The Boring Bit That Actually Works

In this week's edition:

• The Boring Bit That Actually Works - Why companies with AI governance ship 12x more
• Quick Hits - Chrome gets Gemini, Karpathy converts, contract AI scales, DeepMind's Agentic Vision
• Tool of the Week - OpenClaw: my AI assistant, running from WhatsApp
• Weekly Tip - Start your AI agents on a short lead

The Boring Bit That Actually Works

Everyone wants to talk about agents. Nobody wants to talk about governance. Which is exactly why the companies getting governance right are the ones actually shipping.

Your AI Isn't the Problem. Your Process Is.

Two reports landed this week that, taken together, tell a story most AI vendors would rather you didn't hear.

Databricks published data from over 20,000 organisations. The headline numbers are big: multi-agent workflows grew 327% between June and October 2025. Nearly 8 in 10 companies now use at least two different AI model families. The "Supervisor Agent" pattern, where one AI orchestrates a team of specialist sub-agents, has become the dominant architecture. In database testing alone, 97% of environments are now built by AI agents.

Those are impressive stats. But the number that actually matters is buried further down.

Organisations using AI governance tools are putting 12 times more projects into production than those without them. Not 12% more. Twelve times.

Then Deloitte dropped their own data. Only 21% of organisations have implemented any form of governance or oversight for their AI agents. Meanwhile, 23% are already using agents, and that's expected to hit 74% within two years. The ones not adopting at all? Expected to fall from 25% to just 5%.

So the picture is this: nearly everyone's about to deploy AI agents, and almost nobody has the guardrails in place to do it properly.

The YOLO Problem

Sam Altman, of all people, demonstrated the issue perfectly this week. He'd just finished warning everyone about giving AI agents too much access without proper security infrastructure. Two hours later, he'd given Codex full access to everything. His words: "We're all about to YOLO."

It's funny. It's also the entire problem in miniature.

The temptation to skip governance is enormous because governance feels slow. It feels like paperwork. It feels like the thing you do after the exciting bit. But the Databricks data says the opposite. Governance is what turns experiments into actual products. Without it, you get a graveyard of proof-of-concepts that never leave the sandbox.

Deloitte's recommendation is sensible: "tiered autonomy." Start your agents with view-only access. Let them earn action rights over time. It sounds obvious, but almost nobody's doing it.

What This Means for You

If you're a SaaS business with 50 to 200 people, the message is clear. You don't need more AI tools. You probably have enough already. What you need is a framework for deciding what those tools are allowed to do, who's watching them, and what happens when they get it wrong.

The companies that sort this out now won't just avoid disasters. They'll move faster than everyone else. That 12x number isn't a coincidence. When you trust the process, you stop second-guessing every deployment.

I spend half my time telling clients to slow down and the other half telling them they're overthinking it. The truth, as usual, is annoyingly in the middle. Move quickly, but know what your AI is doing and why. If Sam Altman can't resist handing over the keys after two hours, you can assume your team won't either. Plan accordingly.

Quick Hits

Chrome just became an AI browser

Google is embedding Gemini directly into Chrome's sidebar, and it's not just a chatbot bolted on. It reads context across your open tabs, compares prices, drafts emails, and fills forms. The big one: "auto-browse" lets it navigate websites on your behalf, buying things, finding coupons, collecting documents. Rolling out to paid subscribers in the US first. If you use Chrome (and statistically, you do), your browser is about to change quite a bit. Read more

Karpathy went from "AI agents are useless" to "I code mostly in English"

Andrej Karpathy, who built Tesla's AI division, publicly called AI agents useless three months ago. Now he says 80% of his coding is agent-based and calls it "the biggest change to my basic coding workflow in two decades." The tools have crossed a threshold. Even the sceptics are coming round. Read more

Contract AI platform hits 1 million contracts a year

SpotDraft now processes over a million contracts annually, with volumes up 173% year-on-year. Qualcomm is backing them for on-device contract AI. This is the kind of AI adoption that doesn't make flashy demos but saves legal teams hundreds of hours. Contract review is, at this point, a solved problem. Read more

DeepMind teaches AI to actually look at images properly

Google DeepMind's "Agentic Vision" lets Gemini 3 Flash zoom, crop, and inspect images step by step instead of glancing once and guessing. A construction company using it improved blueprint checking accuracy by 5%. For any business dealing with visual documents, invoices, technical drawings, or product images, this is worth watching. Read more

This Week's Tool: OpenClaw - My AI Assistant, Running From WhatsApp

I've been writing about AI tools for months. This week I thought I'd tell you about the one I actually use every day.

OpenClaw (formerly Moltbot, formerly Clawdbot, which is doing a very convincing Prince impression at this point) is an open-source platform that connects AI agents to your messaging apps. WhatsApp, Telegram, Discord, iMessage. You message it like you'd message a colleague, and it messages back. No separate app, no browser tab, no login.

I used it to build Eric (named after Cantona, obviously), my own AI assistant. Eric lives in my WhatsApp. Every morning, he checks my emails, reviews my calendar, flags what needs attention, and tells me what's overdue. He drafts my newsletters (including this one). He chases people who haven't replied. He manages my task lists. When I tell him to send an email, he drafts it and waits for my approval before sending.

Setting it up took an afternoon. You install it on a server or old laptop, scan a QR code to link your WhatsApp, point it at an AI model (Claude, in my case), and give it a workspace with files that tell it who you are and how you work. It reads those files every session, so it remembers context between conversations.

The interesting part is the "workspace" concept. You write a few markdown files describing your preferences, your tools, your recurring tasks, and the agent works from those. It's not a one-size-fits-all chatbot. It's shaped around how you actually work.

Is it polished? No. You need some technical comfort to get it running. The documentation assumes you know what a terminal is. But once it's set up, it genuinely saves me hours a week. The morning briefing alone has changed how I start my day.

Verdict: Not for everyone yet. If you're comfortable with a bit of setup, it's the most practical AI assistant I've used. If "install on a server" makes you nervous, give it six months. This category is moving fast.

I've put together a guide on how I set Eric up and how you could do the same. If you'd like a copy, drop me a message and I'll send it over.

This Week's Tip: Start Your AI Agents on a Short Lead

Before you give any AI agent write access, run it in "shadow mode" for a week. Let it suggest actions without executing them. Log what it would have done. Review those logs daily.

You'll learn two things: where the agent adds genuine value (usually more than you expected), and where it confidently gets things wrong (also more than you expected). After a week, you'll know exactly which actions to automate and which to keep human-in-the-loop. That's your governance framework, built from evidence rather than guesswork.

It takes seven days. It saves you from the kind of mistake that takes seven months to clean up.

That's it for this week. If someone forwarded this to you and you'd like your own copy, subscribe at [Link]. If you're already subscribed and wondering why, fair enough.