ

14

1 Comment

The "Modern Data Stack" is a mess. Or at least, that’s what I thought until I realized I was overcomplicating it. Most people talk about data like it’s rocket science. It’s actually just a high-tech plumbing system. If you’re tired of nodding along in meetings without actually "getting it," here is the Modern Data Stack explained in 60 seconds: 1. The Source (The Water) This is where your data lives before it’s useful. Think Salesforce, Shopify, or your backend SQL database. It’s raw, messy, and scattered everywhere. 2. Ingestion (The Pipes) You need to move that water to one place. Tools like Fivetran, ADF, SSIS or Airbyte act as the pipes. They don't change the data; they just transport it. 3. Storage (The Reservoir) This is your Data Warehouse. Snowflake, Fabric, BigQuery, or Databricks. It’s a massive, centralized lake where all your "water" sits waiting to be cleaned. 4. Transformation (The Filter) Raw data is dirty. You can’t drink it (or make decisions with it). dbt is the industry standard here. It turns "User_ID_992" into "Active Customer" so humans can actually understand it. 5. Business Intelligence (The Tap) This is the part you actually see. Looker, Tableau, or Power BI. It’s the faucet that turns the data into a clean, drinkable chart. The big shift? In the old days, we "Cleaned → Moved → Stored" (ETL). It was slow, expensive, and broke constantly. Today, we "Move → Store → Clean" (ELT). It’s faster, scalable, and allows you to change your mind about how you "filter" the data later. You don't need a PhD to understand data. You just need to understand the flow. Which part of the stack is giving your team the most headaches right now? #DataEngineering #ModernDataStack #DataAnalytics #BusinessIntelligence #DataStrategy

4

One of the most dangerous things in analytics isn’t bad data. It’s good data without context. I’ve seen incredibly sophisticated dashboards lead smart leaders to the wrong conclusions—not because the numbers were wrong, but because the story around the numbers was missing. Two organizations can show the same metric and be in completely different situations: • Different operating models • Different system constraints • Different strategic trade‑offs • Different definitions hiding behind the same label When we compare without context, we don’t just risk misinterpretation, we risk making confident decisions for the wrong reasons. The real work of analytics isn’t just measurement. It’s interpretation, framing, and asking better questions: Why does this look different? What assumptions are embedded in this metric? What would need to be true for this comparison to be meaningful? Data doesn’t speak for itself. People do using data as a tool. Curious how others are thinking about this: Where have you seen context (or lack of it) materially change a decision?

15

I did a double take when I saw a Forbes headline claiming that 80% of databases are now being created by AI agents. Whether it’s one platform today or many tomorrow, this trajectory feels inevitable—and likely already happening in more places than we realize. One line from the article stopped me: “An AI agent working through a complex coding problem may spin up dozens of isolated database environments in parallel, test multiple hypotheses at once, evaluate results, and then tear everything down—all within seconds.” Let that sink in. This isn’t just automation at scale. It fundamentally redefines the data layer itself—governance, cost control, security, lineage, resilience, and ownership are no longer static design decisions. The upside is enormous: ⚡ experimentation at machine speed ⚡ innovation humans simply can’t match The risk is just as real: ⚠ unmanaged sprawl ⚠ opaque decision-making ⚠ architectures we no longer fully understand The risks and rewards are too large to ignore. If AI agents are becoming first-class actors in creating and managing data infrastructure, leaders need to start asking harder questions: • Who governs data environments that exist only for seconds? • How do we secure what’s constantly created and destroyed? • What does “data architecture” even mean in an agent-driven world? And critically: Is this only happening at hyperscalers? How do we support highly regulated environments like finance and healthcare—where auditability, change management, and lineage aren’t optional? Curious how others are thinking about this shift—especially those responsible for data, platform, and risk. https://lnkd.in/gdfQEr-Y #AI #AIAgents #DataArchitecture #AnalyticsLeadership #DataGovernance #CIO #CDO #RiskManagement

53

8 Comments

One thing I keep seeing in enterprise AI rollouts: the highest-risk data usually isn’t in tables it’s in the messy stuff. Tickets, call transcripts, notes, emails… the place where PII shows up casually. That’s why I liked seeing Snowflake move in two practical directions recently: A clearer path to structuring unstructured data with Cortex AI Functions, so teams can extract consistent signals without building a one-off pipeline every time. AI_REDACT reaching GA automated PII detection + redaction for unstructured text, built into the platform. From an AI & Security Platform lens, this is the kind of “boring guardrail” that actually enables speed: you reduce privacy risk upfront, and you make cross-team alignment (engineering / data / security) much less painful. If you’re working with unstructured data today, where do you enforce redaction at ingestion, at query-time, or right before model access?

1

I sharded our database. Best decision ever. Until it wasn't. Last year, our write throughput was dying. Every user action—updates, state changes, activity logs—was slamming the same tables. Classic bottleneck. The solution felt obvious: shard by user_id. And honestly? It worked beautifully 💯 Write throughput went up 3x. Contention basically disappeared. Latency dropped hard. The team could breathe again. I thought we'd solved it. Then product evolved. Suddenly we needed users to collaborate. Share resources. Enforce global rules. Run workflows that touched multiple accounts at once. And that's when everything started breaking in weird ways. Simple queries now hit multiple shards. Transactions that used to be atomic became "eventually consistent" nightmares. I was writing compensating logic, retry handlers, background jobs to "fix up" state after the fact. That's when I realized , We didn't have a scale problem. We had Data Model problem. I'd sharded MutableSttate. Not events. I changed: - Instead of storing "current state" in shards, I made writes append-only. Every change became an event. Current state became a derived view—rebuilt from events. - This let me keep events in sharded storage (they have clear owners), run async pipelines for side-effects, and centralize only what needed to be global. The difference was night and day. Write spikes stopped hurting. Correctness got simpler. New features didn't need architecture rewrites. Sharding isn't a database trick. It's a domain decision. - If your boundaries are artificial (like "let's just split by user_id"), the system will expose that—eventually. - High write load ≠ shardable workload. - Fix the model first. Scale second. Yeah, event sourcing added complexity—schemas, replays, eventual consistency. But those problems were manageable. Cross-shard transactions? Not so much.

32

5 Comments

Most analytics systems have a blind spot. We'll talk a lot about data quality and model accuracy -- but very little about the layer between forecasts and operations. That layer is where: - units quietly change - constraints are assumed instead of enforced - costs are treated as symmetric when they're not - outputs become "guidance" rather than decisions Nothing crashes. Dashboards still update. Models still score well. But decisions slowly drift out of alignment with reality. More to come.

11

Most metrics don’t change because reality changed. They change because the denominator moved. ✏️ New eligibility rules. 📅 Different time window. ☑️ A backlog cleared. 🧹 A coding cleanup. ❓ Then leadership asks: “Why did performance drop?” Sometimes the answer is: 😬 It didn’t. The measurement did. My favorite discipline in analytics isn’t building dashboards. It’s protecting decision-makers from false confidence. For better decisions - Track the metric and the mechanics behind it: eligible population size data freshness / completeness backlog volume exception rates definition changes 〰️ That’s how you separate signal from noise. #analytics #metrics #leadership #dataquality #transformation

8

🍌 The biggest data quality mistake I see people make is not unlike baking banana bread. 🤢 What can you do with leaky bananas? Honestly, not much. It's too late.  Being intentional about data quality is just as crucial as being intentional about baking banana bread. You can't "fix" a dataset if you weren't intentional about quality from the start. It's increasingly difficult, if not impossible 😭, to spot bad actors after the fact. I think of it as layers of protection. None of the layers are perfect, but we have the best chance to prevent poor data quality when we apply them all together. 🔍𝐅𝐫𝐚𝐮𝐝 𝐝𝐞𝐭𝐞𝐜𝐭𝐢𝐨𝐧 𝐬𝐨𝐟𝐭𝐰𝐚𝐫𝐞 needs to detect advanced browser/device fingerprinting, monitor in-survey mouse movements and assess open-end for programmatic and cut/paste. You may need multiple tools, not just one. ✅𝐈𝐧-𝐬𝐮𝐫𝐯𝐞𝐲 𝐜𝐡𝐞𝐜𝐤𝐬 to term as many poor quality respondents before they register as completes. 📊𝐏𝐨𝐬𝐭-𝐬𝐮𝐫𝐯𝐞𝐲 𝐜𝐡𝐞𝐜𝐤𝐬 to sense-check the data (and data coherence). The challenge is that, at this stage, you can identify issues in aggregate, but it's much harder to pinpoint issues at the respondent level if you didn’t use the right technical tools. Preventing poor data quality is easier than trying to ‘fix’ bad data. Without these measures, no statistical magic can help you on the back end. What does your data quality stack look like? #mrx #dataquality

58

19 Comments

When estimating compound growth, the fact that business analysts call what a demographer would call a rate a factor and then call a transformation of that rate into a percent change a rate... ... this will always piss me off more than it should.

2