Anatoly Volkhover

The fastest way to waste $150,000 on data? Hire one person and expect them to do everything. We see this mistake all the time in mid-market companies. The CEO, CIO, or CFO says: “We need data.” They spend $10k on JD, for some engineer, and then... This is how it usually plays out: - Day 1, they’re asked to “set up the warehouse from the ground up, build ETLs, pipelines, dashboards, AI, and get leadership real insights.” - Within weeks, they’re buried in ad-hoc requests and Jira tickets. - The data becomes an endless project or a duct-taped solution. 12 months later, execs still don’t trust the numbers, and your first data hire is burned out or gone. This isn’t a talent issue. It’s a hiring issue. Here’s how we guide mid-market clients (50–500 FTEs) to avoid the trap: 1. Define the Business Problem First. Don’t start with “we need Snowflake” or “build a dashboard.” Start with: What are the 2–3 business questions we can’t answer today that cost us money? Agree on the definitions aligning with company objectives. Data architecture without a data strategy is just a cost center. It's like building without a map. 3. Hire for the Right Skill at the Right Time. Not a “Data Scientist.” Not an “ML Engineer.” Your first role should almost always be an Analytics Engineer. A strong generalist who can wear a few hats and deliver quick wins but not carry the whole strategy alone. A big recommendation is engaging with fractional leaders or firms at the start. They’ll set you up with a foundation from the get-go and help avoid million-dollar mistakes. 4. Future hires Adding new labor won't make you data-driven if there is no strategy. So invest in fractional leadership, as most do it the other way when it's too late and deep in cloud costs and no guidance. Role #2 should be a Data Engineer or analyst most cases. Focused on integrations, scalability, and cost optimization. Only hire when there is enough logistical complexity and it is a core problem. 4. Set a 90-Day Delivery Window. First hire should not disappear for 12 months “building the stack.” They should deliver a trusted, high-value dataset model within the first few weeks. Within 90 days, assess if your data investments are “cost" or “growth.” Quick wins build trust → 5. Place Data in the Right Reporting Line. If your data hire reports into IT, they become ticket-takers. When they report to COO/CFO/CEO, they stay focused on ROI, not infrastructure. Honestly... Hiring one person to wear six hats isn’t efficient. But at the same time, hiring more engineers doesn't make you data-driven. It doesn't matter if there is no data strategy and leadership. Most data issues are related to process, adoption, and trust. Comment "Data Team" and I'll send a playbook on how to build data team for success. 🧬 Repost if you think wearing all the hats as a “role" is not ideal. → Follow for more insights on leveraging data and how to become a truly data-driven organization.

38

10 Comments

🚀 Unlocking Agentic AI: From MCP to A2A, ACP, ACNBP & Beyond Agent-based systems are no longer theoretical. They're here—and protocols are the glue that makes them collaborate, delegate, and scale. If you're building with AI agents, here’s a quick guide to the evolving communication stack that powers next-gen autonomy 👇 🔌 MCP (Model Context Protocol) Think of this as the USB-C port for agents. It standardizes agent-to-tool communication via JSON-RPC, letting agents access APIs, databases, or search—ideal for stateless tool use and injecting external knowledge. 🤝 A2A (Agent‑to‑Agent Protocol) Built for collaboration across boundaries. Developed by Google, now open under the Linux Foundation, A2A enables task delegation, capability discovery, and real-time coordination between agents via HTTP/SSE and JSON-RPC. 🧠 ACP (Agent Collaboration Protocol) Structured conversations between agents. ACP introduces intent-based messaging (like REQUEST, INFORM, PROPOSE), shared ontologies, error handling, and multi-modal payload support—ideal for high-trust, real-time interactions. 🔐 ACNBP (Agent Capability Negotiation & Binding Protocol) Enabling smart agent matchmaking. This newer protocol supports secure discovery, negotiation, and commitment exchange between agents—essential when agents need to select trustworthy collaborators on the fly. 🌐 ANP + AG‑UI (Emerging standards) ANP handles agent discovery in dynamic ecosystems. AG‑UI powers real-time agent-to-user streaming (SSE/WebSockets), making agents feel alive in UI environments. 🧩 How it all fits together Imagine an agent workflow: Agent uses ANP to discover peers Agents negotiate with ACNBP Tasks are delegated via A2A External tools are accessed over MCP Conversations flow using ACP UIs stream results in real-time using AG‑UI Each protocol plays a unique role—but together, they enable resilient, intelligent, multi-agent systems. 🎯 Why this matters ✅ Modular: Plug in/out tools and agents 🛠️ Secure: Discovery, binding, identity baked in 📡 Scalable: Works across local, cloud, and hybrid systems 💬 Multimodal: Text, code, audio, and visuals 📈 Enterprise-ready: Backed by industry adoption Are you exploring multi-agent architectures yet? Curious how MCP, A2A, or ACP could reshape your enterprise workflows? Let’s connect and compare notes—agentic design is moving fast, and collaboration protocols are the key to making it work. #AgenticAI #A2A #MCP #MultiAgentSystems #AIProtocols #AIArchitecture #AIEngineering #FutureOfWork #EnterpriseAI #LLMops #AutonomousAgents

37

2 Comments

Most AI agents fail for the same reason most documentation fails. They don't understand what the user actually needs in that moment. And even when they do, they don't have the right context loaded to help. The Diataxis framework breaks documentation into four modes: tutorials, how-to guides, reference, and explanation. Each serves a different user intent. Mix them up and you confuse people. The brilliance of Diataxis is recognizing that the same information, presented in the wrong mode, becomes useless. Agents need the same discipline. Both in how they respond AND in what context they load. Sometimes the user wants to learn. Walk me through this. Explain what's happening. The agent needs teaching context: examples, analogies, step-by-step breakdowns. Sometimes they want to execute. Just do the thing. Don't explain. Don't ask questions. The agent needs action context: schemas, APIs, current state, permissions. Sometimes they need reference. What are my options here? What fields are available? The agent needs comprehensive documentation loaded, not tutorials. Sometimes they want to understand. Why did this fail? What's the reasoning? The agent needs diagnostic context: logs, history, system state. Most agents default to one mode with one context. Usually "helpful assistant" mode where they explain everything all the time. Which is great when you're learning. And incredibly annoying when you just want something done. The best agent experiences switch modes AND load different context based on the situation. Short command? Execute mode, action context. Question mark? Explain mode, teaching context. Error state? Diagnostic mode, system context. This is why context engineering matters more than prompt engineering. It's not just about how you phrase the request. It's about what knowledge the agent has access to in that moment, aligned with what the user actually needs. Building great agents isn't just about what they do. It's about knowing how to show up: right mode, right context, right moment. #AI #Agents #ContextEngineering #DataEngineering

7

If you’re still arguing last-touch vs. MTA, you’re leaking margin. Use the right tool for the right job: Your stack: MTA for daily creative/audience optimizations (micro moves). MMM for quarterly budget allocation (macro moves). Incrementality tests to validate what the models “think.” Customer surveys to capture dark social & word-of-mouth. Last-touch as a simple sanity check for short cycles. Guardrails: Join everything on margin LTV + payback, not just revenue. Don’t judge MMM by daily swings; don’t judge MTA by offline truth. Run one lift test per quarter to keep everyone honest. Result: less noise, faster decisions, higher ROI. #DTC #Attribution #ecommerce #MMM #MTA #Incrementality

1

The "Meaning Layer" is no longer optional. 🧠 Most SaaS platforms are drowning in data, but they’re starving for context. Traditional keyword search only takes you so far—it finds what you typed, not what you meant. To build truly intelligent applications, you need to understand the relationship between data points. This is where Vector Embeddings come in. By turning text, images, and user behavior into mathematical vectors, we can unlock: * Semantic Search: Finding results based on intent, even if the keywords don't match. * Hyper-Personalization: Recommendations that actually understand user preferences. * Clustering & Discovery: Identifying patterns in your data that were previously invisible. I’m thrilled to announce that we have officially added Vector Embedding capabilities to dotDash! We’ve built this to be seamless, allowing you to transform your unstructured data into searchable, "AI-ready" vectors with just a few clicks. Check out this short demo to see it in action: 👇 [Video Attached] #VectorEmbeddings #SaaS #AI #DataEngineering #dotDash #fluentData #MachineLearning

10

Founder: "I want an AI agent to do my ARR forecasting." Me: "Show me your data infrastructure." The reality check: If your data is scattered across Stripe, HubSpot, and Google Sheets... You don't need more AI. You need a Data Warehouse. To build a "AI-Native" business, you have to fix your messy data first. The old way: ❌ marketing calculates spend in spreadsheets. ❌ product counts users in Mixpanel. ❌ revops tracks ARR in Stripe. → Problem: when data is disconnected, AI lacks context. Here's why that matters: 1. Conflicting definitions.    Marketing defines a "Lead" differently than Sales. The AI doesn't know which logic to trust. 2. Hallucinations. When an agent can't find a single source of truth, it guesses. Guessing = hallucination. 3. Token bloat. Your AI has to burn tokens trying to understand three conflicting data sources instead of answering the question. The new way: ✅ centralize data in a Data Warehouse. ✅ add context + definitions. ✅ deploy AI agents on top. → Result: consistent, accurate, and cheap answers. An AI agent is only as smart as the data it feeds on. If you can't trace a customer from ad click to renewal, neither can your AI. Fix the data first. The agents come second.

20

2 Comments

More BFCM insights on incrementality tests run prior to the holiday: - Total efficiency improved by more than 75% when including the latent impact from BFCM. For example, if a brand saw a $1 iROAS before BFCM, it increased to $1.75 during the post-treatment window. - For experiments that started prior to BFCM and tracked the PTW lag effect into BFCM, 41% of the incremental value of media showed up in the post-treatment window. - In 44% of experiments, efficiency doubled (or better!) when measuring the BFCM lag. - Delayed effects of advertising during BFCM are nearly triple as much compared to evergreen PTWs throughout the year. Outside of BFCM, the delayed lift during the PTW for these same brands added only 26% more incremental value, meaning the delayed effects of advertising are especiallysignificant around BFCM. - Overall, spend before BFCM – including spend from previous months before November – continued to generate conversions well into November. In other words, if a brand cuts spend based on Day 1 ROAS, they’re cutting the revenue that materializes on Day 30.

31

New book from Packt Publishing >> "The Profitable AI Advantage" — A business leader's guide to designing and delivering #AI roadmaps for measurable results: https://amzn.to/4oaRhtl by Tobias Zwingmann, with Foreword by Armand Ruiz   𝓚𝓮𝔂 𝓕𝓮𝓪𝓽𝓾𝓻𝓮𝓼: 🔵Simplify complex AI concepts with clear explanations tailored for business leaders 🔵Learn how to assess AI opportunities and align them with business goals for maximum impact 🔵Scope and build your first AI prototypes to gain traction quickly Readers will learn to: 🔵Identify and leverage AI opportunities in your business 🔵Create actionable AI roadmaps and implement them incrementally 🔵Scale AI operations to enhance business functions 🔵Design and evaluate AI use cases through feasibility assessments 🔵Recognize AI trends to quickly prototype and iterate AI solutions 🔵Manage AI projects and teams effectively without deep technical skills 🔵Understand fundamental concepts of modern AI, including Generative AI, LLMs, and predictive models See the book at: https://amzn.to/48oASgf

21

7 Comments