How Poor Data Affects AI Results

𝐑𝐮𝐥𝐞 𝐨𝐟 𝐭𝐡𝐮𝐦𝐛: 𝐀𝐈 𝐬𝐮𝐜𝐜𝐞𝐬𝐬 𝐢𝐬 𝟐𝟎% 𝐦𝐨𝐝𝐞𝐥 𝐚𝐧𝐝 𝟖𝟎% 𝐝𝐚𝐭𝐚 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞. I've reviewed dozens of failed AI initiatives. The pattern is often the same. The POC worked beautifully on clean, curated sample data. Then production happened. 𝐓𝐡𝐞 𝐝𝐚𝐭𝐚 𝐰𝐚𝐬𝐧'𝐭 𝐰𝐡𝐞𝐫𝐞 𝐭𝐡𝐞𝐲 𝐭𝐡𝐨𝐮𝐠𝐡𝐭 𝐢𝐭 𝐰𝐚𝐬. The real records are split across many systems. Product data with three different schemas. No master data management. 𝐓𝐡𝐞 𝐝𝐚𝐭𝐚 𝐰𝐚𝐬𝐧'𝐭 𝐰𝐡𝐚𝐭 𝐭𝐡𝐞𝐲 𝐭𝐡𝐨𝐮𝐠𝐡𝐭 𝐢𝐭 𝐰𝐚𝐬. Labels were inconsistent. Historical records are incomplete. Business rules encoded in people's heads, not systems. 𝐓𝐡𝐞 𝐝𝐚𝐭𝐚 𝐰𝐚𝐬𝐧'𝐭 𝐫𝐞𝐚𝐝𝐲 𝐟𝐨𝐫 𝐬𝐜𝐚𝐥𝐞. Batch pipelines that took 8 hours. No real-time feeds. No versioning for training data. Many AI teams report spending around 80% of their time on data preparation. That's not an AI failure. It's a failure of data infrastructure. And here's the painful part: surveys have found that around 42% of data scientists say their results aren't used by business decision makers. The models work. The trust doesn't. The organizations winning at AI aren't the ones with better models. They're the ones who fixed their data platform first. Before your next AI initiative, ask: - Do we have a single source of truth for this domain? - Can we access this data reliably at scale? - Is the data quality sufficient for production decisions? 𝐖𝐡𝐚𝐭'𝐬 𝐭𝐡𝐞 𝐛𝐢𝐠𝐠𝐞𝐬𝐭 𝐝𝐚𝐭𝐚 𝐠𝐚𝐩 𝐛𝐥𝐨𝐜𝐤𝐢𝐧𝐠 𝐲𝐨𝐮𝐫 𝐀𝐈 𝐚𝐦𝐛𝐢𝐭𝐢𝐨𝐧𝐬?

Bad data doesn’t break loudly. It smiles, gets approved, and quietly ruins decisions. That’s the real problem. You know what’s worse than broken pipelines? Pipelines that run flawlessly while feeding you lies. Data that passes checks. Gets signed off. Ships to dashboards and models. Then nudges pricing, forecasts, and product calls just far enough off to burn real money. That’s how companies lose trust without noticing. Data quality isn’t about chasing perfection. It’s about one question: Would you stake your reputation on this number? If the answer is no, the rest doesn’t matter. Quality comes down to basics. → Does the data reflect reality? → Is anything missing that would change a decision? → Does the same field mean the same thing everywhere? → Is it recent enough to act on? → Does it follow the rules you agreed on? → Are duplicates quietly inflating counts? → Break any one of those and your conclusions wobble. This is why bad data hurts more than bad tools. Gartner pegs the average loss at $12.9M per year. Revenue slips. Forecasts drift. Models learn patterns that don’t exist. When AI fails, it’s rarely the math. It’s the input. Ownership doesn’t sit with one team either. Business sets the rules. Engineers build checks. Stewards watch signals. Source teams keep inputs clean. If everyone shrugs, quality decays by default. Start small. Track freshness, row counts, schema changes, missing values, duplicates. Turn them into simple checks and run them every time. Stable beats perfect. Check early. At intake. During transforms. Before storage. Before anything touches a report or model. Late checks assign blame. Early checks save you. Data quality isn't a project. It's a reflex. It’s a habit. Visible. Automated. Shared. If you're not measuring it, you don't control it. If you don't control it, it's controlling you.

#UltrasoundAI didn’t struggle because the models were weak. It struggled because the images were. For years, we kept asking smarter algorithms to interpret noisy, operator-dependent ultrasound data and then blamed AI when results didn’t translate clinically. I was proud to be part of a newly published, peer-reviewed study with the Mayo Clinic that took a different approach: Fix the data foundation first. Using over 62,000 real-world breast ultrasound scans, we evaluated what happens when raw B-mode images are combined with quality-improved and enhanced ultrasound representations before AI ever sees them. Here’s what stood out: • Same patient. Same scan. Different outcome. Multifeature ultrasound improved diagnostic accuracy by +66% versus B-mode alone. • One size doesn’t fit all in medicine and that’s a good thing. • Graph models prioritized sensitivity (91.7%) for screening • Masked autoencoders prioritized specificity (100%) for diagnosis • The breakthrough wasn’t “better AI.” It was giving AI clearer, more consistent signal to work with. • This matters beyond academic centers. Image quality has been the limiting factor for AI at the point of care. Improve the signal, and AI becomes usable where patients actually are. This work reinforced something clinicians already know intuitively: You can’t out-optimize poor inputs. If we want AI diagnostics to scale safely, equitably, and responsibly, image quality has to be treated as infrastructure not an afterthought. Full disclosure: I serve as an advisor to #PONSAI, and it’s been encouraging to work alongside leaders like soner ozkan and Ilker Hacihaliloglu who are focused on strengthening the foundations of clinical AI, not just chasing benchmarks. Where do you see AI breaking down most today the algorithm, or the data we’re feeding it? #HealthcareAI #MedicalImaging #Ultrasound #ClinicalAI #DigitalHealth #AIinMedicine #PointOfCare #HealthTech #AIValidation Eric Topol, MD this ironically your post today.

Everyone’s talking about AI models, but here’s the truth most overlook: Your AI is only as smart as your data. As the founder of DataGardener, I’ve seen AI transform how #businesses operate—but I’ve also seen promising models fall flat because the data wasn’t good enough. Why Data is the Real Power Behind AI Algorithms don’t work magic. They learn patterns from data. So if your data is: ✔️ Outdated ✔️ Incomplete ✔️ Inaccurate …you’ll get flawed predictions and risky decisions. No matter how advanced the model. #AI learns from patterns. The more diverse and representative your #dataset, the better your models can generalise to real-world scenarios. Two Things Every Business Needs: 1. Accuracy "Garbage in, garbage out" is real. Clean, correct data is the only way to get trustworthy insights. Insufficient data doesn’t just mean bad business—it can lead to bias, compliance risks, and lost revenue. 2. Data Volume More data = better pattern recognition. Large datasets make models more robust, less prone to overfitting. #Diversity in data ensures insights reflect reality—not just a narrow view. How Key Data Attributes Impact AI Quality: #Accuracy → Produces trustworthy, actionable results #Volume → Enables richer insights and model resilience Real-World Impact Real-World Impact At DataGardener, our clients use AI built on verified, comprehensive company data. That’s how they: Make smarter credit decisions Uncover leads others miss Mitigate risks before they become costly The difference? It’s the data. Takeaway for Business Leaders Treat your data like an asset—not a byproduct: invest in data collection, cleaning, and validation. Before chasing the next AI model, fix your foundation. Remember: AI is only as good as the data it learns from. In the age of AI, data stewardship isn’t just IT’s job—it’s a boardroom priority. Curious how high-quality data can power better AI decisions in your business? Let’s talk. Let’s build smarter—starting with the right data. #SmartData #AIDrivenDecision #Data #BusinessLeader #ComplianceRisks #CreditDecisions #AIDecisions

Garbage in, garbage out. Picture trying to do your taxes using receipts jammed into an old, messy filing cabinet. Nothing’s labelled. Some receipts are handwritten. Others are duplicated. One’s in a different currency.   You could add it all up. But your final numbers? They're probably wrong.   That’s what it’s like when AI systems are trained on poor-quality data.   If the input is a mess the output just won’t be trustworthy. And that's true even if the algorithm is super advanced.   In real life, this can look like: → A model using old or inconsistent customer data → A chatbot trained on forums full of misinformation → A hiring tool learning from biased, incomplete records   The problem is, the AI output may look polished. It might even sound confident. But when the foundation is flawed, the results will be too. So, if you're starting to look beyond ChatGPT to custom workflow automations and AI agents, then you're going to need to pay attention to your data. What does good data practice look like? → Clean and check data first → Flag gaps, errors, and duplicates → Keep inputs current and consistent → Involve people who truly know the data   Even the smartest AI won't be able to make sense of a filing cabinet of chaos. Well, perhaps that's a challenge...   ⚛️ I’m Sarah Mitchell, PhD, AIGP and founder of Anadyne IQ. I help teams build AI literacy, develop smart adoption strategies, and manage risks responsibly. Follow along for regular AI governance stories, news, and insights.

You can’t outsmart bad data. Not with agents. Not with GenAI. Not even with humanoids that salsa and write poetry. If your data is trash, your AI will just make… 💩 but faster. 💩 but with a unicorn emoji. 💩 at scale. Everyone’s chasing copilots, agents, and AGI. But skip the “boring” stuff like data quality? And your AI becomes a very expensive fiction writer. Here’s what most don’t realize: 🔹 Machine Learning needs labeled, trustworthy data 🔹 Generative AI needs high-context, human-like data 🔹 Agentic AI needs clean, structured, real-time data 🔹 Superintelligence? Still needs a foundation it can trust What to do instead: → Map your messy sources → Define “good enough” for each use case → Align teams on what “clean data” actually means → Govern it like your product depends on it (because it does) AI doesn’t fail because the models are broken. It fails because the data is. 📌 Save this if you’re building or cleaning up after AI. 🔁 Repost to save a team from making a $500k data mistake. 🧭 Follow Gabriel Millien for practical, no-hype AI strategy and insights. Visual credit: Eduardo Ordax

If you’re not collecting the right data, AI will only accelerate your mistakes. Because AI doesn’t make you smarter. It just scales whatever assumptions you feed it good or bad. Everyone loves to talk about model performance, neural networks, and GenAI breakthroughs. But in the real world in industries like transport, logistics, healthcare, and public safety AI is only as powerful as the data foundation it sits on. If your data is noisy, biased, or incomplete your AI will be too. And when decisions are automated, those mistakes happen at speed and scale. As per Global context: – $100B+ lost annually due to bad data across industries (IBM) – 85% of AI projects fail to move beyond proof-of-concept (Gartner, 2023) – 96% of companies face data quality issues that impact AI performance (MIT Sloan) – In safety-critical domains, a false prediction from bad data can mean human lives at risk, not just business inefficiency While in Saudi Arabia: – As Vision 2030 accelerates AI adoption across sectors, from SDAIA to smart city infrastructure - clean, localized, context-aware data is becoming the bottleneck – Most imported datasets don’t reflect the unique driving patterns, heat, culture, or behavioral nuances of KSA’s real-world environments – Without investment in local data ops, many AI tools here will remain impressive in demo - and risky in deployment So, what can be done? – Map the right problem first. Don’t collect “big data,” collect the right data. – Label for local context. Your AI can’t make sense of behavior it wasn’t trained to see. – Build feedback loops. Your AI should learn and evolve from real-world conditions. – Govern your data like its product. Because it is. Garbage in, garbage out still applies. AI isn’t dangerous because it’s too powerful. It’s dangerous when it’s trusted too early on the wrong data. If you're not disciplined about your data today, your AI won’t be intelligent tomorrow. 📌 Do you use AI? If yes, for which purpose. ♻️ Repost to share insights with your network.

Garbage In = Garbage Out — No Matter How Smart the AI Is! It doesn’t matter if you're using ML, GenAI, or autonomous agents - if your data is bad, your results will be worse. Here’s how it breaks down: 1. Machine Learning Messy CSVs + algorithms = slightly organized garbage. You get patterns, but they're built on noise. 2. Classical AI Flawed data leads to polished interfaces but wrong decisions — the system appears smart but makes poor calls. 3. Generative AI Even LLMs can't fix low-quality inputs. They’ll give you impressive-sounding nonsense — creative garbage at scale. 4. Agentic AI This is where it gets risky. With bad inputs, agents don’t just make bad calls — they act on them autonomously. That’s automated chaos. The lesson? Good data hygiene isn't optional, its mandatory. You’re not just training a model - you're teaching it how to think and act. Save this if you're working with AI systems and want to avoid scaling the wrong signals.

Most industrial AI projects fail. Not because of the algorithm. Not because of the data scientists. They fail because of data quality. After more than 30 years working in manufacturing and automation, I’ve seen the same pattern repeatedly. Companies invest heavily in: • data lakes • AI platforms • digital twins • advanced analytics But they often skip the most basic question: Is the underlying data actually reliable? In many plants today you’ll find: • dead historian tags • stale sensor values • flatlined transmitters • poor compression settings • missing engineering context In one industrial dataset we analyzed recently, more than 20% of tags had quality issues. Yet organizations expect AI models to generate meaningful insights from this data. AI cannot fix bad inputs. If Industry 4.0 is going to deliver real value, data quality has to come first. Curious how many companies are actually measuring the quality of their industrial data today.

𝟰𝟯% 𝗼𝗳 𝗔𝗜 𝗽𝗿𝗼𝗷𝗲𝗰𝘁𝘀 𝗳𝗮𝗶𝗹 𝗯𝗲𝗰𝗮𝘂𝘀𝗲 𝗼𝗳 𝗱𝗮𝘁𝗮 𝗾𝘂𝗮𝗹𝗶𝘁𝘆 Yet most organizations spend 80% on models and 20% on data. Your AI is only as smart as your data is clean. The pattern repeats across industries 👇 📊 𝗧𝗵𝗲 𝗗𝗮𝘁𝗮 𝗤𝘂𝗮𝗹𝗶𝘁𝘆 𝗖𝗿𝗶𝘀𝗶𝘀 Informatica's 2025 CDO survey found: ➜ 43% cite data quality as #1 obstacle to AI success ➜ 57% report data is NOT AI-ready ➜ Only 5% of organizations have comprehensive data governance 📉 𝗪𝗵𝗮𝘁 𝗕𝗮𝗱 𝗗𝗮𝘁𝗮 𝗟𝗼𝗼𝗸𝘀 𝗟𝗶𝗸𝗲 The data exists but: → Lives in 47 different systems with no integration → Uses inconsistent formats and definitions → Contains unknown biases that propagate through AI → Lacks lineage—nobody knows where it came from → Has quality issues discovered only after deployment Gartner predicts 30% of GenAI projects abandoned by end of 2025 due to poor data quality. 𝗧𝗵𝗲 𝗗𝗮𝘁𝗮 𝗘𝘅𝗰𝗲𝗹𝗹𝗲𝗻𝗰𝗲 𝗙𝗿𝗮𝗺𝗲𝘄𝗼𝗿𝗸 Organizations achieving production AI allocate 50-70% of timeline and budget to data readiness. Here's what they build: 1. 𝗖𝗼𝗺𝗽𝗿𝗲𝗵𝗲𝗻𝘀𝗶𝘃𝗲 𝗔𝘀𝘀𝗲𝘀𝘀𝗺𝗲𝗻𝘁 Completeness: Do you have sufficient volume? Accuracy: Is the data correct? Consistency: Do definitions match across systems? Timeliness: Is data current enough for decisions? Validity: Does data conform to business rules? 2. 𝗟𝗶𝗻𝗲𝗮𝗴𝗲 & 𝗣𝗿𝗼𝘃𝗲𝗻𝗮𝗻𝗰𝗲 For every data point: Where did it originate? How was it transformed? What systems touched it? When was it last validated? You can't trust AI you can't trace. 3. 𝗕𝗶𝗮𝘀 𝗗𝗲𝘁𝗲𝗰𝘁𝗶𝗼𝗻 & 𝗠𝗶𝘁𝗶𝗴𝗮𝘁𝗶𝗼𝗻 identify: Sample bias (unrepresentative training data) Historical bias (past discrimination baked in) Measurement bias (flawed data collection) Aggregation bias (combining incompatible data) Then engineer mitigation before deployment. 4. 𝗔𝗜 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 requires: Model-specific data requirements documentation Continuous data quality monitoring Automated drift detection Regular revalidation cycles 5. 𝗗𝗮𝘁𝗮 𝗣𝗿𝗲𝗽𝗮𝗿𝗮𝘁𝗶𝗼𝗻 𝗜𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 Build platforms that enable: Extraction from source systems Normalization and transformation Quality dashboards with real-time monitoring Retention controls meeting compliance requirements API access for AI consumption Data readiness is NEVER "complete." It's continuous discipline requiring dedicated ownership. The Data Excellence Test: Ask yourself these questions: ✓ Can you trace any data point from source to consumption? ✓ Can you explain its quality metrics and bias profile? ✓ Do you have automated systems detecting data drift? ✓ Can you demonstrate data governance to regulators? ✓ Do you spend more on data infrastructure than AI models? If you answered "no" to any of these, you're building on quicksand. ♻️ Repost if you've seen AI fail due to data problems ➕ Follow for Pillar 4 tomorrow: Governance & Risk 💭 What percentage of your AI budget goes to data readiness?