Long reads on AI for B2B.
Frameworks, deep dives, opinion. Field notes are the short ones — these are the essays.
Blog · long-form
AI Agents for Business in Ukraine: Who Builds Them, What They Cost, What to Check (2026)
What an AI agent does in a Ukrainian company, pilot prices in EUR, local specifics (language, CRM, telephony, liability) and how to pick a builder.
AI Agents for Small Businesses in Spain: Who Builds Them, What They Cost, What to Check (2026)
What an AI agent does for a Spanish pyme, pilot prices in EUR, the Spain-specific checks (VeriFactu, RGPD, AI Act, Kit Consulting) and how to pick a builder.
AI Consulting for Small Business: What You Get, What It Costs, and When to Skip It (2026)
What AI consulting for a small business should deliver, three pricing models and where the risk sits, what a good audit report contains, and when to skip it.
How to Choose an AI Agent Development Company for a Small Business (2026)
What an AI agent development company should deliver, what a pilot costs in EUR, seven questions that expose a prompt shop, and when not to hire one at all.
The agent made a mistake: who is legally liable
EU high-risk AI obligations were pushed to December 2027; the liability directive was withdrawn. Who answers for an agent's mistake and what to decide now.
Correcting speech recognition errors: 3 steps
Correcting speech recognition errors means targeted edits to the transcript, not a rewrite. A breakdown of Agentic ASR and why WER cannot see the difference.
AI voice agents in Ukrainian: where speech recognition breaks
Three public numbers on Ukrainian ASR were measured under different conditions, and none is about a phone call. What that means for deploying a voice agent.
Vertical AI vs Generic AI: Why Narrow Wins, and Why It Isn't the Model
Harvey is worth $11B and runs the same frontier models you can call. What vertical AI actually owns, and how to build those four layers at SMB scale.
What your process does when the model says no
LLM guardrails block legitimate requests and the process stops. The Hugging Face incident, vendor lock-in as a class of risk, and how to build a fallback route.
Multi-agent system failures: three patterns we found on our own board
Five months of Grow2.ai's multi-agent editorial system: not one failure happened inside an agent, all of them at the handoffs. Three failure patterns with numbers from our own board.
An autonomous AI agent: 14 hours instead of 2–17 weeks of work
Epoch AI and METR measured how long AI works without a human in the MirrorCode benchmark: a flagship model rewrote a program in 14 hours for $251. The Grow2.ai read on what carries over to SMBs.
The Every case: 25 people and four AI agents in management
Every handed prioritization, meetings, OKRs and growth reporting to four AI agents. What carries over to an SMB and what needs the data sorted first.