Omdena #AI

Humans have an unfair advantage over AI: we know what isn't written down. You read a document and immediately know: “That's outdated.” “That decision was reversed.” “Ask Sarah — she owns this.” “That number doesn't look right.” An AI agent doesn't automatically know any of that. So the problem isn't always giving AI more data. It's giving AI the background context humans naturally understand. And that's a big part of what context engineering is about.

You can have 10 years of data and still not be ready for AI. The problem isn't always that data is missing. Sometimes: Two systems show different customer information. Important documents are outdated. Nobody knows which dataset is authoritative. The AI can't access the information it actually needs. Then the AI gives an unreliable answer. We change the model. Rewrite the prompt. Tune the retrieval. But AI can't make unreliable data reliable. The real work is making sure the right data is accurate, accessible, contextual, and trustworthy for the job you're asking AI to do.

The most dangerous AI failure might be the one nobody reports. No outage. No error message. No failed deployment. The agent is still answering users. But the answers are getting worse. Retrieval quality drops. Data becomes stale. The model starts making more corrections. Users quietly stop trusting it. Everything looks green on the dashboard. Until someone asks: Is the AI still doing its job? That’s the observability problem production AI teams need to solve.

Give an AI 500 pages of context and you haven’t necessarily made it smarter. You may have just made it harder to find the one paragraph that actually matters. Good retrieval is about getting the right information, not the most information. Sometimes the better AI system isn’t the one with the bigger model. It’s the one that knows what information to give the model in the first place.

“AI-ready data” is becoming a popular phrase. But what does it actually mean? Clean data? Governed data? Fresh data? Accessible data? Contextual data? The harder part is that there isn't one universal definition. Data that's ready for RAG might not be ready for an autonomous agent. The more responsibility we give AI, the more we need to think about the data behind it — its quality, context, permissions, freshness, and provenance. So perhaps the better question isn't: “Is our data AI-ready?” It's: “Is our data ready for what we're asking AI to do?” That distinction can change how organisations approach AI readiness.

1/6 Your AI can be running perfectly and still be failing. 99.9% uptime. No infrastructure alerts. Good latency. And yet the answers are getting worse. That's one of the biggest problems with monitoring AI in production. 🧵👇