Data Quality for AI

𝗜𝗳 𝘆𝗼𝘂 𝘄𝗮𝗻𝘁 𝘁𝗼 𝗯𝘂𝗶𝗹𝗱 𝗮𝗻 𝗔𝗜 𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆 𝗳𝗼𝗿 𝘆𝗼𝘂𝗿 𝗰𝗼𝗺𝗽𝗮𝗻𝘆, 𝘆𝗼𝘂 𝗳𝗶𝗿𝘀𝘁 𝗻𝗲𝗲𝗱 𝘁𝗼 𝗯𝘂𝗶𝗹𝗱 𝗮 𝘀𝗼𝗹𝗶𝗱 𝗱𝗮𝘁𝗮 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝗮𝗻𝗱 𝗲𝗻𝗳𝗼𝗿𝗰𝗲 𝘀𝘁𝗿𝗶𝗰𝘁 𝗱𝗮𝘁𝗮 𝗵𝘆𝗴𝗶𝗲𝗻𝗲. Getting your house in order is the foundation for delivering on any AI ambition. The MIT Technology Review — based on insights from 205 C-level executives and data leaders — lays it out clearly: 𝗠𝗼𝘀𝘁 𝗰𝗼𝗺𝗽𝗮𝗻𝗶𝗲𝘀 𝗱𝗼 𝗻𝗼𝘁 𝗳𝗮𝗰𝗲 𝗮𝗻 𝗔𝗜 𝗽𝗿𝗼𝗯𝗹𝗲𝗺. 𝗧𝗵𝗲𝘆 𝗳𝗮𝗰𝗲 𝗰𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲𝘀 𝗶𝗻 𝗱𝗮𝘁𝗮 𝗾𝘂𝗮𝗹𝗶𝘁𝘆, 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲, 𝗮𝗻𝗱 𝗿𝗶𝘀𝗸 𝗺𝗮𝗻𝗮𝗴𝗲𝗺𝗲𝗻𝘁. Therefore, many firms are still stuck in pilots, not production. Changing that requires strong data foundations, scalable architectures, trusted partners, and a shift in how companies think about creating real value with AI. Because pilots are easy, BUT scaling AI across the enterprise is hard. 𝗛𝗲𝗿𝗲 𝗮𝗿𝗲 𝘁𝗵𝗲 𝗸𝗲𝘆 𝘁𝗮𝗸𝗲𝗮𝘄𝗮𝘆𝘀: ⬇️ 1. 95% 𝗼𝗳 𝗰𝗼𝗺𝗽𝗮𝗻𝗶𝗲𝘀 𝗮𝗿𝗲 𝘂𝘀𝗶𝗻𝗴 𝗔𝗜 — 𝗯𝘂𝘁 76% 𝗮𝗿𝗲 𝘀𝘁𝘂𝗰𝗸 𝗮𝘁 𝗷𝘂𝘀𝘁 1–3 𝘂𝘀𝗲 𝗰𝗮𝘀𝗲𝘀:   ➜ The gap between ambition and execution is huge. Scaling AI across the full business will define competitive advantage over the next 24 months. 2. 𝗗𝗮𝘁𝗮 𝗾𝘂𝗮𝗹𝗶𝘁𝘆 𝗮𝗻𝗱 𝗹𝗶𝗾𝘂𝗶𝗱𝗶𝘁𝘆 𝗮𝗿𝗲 𝘁𝗵𝗲 𝗿𝗲𝗮𝗹 𝗯𝗼𝘁𝘁𝗹𝗲𝗻𝗲𝗰𝗸𝘀: ➜ Without curated, accessible, and trusted data, no AI strategy can succeed — no matter how powerful the models are. 3. 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲, 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆, 𝗮𝗻𝗱 𝗽𝗿𝗶𝘃𝗮𝗰𝘆 𝗮𝗿𝗲 𝘀𝗹𝗼𝘄𝗶𝗻𝗴 𝗔𝗜 𝗱𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁 — 𝗮𝗻𝗱 𝘁𝗵𝗮𝘁 𝗶𝘀 𝗮 𝗴𝗼𝗼𝗱 𝘁𝗵𝗶𝗻𝗴:   ➜ 98% of executives say they would rather be safe than first. Trust, not speed, will win in the next AI wave. 4. 𝗦𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘇𝗲𝗱, 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀-𝘀𝗽𝗲𝗰𝗶𝗳𝗶𝗰 𝗔𝗜 𝘂𝘀𝗲 𝗰𝗮𝘀𝗲𝘀 𝘄𝗶𝗹𝗹 𝗱𝗿𝗶𝘃𝗲 𝘁𝗵𝗲 𝗺𝗼𝘀𝘁 𝘃𝗮𝗹𝘂𝗲:  ➜ Generic generative AI (chatbots, text generation) is table stakes. True differentiation will come from custom, domain-specific applications. 5. 𝗟𝗲𝗴𝗮𝗰𝘆 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝗮𝗿𝗲 𝗮 𝗺𝗮𝗷𝗼𝗿 𝗱𝗿𝗮𝗴 𝗼𝗻 𝗔𝗜 𝗮𝗺𝗯𝗶𝘁𝗶𝗼𝗻𝘀:  ➜ Firms sitting on fragmented, outdated infrastructure are finding that retrofitting AI into legacy systems is often more costly than building new foundations. 6. 𝗖𝗼𝘀𝘁 𝗿𝗲𝗮𝗹𝗶𝘁𝗶𝗲𝘀 𝗮𝗿𝗲 𝗵𝗶𝘁𝘁𝗶𝗻𝗴 𝗵𝗮𝗿𝗱: ➜ From GPUs to energy bills, AI is not cheap — and mid-sized companies face the biggest barriers. Smart firms are building realistic ROI models that go beyond hype. 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗮 𝗳𝘂𝘁𝘂𝗿𝗲-𝗿𝗲𝗮𝗱𝘆 𝗔𝗜 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝗶𝘀𝗻’𝘁 𝗮𝗯𝗼𝘂𝘁 𝗰𝗵𝗮𝘀𝗶𝗻𝗴 𝘁𝗵𝗲 𝗻𝗲𝘅𝘁 𝗺𝗼𝗱𝗲𝗹 𝗿𝗲𝗹𝗲𝗮𝘀𝗲.   𝗜𝘁’𝘀 𝗮𝗯𝗼𝘂𝘁 𝘀𝗼𝗹𝘃𝗶𝗻𝗴 𝘁𝗵𝗲 𝗵𝗮𝗿𝗱 𝗽𝗿𝗼𝗯𝗹𝗲𝗺𝘀 — 𝗱𝗮𝘁𝗮, 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲, 𝗴𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲, 𝗮𝗻𝗱 𝗥𝗢𝗜 — 𝘁𝗼𝗱𝗮𝘆.

According to Gartner, AI-ready data will be the biggest area for investment over the next 2-3 years. And if AI-ready data is number one, data quality and governance will always be number two. But why? For anyone following the game, enterprise-ready AI needs more than a flashy model to deliver business value. Your AI will only ever be as good as the first-party data you feed it, and reliability is the single most important characteristic of AI-ready data. Even in the most traditional pipelines, you need a strong governance process to maintain output integrity. But AI is a different beast entirely. Generative responses are still largely a black box for most teams. We know how it works, but not necessarily how an independent output is generated. When you can’t easily see how the sausage gets made, your data quality tooling and governance process matters a whole lot more, because generative garbage is still garbage. Sure, there are plenty of other factors to consider in the suitability of data for AI—fitness, variety, semantic meaning—but all that work is meaningless if the data isn’t trustworthy to begin with. Garbage in always means garbage out—and it doesn’t really matter how the garbage gets made. Your data will never be ready for AI without the right governance and quality practices to support it. If you want to prioritize AI-ready data, start there first.

In the last 3 months at Ahrefs, we analyzed over 1 billion data points across 11 studies*. Here's what we learned about AI search optimization: 1. YouTube mentions are the single strongest predictor of AI visibility (correlation: 0.737) – stronger than Domain Rating, backlinks, or any traditional SEO factor. YouTube is heavily cited in AI responses, and both Google and OpenAI train on YouTube content. 2. For a given query, AI Mode and AI Overviews reach the same conclusions 86% of the time – but cite almost entirely different sources (only 13.7% citation overlap). AI Mode responses are 4x longer and mention 3x more entities. 3. Content length has essentially zero correlation with AI citations (0.04). 53% of all AI Overview citations go to pages under 1,000 words. Writing ultra-long contentisan't necessary for AI visibility. 4. Google still sends 345x more traffic than ChatGPT, Gemini, and Perplexity combined – but ChatGPT accounts for 80%+ of all AI-driven website traffic. 5. AI Overviews have a 70% chance of changing from one observation to the next, with content lasting an average of just 2.15 days. But semantic meaning stays remarkably consistent (0.95 cosine similarity). 6. "Best X" blog lists make up 43.8% of all page types cited in ChatGPT responses. 35% of those lists come from low-authority domains. 7. 79% of blog lists cited by ChatGPT were updated in 2025, and 76% of top-cited pages were refreshed within the last 30 days. Freshness matters more than ever. 8. When asked questions without valid answers, AI systems choose fabricated content with specific numbers almost every time. ChatGPT resisted best (84% accuracy), but Grok and Copilot were fully manipulated. 9. Domain Rating correlates weakly with AI visibility (just 0.266-0.326 across platforms). Number of site pages is even weaker at 0.194. 10. 67% of ChatGPT's top 1,000 citations are essentially off-limits to marketers – Wikipedia alone accounts for 29.7%, followed by homepages (23.8%) and educational content (at just 19.4%). *i'll share all the study links in a comment!

Last week, I described four design patterns for AI agentic workflows that I believe will drive significant progress: Reflection, Tool use, Planning and Multi-agent collaboration. Instead of having an LLM generate its final output directly, an agentic workflow prompts the LLM multiple times, giving it opportunities to build step by step to higher-quality output. Here, I'd like to discuss Reflection. It's relatively quick to implement, and I've seen it lead to surprising performance gains. You may have had the experience of prompting ChatGPT/Claude/Gemini, receiving unsatisfactory output, delivering critical feedback to help the LLM improve its response, and then getting a better response. What if you automate the step of delivering critical feedback, so the model automatically criticizes its own output and improves its response? This is the crux of Reflection. Take the task of asking an LLM to write code. We can prompt it to generate the desired code directly to carry out some task X. Then, we can prompt it to reflect on its own output, perhaps as follows: Here’s code intended for task X: [previously generated code] Check the code carefully for correctness, style, and efficiency, and give constructive criticism for how to improve it. Sometimes this causes the LLM to spot problems and come up with constructive suggestions. Next, we can prompt the LLM with context including (i) the previously generated code and (ii) the constructive feedback, and ask it to use the feedback to rewrite the code. This can lead to a better response. Repeating the criticism/rewrite process might yield further improvements. This self-reflection process allows the LLM to spot gaps and improve its output on a variety of tasks including producing code, writing text, and answering questions. And we can go beyond self-reflection by giving the LLM tools that help evaluate its output; for example, running its code through a few unit tests to check whether it generates correct results on test cases or searching the web to double-check text output. Then it can reflect on any errors it found and come up with ideas for improvement. Further, we can implement Reflection using a multi-agent framework. I've found it convenient to create two agents, one prompted to generate good outputs and the other prompted to give constructive criticism of the first agent's output. The resulting discussion between the two agents leads to improved responses. Reflection is a relatively basic type of agentic workflow, but I've been delighted by how much it improved my applications’ results. If you’re interested in learning more about reflection, I recommend: - Self-Refine: Iterative Refinement with Self-Feedback, by Madaan et al. (2023) - Reflexion: Language Agents with Verbal Reinforcement Learning, by Shinn et al. (2023) - CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing, by Gou et al. (2024) [Original text: https://lnkd.in/g4bTuWtU ]

Before You Obsess Over MCP or A2A… Fix Your Data Everyone’s talking about agent protocols—MCP, A2A, interoperability, orchestration layers… and yes, those are important. But don’t miss the most important part. It doesn’t matter how smoothly your agents talk to each other if they’re all speaking garbage. Protocols help agents communicate and/or interact with tools, but data is what they think with. An AI agent is only as good as the data it operates on. Feed it incomplete, outdated, or inconsistent data, and 𝐢𝐭 𝐰𝐢𝐥𝐥 𝐟𝐚𝐢𝐥—𝐟𝐚𝐬𝐭, 𝐚𝐧𝐝 𝐰𝐢𝐭𝐡 𝐜𝐨𝐧𝐟𝐢𝐝𝐞𝐧𝐜𝐞. Some common pitfalls? ▪️Agents disconnected from real-time operational data (hello, CRM silos). ▪️Structural errors and inconsistent formats. ▪️Snapshots of old data trying to guide dynamic decisions. And yet, most teams spend more time wiring up protocols than cleaning their inputs. Data quality isn’t a nice-to-have—it’s the foundation. Want to build smart agents? ✔️Standardize and clean your structured data. ✔️Integrate real-time sources. ✔️Create feedback loops to refine over time. ✔️Prioritize data engineering over just protocol engineering. In short: MCP might make your agents sound smart. Clean data makes them actually smart. Build agents that reason, not just talk. #ai #data #genai #agents #mcp

AI is not failing because of bad ideas; it’s "failing" at enterprise scale because of two big gaps: 👉 Workforce Preparation 👉 Data Security for AI While I speak globally on both topics in depth, today I want to educate us on what it takes to secure data for AI—because 70–82% of AI projects pause or get cancelled at POC/MVP stage (source: #Gartner, #MIT). Why? One of the biggest reasons is a lack of readiness at the data layer. So let’s make it simple - there are 7 phases to securing data for AI—and each phase has direct business risk if ignored. 🔹 Phase 1: Data Sourcing Security - Validating the origin, ownership, and licensing rights of all ingested data. Why It Matters: You can’t build scalable AI with data you don’t own or can’t trace. 🔹 Phase 2: Data Infrastructure Security - Ensuring data warehouses, lakes, and pipelines that support your AI models are hardened and access-controlled. Why It Matters: Unsecured data environments are easy targets for bad actors making you exposed to data breaches, IP theft, and model poisoning. 🔹 Phase 3: Data In-Transit Security - Protecting data as it moves across internal or external systems, especially between cloud, APIs, and vendors. Why It Matters: Intercepted training data = compromised models. Think of it as shipping cash across town in an armored truck—or on a bicycle—your choice. 🔹 Phase 4: API Security for Foundational Models - Safeguarding the APIs you use to connect with LLMs and third-party GenAI platforms (OpenAI, Anthropic, etc.). Why It Matters: Unmonitored API calls can leak sensitive data into public models or expose internal IP. This isn’t just tech debt. It’s reputational and regulatory risk. 🔹 Phase 5: Foundational Model Protection - Defending your proprietary models and fine-tunes from external inference, theft, or malicious querying. Why It Matters: Prompt injection attacks are real. And your enterprise-trained model? It’s a business asset. You lock your office at night—do the same with your models. 🔹 Phase 6: Incident Response for AI Data Breaches - Having predefined protocols for breaches, hallucinations, or AI-generated harm—who’s notified, who investigates, how damage is mitigated. Why It Matters: AI-related incidents are happening. Legal needs response plans. Cyber needs escalation tiers. 🔹 Phase 7: CI/CD for Models (with Security Hooks) - Continuous integration and delivery pipelines for models, embedded with testing, governance, and version-control protocols. Why It Matter: Shipping models like software means risk comes faster—and so must detection. Governance must be baked into every deployment sprint. Want your AI strategy to succeed past MVP? Focus and lock down the data. #AI #DataSecurity #AILeadership #Cybersecurity #FutureOfWork #ResponsibleAI #SolRashidi #Data #Leadership

The AI Wave Finally Hit - and AI Data Readiness will be next. Over the past few weeks, something has shifted. OpenClaw has been making waves in the open-source world. Claude Cowork's plugins landed with a thud, particularly the automated contract-analysis tool. In days, hundreds of billions in market value were wiped off established software and IT services stocks. Markets don't reprice like that over a single product. They reprice when a deeper assumption breaks. 🔵 The Threshold Those of us working closely with systems like Claude Code have seen this coming - especially since Opus-class models made agentic workflows viable. But this feels like the moment the conversation crossed a threshold. What was niche is now mainstream. Here is a prediction: people will soon wake up to the importance of how these systems are grounded. Frameworks like OpenClaw are powerful, but rely on emergent behaviour over loosely structured context. An ontology-backed data structure gives you something tighter: clearer constraints, more predictable reasoning, and far less ambiguity about what the system is allowed to conclude. That difference shows up as reliability, and it will become impossible to ignore once people start to engage seriously. 🔵 A Personal Resonance For years, my argument was simple: AI is coming, and organisations need to get their data ready. It was never about chasing the latest technology. It was about recognising that once AI arrived, the limiting factor would not be the models - it would be the data. What I didn't anticipate is how unprepared I would be when that moment truly arrived. Last week, as I fully "wire-headed" into one of our internal agents - with direct access to our ontology and knowledge graph - something crystallised. The speed with which organisational context became usable, the way complex structures turned into leverage, was both exhilarating and unsettling. We are not psychologically prepared for what this is going to feel like. You can see this by the slightly manic look in the eyes of those who have already wire-headed. 🔵 The Principles That Still Hold As things get increasingly volatile, it's worth returning to core principles I've been repeating for a decade. First, focus on your data. AI is like an iceberg: what you see above the surface gets the attention, but what matters is what sits underneath. Second, "getting your data ready" means two things: linking it together richly, and organising it semantically. Without that, AI systems either underperform or produce confident nonsense. Finally, stick to open standards. They are the only reliable way to maintain flexibility as tools, vendors, and architectures change faster than organisations can react. The recent market reaction wasn't panic over a single tool. It was a delayed recognition of a reality building for years. AI didn't arrive overnight. But now that it's here, the cost of not being data-ready will become visible - all at once.

When a dashboard crashes, the finger-pointing starts. Is it the Engineer? The Analyst? The Steward? Think of data governance like building a bridge. One engineer can design brilliant steel beams, but if the foundation team uses weak concrete and the inspection team skips safety checks—the bridge collapses. You can't blame the steel. 🎬 𝗗𝗮𝘁𝗮 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 = 𝗔 𝗦𝘆𝗺𝗽𝗵𝗼𝗻𝘆, 𝗡𝗼𝘁 𝗮 𝗦𝗼𝗹𝗼 🔧 Engineers: Build the pipelines (the stage) 📊 Analysts: Define the metrics (the script) 🔮 Scientists: Extract insights (direct the plot) 📜 Stewards: Own data quality (manage backstage) 📈 Business: Drive decisions (deliver the finale) Miss one cue? The entire show derails. 💥 What Happens Without Governance? ❌ Wrong data flows into dashboards → Bad decisions made with confidence ❌ Silos form between teams → Duplicate work, conflicting sources ❌ Finger-pointing replaces fixing → Problems fester, trust erodes ❌ Reactive patches, not root fixes → Same fires, different day The Damage: When Governance Fails → $12.9M lost per organization annually (as per Gartner's research) Wasted spend. Bad decisions. Endless rework. Missed opportunities. → 15-25% revenue leakage Decisions made on incomplete data. Inconsistent sources. Duplicate records. → $4.5M average breach cost (2025) U.S. and U.K.? Often $9-10M per incident. Security isn't optional anymore. → $3.1T drained from U.S. economy Failed initiatives. Wasted effort. Poor quality compounds across industries. 𝘕𝘰𝘵 𝘢 𝘵𝘦𝘤𝘩 𝘱𝘳𝘰𝘣𝘭𝘦𝘮. 𝘈 𝘵𝘦𝘢𝘮𝘸𝘰𝘳𝘬 𝘱𝘳𝘰𝘣𝘭𝘦𝘮. 💡 The Fix: Build Systems, Not Silos ✅ Automate quality checks → Catch issues before analysts do ✅ Track lineage & metadata → Know where data comes from, where it goes ✅ Design for observability → Monitor pipelines like you monitor apps ✅ Embed compliance early → Privacy isn't a checkbox—it's architecture ✅ Break down role barriers → Engineers, analysts, stewards—one team 🎯 The Bottom Line Governance isn't bureaucracy. It's the blueprint for data systems that don't crumble under pressure. 𝘎𝘳𝘦𝘢𝘵 𝘥𝘢𝘵𝘢 𝘪𝘴𝘯'𝘵 𝘣𝘶𝘪𝘭𝘵 𝘪𝘯 𝘪𝘴𝘰𝘭𝘢𝘵𝘪𝘰𝘯—𝘪𝘵'𝘴 𝘰𝘳𝘤𝘩𝘦𝘴𝘵𝘳𝘢𝘵𝘦𝘥 𝘵𝘰𝘨𝘦𝘵𝘩𝘦𝘳. What's your biggest data governance challenge? Drop it below. 👇

The new consulting edge isn't AI. It's knowing when your AI is wrong. Every consultant has been there: You ask AI to analyze documents and generate insights. During review, you spot a questionable stat that doesn't exist in the source! AI hallucinations are a problem. The solution? Implementing "prompt evals". → Prompt evals: directions that force AI to verify its own work before responding. A formula for effective evals: 1. Assign a verification role → "Act as a critical fact-checker whose reputation depends on accuracy" 2. Specify what to verify → "Check all revenue projections against the quarterly reports in the appendix" 3. Define success criteria → "Include specific page references for every statistic" 4. Establish clear terminology → "Rate confidence as High/Medium/Low next to each insight" Here is how your prompt will change: OLD: "Analyze these reports and identify opportunities." NEW: "You are a senior analyst known for accuracy. List growth opportunities from the reports. For each insight, match financials to appendix B, match market claims to bibliography sources, add page ref + High/Med/Low confidence, otherwise write REQUIRES VERIFICATION.” Mastering this takes practice, but the results are worth it. What AI leaders know that most don't: "If there is one thing we can teach people, it's that writing evals is probably the most important thing." Mike Krieger, Anthropic CPO By the time most learn basic prompting, leaders will have turned verification into their competitive advantage. Steps to level-up your eval skills: → Log hallucinations in a "failure library" → Create industry-specific eval templates → Test evals with known error examples → Compare verification with competitors Next time you're presented with AI-generated analysis, the most valuable question isn't about the findings themselves, but: 'What evals did you run to verify this?' This simple inquiry will elevate your teams approach to AI & signal that in your organization, accuracy isn't optional.

Many people, especially tech journalists and bloggers, don't understand the purpose of benchmarks in AI. Performing well on a benchmark doesn't prove anything but the fact that your AI works well on the types of problems this benchmark covers. If you work on a general-purpose AI, then it could perform well on one benchmark by chance. This is why we usually have multiple benchmarks, each focusing on a specific type of problem. Performing well on all benchmarks, made by different people testing different capabilities, is possible by chance, but this chance is very slim, close to 0.00000(...)0001. This is why we need many benchmarks. As a consequence, when our general-purpose AI performs well on multiple benchmarks made by different people to test different capabilities, we can safely extrapolate that it also performs well on other types of problems, not covered by benchmarks, because the chance that our choice of benchmarks and the range of problems our AI is capable of solving coincide exactly is very slim too. However! If you take a neural network and finetune it on the data from the benchmarks (or on a synthetic dataset that resembles the benchmark examples), then your model will perform well on all the benchmarks. But now we are not talking about performing well by chance; we are talking about performing well by design. And this defeats the whole purpose of the benchmarks. Now we cannot (and must not) extrapolate the performance of our neural network to other domains. This would be (and is) anti-scientific.