Alessandro Pireno

I am speaking soon at Tata Consultancy Services about AI and how to activate your data. I have been trying to polish the live demo that goes with my talk track. It showcases how to reach fragmented enterprise data in real time and it runs on Claude. The only problem is that Claude is currently down. So here is the pitch: come show up and see if we can get a demo to run at all. If the API gods are not smiling on us, I will resort to doing magic tricks and singing. Your loss either way. But seriously, we will be having a human to human conversation about the real bottleneck in enterprise AI. AI models are not magic. They are refineries. Without continuous access to the raw data, they will never produce the insights your business is counting on. I will be sharing how Attimis and OneBucket solve this exact problem to orchestrate across hybrid environments. It is a small room and a curated guest list designed for CDOs, CAIOs, CIOs, and data leaders. Grab one of the last minute spots here: https://lnkd.in/gPrUV9Bk

Grab a seat, then read on → https://luma.com/t8zmzm2l Every enterprise has an AI pilot. Almost none have it in production. What kills it is boring and brutal: the agent cannot reach the data that matters. Less than 10% of enterprise data ever gets to the models. That is the pilot trap. It is not a model problem, it is a data-access problem. You cannot prompt your way past data locked in another region, another system, another team's control. On June 24 in NYC we are not going to talk about it, we are going to show it. Claude reaching live enterprise data across three global locations, in real time, no copying, no migration. Real data, real demo, real questions, with TCS and Attimis. Built for the people who actually own this problem: CDOs, CAIOs, enterprise architects. Small room, going fast. Bonus for me; I personally can't wait to hang with Shayne Stubbs, Siddhartha Bala, Damian Kowalewski, Lakshmi Pingali. Honest question for you: what percent of your enterprise data can your agents reach today? Come see what the other side looks like. Seats → https://luma.com/t8zmzm2l

I have a new hobby. It is called whack-a-mole, and the moles are my AI vendors. Two weeks ago Anthropic was about to change the Agent SDK billing under me. I refactored, posted about it, they postponed. This week Google whacked the next one. Gemini CLI for individuals stopped serving requests on June 18. You sign in, get told the client is no longer supported, and get herded into Antigravity. No grace period, no feature parity at launch. It was one of the review engines in my stack, so it went dark mid-workflow. I swapped that step to Claude sub-agents, changed the one line that names the engine, and kept moving. Fix time: an afternoon. The model wrote its own migration. Here is the part the hyperscalers have not priced in. Every one of these moves is an attempt to control how you use the tool. Deprecate the CLI, herd you into the new suite, retune the billing. That playbook only works when switching is hard, and LLMs made switching trivial. The same model they want you locked into is the one that writes your escape hatch before lunch. They are running a lock-in strategy on top of the exact technology that dissolves lock-in. And let's name what the Gemini move actually is. Push individuals off a clean headless CLI and onto a consumer suite, where the free tier's deal has always been the old one: you are not the customer, you are the training set. That bargain works great when you have no alternative. Developers have alternatives now. Forcing the migration with worse parity and no grace period is not a nudge, it is a tell. Contrast the two this month. Google is spending developer goodwill to capture behavior and data. Anthropic walked back a pricing change specifically to keep developers onside. In a market where the only durable moat is developer mindshare, one of those compounds trust and the other compounds resentment. The gap does not show up in a quarter. It shows up when the next default gets chosen. For individuals, all of this is a shrug. Long gone are the days a deprecation meant real downtime. For enterprises it is a warning. Change the surface people built on at this clip, with no parity and no grace period, and it stops reading as innovation and starts reading as instability. Procurement has a long memory. If your stack breaks because a vendor changed a product page, you do not have a stack. You have a dependency wearing a stack costume. Build engine-agnostic, keep every call swappable, and let the next mole pop up. Whack-a-mole is only stressful if you are attached to one mole.

*FOR NOW. Anthropic just postponed the Agent SDK billing change that was supposed to land today. I spent the last few weeks refactoring around it. Jumped the gun. Reprieve granted. Postponing a billing change at the last minute is not free. It signals they are trying to get the model right, not just ship it. Respect. I am keeping the architecture. Pricing whiplash is not a one-off, it is the operating environment. The question is not whether the bill changes. The question is whether your stack survives when it does. Mine does now. Billing-model agnostic. Skills, loops, hooks, multi-agent orchestration, switchable between subscription and API, fully observable. If subs survive, it works. If they get metered, it works. If the heavy models move to API, it works. Same window I shipped this, Anthropic's Boris Cherny named hooks, loops and skills the most powerful Claude Code features, and Karpathy described the spec, verifier, environment loop. Three independent reads of the same shape. I genericized the whole pattern and pushed it as an open template. Fork it, run it, save yourself the next refactor: https://lnkd.in/g-i_nbHA One ask for whoever is reading this at Anthropic. Keep claude -p as a first-class entry point. It is the cleanest hand-off into an agent stack on macOS. The osascript alternative steals focus every time it fires, which is a non-starter for anything that runs on a schedule or in the background. claude -p stays in its lane, returns a value, and gets out of the way. That is the right shape for a tool that lives inside someone's real workflow. If your AI stack can break because of one line on someone else's pricing page, you do not have a stack. You have a dependency wearing a stack costume. Pricing changes are temporary. Vendor whiplash is permanent. Architect for both.

Your AI strategy can be shut off by a memo you'll never see. Not hypothetical. A frontier model got yanked this week. For everyone. Overnight. The vendor flipped the switch to comply, and hundreds of millions of users were gone by lunch. Sit with that. The smartest thing in your stack is a rental, and the landlord answers to people who've never heard of your roadmap. Here's what nobody wants to say out loud: you don't have an AI strategy. You have an AI dependency. If a government, a vendor, or a board can switch your intelligence off, it was never yours. "But it's the best model" is exactly how you end up hostage to it. The fix isn't a bigger model. It's owning the one thing no one can revoke: your data, and the pipes that move it. Control that, and a small, swappable, even on-prem model takes you further than your competitor's frontier model does, right up until the morning theirs goes dark. The model is borrowed. The data is yours. Build like you mean it. So be honest: are you running an AI strategy, or an AI dependency?

We treat AI models like magic. They're refineries, and a refinery with no oil produces nothing. The hard part of enterprise AI was never the model. It's that most of your data never reaches the agent. It stays locked in the object stores and systems it already lives in, and that's where pilots die on the way to production. On June 24 I'm speaking in NYC at a small, candid evening hosted by Tata Consultancy Services and Attimis for data and AI leaders. I'll walk through the nuts and bolts of the AI data pipeline and run a live demo with Claude on a real IT use case, hitting fragmented enterprise data in real time at 40X efficiency and 10X lower latency. No vendor pitch, no handwaving. Just me, a real problem, and a room of people who deal with it. The leaders who solve data access first are the ones whose agents reach production. Come join me and Shayne Stubbs to hear how. And bring your hardest questions! 👇 📍 NYC · June 24 · https://luma.com/t8zmzm2l

If a frontier lab shipped your core feature tomorrow, ten times cheaper, are you dead or ten times better? Most AI strategies cannot answer that, because they are built on the wrong asset. The database everyone calls a moat is table stakes now. The real moat is the data nobody put in the warehouse. The two questions at the end are worth asking your team Monday morning.

I’m excited to share that I’ve joined Attimis Corporation as an Advisor. As the industry races to secure compute for AI, the real bottleneck has shifted to data delivery. It’s no longer just about storing massive datasets; it’s about Just-In-Time data management at petabyte scale. Attimis is solving exactly this, ensuring the right data hits the right GPUs at the right time. I’m looking forward to bringing my experience in AI and data architecture to the table to help refine their technical and GTM strategies. A huge thank you to Damian Kowalewski and Shayne Stubbs for bringing me into the fold, and thanks George Yakovlev for guidance on the tech and architecture side. Excited for what we are building!

🚨 Seeking a Venue for NY TechWeek Our original venue fell through, so we are looking for an ecosystem partner, VC firm, or tech company interested in hosting a 50–100 person event for an evening session during TechWeek. Yong-Gil Son and I will be hosting and presenting primary research and an actionable framework on how B2B tech companies survive the transition from simple AI wrappers to enterprise-grade systems of action. 📊 The Presentation: The Agentic Company Playbook: Why Data is Still the Moat. We are shifting rapidly from AI that merely observes to AI that executes. However, 95% of GenAI pilots are currently stalling because large language models hit a ceiling when dealing with complex, proprietary enterprise schemas. This talk covers the structural shifts happening right now at the Series B milestone, including: The SQL Ceiling: Why LLMs hit 90% on generic benchmarks but drop below 50% on enterprise data architectures. Shadow Context: How the true AI moat is moving away from basic customer data toward proprietary interaction logs and execution histories. The Grounding Dividend: Transitioning from probabilistic guessing to verifiable execution via Knowledge Graphs and the Model Context Protocol (MCP). The Token Paradox: Why enterprise AI bills are rising by 320% even as individual token unit costs crash. 🏢 Venue & Crowd Requirements: Capacity: Space to comfortably host 50–100 attendees. Vibe: A space suitable for an after-work presentation followed by casual networking. Location: NYC What's in it for you? Co-branding, exposure to a high-caliber crowd of founders, technical executives, and operators, and a chance to support the local tech ecosystem during its busiest week. If your office or portfolio space has availability next week and you'd like to co-host, drop a comment below or DM us directly! 📥 Want a copy of the whitepaper handout ahead of time? Let us know and we'll send it over.