Transform AI Policy from Handbook to Living Practice

By giving machines human skills, we make them smarter. But do we become better ourselves? Today, a lot can be delegated to AI: writing, analyzing, shaping thoughts… It’s convenient. Fast. Efficient. But what is the side effect of this? Is it only important what the tool gives? And what it takes away — is that not important? What happens to our skills, depth of thinking, understanding of what we do, our authenticity? For example: when was the last time you wrote by hand instead of using a device? Is that skill still easy for you? Where is the line for you between assistance and replacement?

How do you 8x your output without losing quality? It’s not just about AI. It’s about the "Paper-First" approach. • The Problem: High-end tools often lead to "aimless editing" and talking in circles. • The Solution: A pencil and a blank page. • The Formula: Use the analog brain to build the format, then use the digital tools to execute the vision. Experience gives you the talent, but a solid formula gives you the speed. When you marry decades of production skill with modern tech, the "impossible" timelines become your new standard.

One thing I’ve learned from building AI products: people do not buy “intelligence.” They buy a clearer result. Faster support. Better follow-up. Less manual work. More consistency. Everything else is packaging.

Overloading an AI agent with every piece of company information is one of the most common mistakes businesses make, according to Matthias Lübken of TAVON.ai. The field addressing this is called context engineering, which focuses on giving agents the right instructions at the right point in time. The typical failure mode Matthias describes is companies feeding an agent their entire marketing strategy, sales strategy, and company background at once and expecting it to perform. That approach does not work. The better method is to be specific and structured. Matthias compares it to writing a longer message to a language model, but organized as a markdown file. Different tools implement this differently, but the core idea is the same. Agents need targeted context, not an information overload.

Most AI outputs aren't wrong enough to catch. That's the problem. In two or three years, the experts who win won't be the ones who automated the most — they'll be the ones who figured out which work actually needed to be right. Lead lists where a hallucinated field costs a deal. Research where a fabricated source costs credibility. Data where 3% error means the whole thing is garbage. Tendem by Toloka is building the infrastructure for that layer. AI handles the speed. A vetted expert handles the truth. The output is signed off on before it reaches you. Verified work has a pace. It turns out that pace is worth paying for.

Engineers don't need AI to read a datasheet. Our first feature let you upload a datasheet and ask it questions. Sounds useful. Except it took 4x longer than just finding the answer yourself. Turns out humans are good at reading datasheets. After a few months of practice you're trained up enough to skim 100's of pages in under three minutes. Reading one datasheet wasn't the problem. The problem is humans only have one set of eyes, and can only read one datasheet at a time. But when you're picking a part, you're comparing hundreds of options across different manufacturers, each with different structures and different ways of presenting similar information. That's where AI actually helps. Not reading one datasheet. Searching thousands of them at once. That's what Zenode does now. Think of it as spinning up hundreds of "eyes" on your behalf 👀👀👀👀👀👀 What's the most tedious part of your component search process?

I don't think the AI-bros understand what the word "efficiency" means. They use the term constantly, but they're only focused on what AI can do for quantity. They brag about generating a thousand landing pages. A 500-email drip. A "create blog" button that fires off posts at scale. But efficiency requires quality. It's about progress toward a goal. Just like five years ago, "more stuff" is rarely the unlock. It's just the most obvious thing to do with a tool that's good at producing things. Efficiency isn't only about quantity. It's also about quality. It's something like: Efficiency = (Quality × Quantity) ÷ Input AI can compress the time it takes to do something well. The real question is what you do with the time it saves you. In most cases, the answer is more iterations, sharper testing, and tighter analysis on the work you were already going to do. Quantity isn't unimportant. It's just not the only part of the equation. If you're judged on outcomes, 100x more stuff is rarely the answer.

A question I'd ask before buying any construction AI: what does it actually remember, and can you trust that memory? Most of what's marketed as AI in this industry is search with better manners — it answers faster, but it's reaching into the same fragmented pile of files, spreadsheets, and systems that don't talk to each other. Faster access to messy data isn't intelligence. It's just a louder version of the problem. The shift that matters has nothing to do with chatbots. It's whether your capital project has one verified record — a single source the whole team works from — instead of a dozen competing ones. Get that right and the useful things become possible: catching a schedule slip early enough to do something about it, handing the repetitive paperwork to something that can actually take action, and using what past projects taught you so the same mistake doesn't cost you twice. That's true whether you run Procore, something else, or three things stitched together. The foundation comes before the agent, every time. Good read on this.

Struggling to find the AI quick-win that *actually* lasts? It's not about the tech, it's about the *people* and the *problem*. Focus on a clear business need and ensure buy-in. Then, watch the magic happen! ✨ → Learn more about our AI Readiness Assessment at unleashingdata.com

Thinking about simplifying what makes a functional AI scaffolding and how to think about building it up. Seems like you can reduce most approaches into three components: 1. A memory system: Contains relevant context about the domain the AI is operating on. What's going on in your business, your projects, your people. Most people use markdown files (CLAUDE.md, AGENTS.md). More complex setups exist depending on the use case. 2. Skills: a specific workflow or specific knowledge, encoded once and reused when the task calls for it. Drafting a newsletter, closing the books, processing call notes into next actions. 3. Bespoke software: where the work can be made fully deterministic, just have the AI write code. The model is great at judgment and synthesis, bad at anything that needs the same answer every time. If a script can handle it reliably, write the script.