The best use of AI is as little use of AI as possible

Friends,

your weekly AI briefing is here - designed to help you respond to AI, not react to the noise. No curveballs. No chaos. Just clarity.

This is a whole day on AI in which you will not touch AI. Knowledge used to be power. As machines take on more of the knowing, the value moves to what they can't do - judgement, presence, relationships, the willingness to put your name to a decision. None of that gets practised on a laptop.

Nobody knows what this technology can do yet - not the analysts, not its own builders. It's startling at things you'd never guess, hopeless at things you'd expect, and that uncertainty lands in an organisation as fear. AI has become a lightning rod for plenty besides: the future arriving, for some; everything they already distrust about work - faster, leaner, less human - for others. Waiting for agreement is not a strategy. Neither is sitting it out. Nor is forcing the pace from the top, past the front line where the real experiments are happening. On AI, there's no innocent position - which is exactly why the day runs on games, play and old-fashioned storytelling: how you make real moves when nobody can tell you what's possible, how you take a stance that's authentic and practical, how you stay connected to people moving at very different speeds. No slides, no laptops.

These decisions are hard, and AI is divisive. Spending a day with other humans doing human work is the way through. For leaders, and the people carrying AI change for them.

(Disclosure: Mayvin are AI Optimist partners - they work the people side of AI adoption alongside us.)

📰 This was the week AI cleared the compliance and data rules that used to keep it out

Ben and I have called this for eighteen months. Since early 2025, across hundreds of client conversations and cohort after cohort at AI Night School, we've held one position: use today's most capable AI while it's priced well below what it costs to run, and build so that the day it makes sense to run your own instead, you're ready for it. That was a patient bet for a long stretch. It's now the live conversation in nearly every client meeting we have, at every size of business - alongside a harder one, which is how an organisation actually agrees on what AI is worth to it.

This week is why the patience is paying off. Anthropic extended its Compliance API - the feature that gives IT and compliance teams a full audit trail of everything an AI touches - to Claude Cowork and Claude Code, its tools for team collaboration and coding, and opened Claude Code to run on a business's own servers instead of Anthropic's, in public beta, just days later. The paperwork an enterprise needs to say yes, and the option to keep everything in-house, arriving within a week of each other. Amazon's agentic AI, called Quick, reached AWS GovCloud too - the locked-down version of Amazon's cloud built for US government agencies that can't compromise on where data lives. If agents are landing there, they can land in your compliance department.

Mistral shipped in-region inference - AI processing that never leaves a country's own servers - alongside open models built specifically for sovereign AI. It's the one worth pausing on: the reminder that this movement runs well beyond America's borders too. For plenty of organisations, that isn't a nice-to-have. The health charities I work with have to keep patient data on UK soil, no exceptions, whatever model happens to be the best that month. Until now, using the strongest AI available and keeping that promise were two separate choices - and that trade-off didn't only complicate the paperwork, it shaped what the work could do and how it got done. Mistral shipping AI that runs and stays inside a country's own borders means, for the first time, they don't have to pick one.

And joining them, last, was xAI - a company whose approach to AI safety has drawn real scrutiny - with Grok Bot, agents pitched explicitly as teammates you hand ongoing work to and trust to run with it. Four companies, one direction, and it happened in the same week.

Four weeks ago this newsletter asked where to point AI - one process, one owner, one metric. That question hasn't gone anywhere. This week is about what surrounds the agent once you've pointed it: whether anyone actually designed the system it's standing in front of, or just handed it a licence and hoped.

Let's get into it.

🔥 Urgent Priorities

• No fires to fight this week

• Whether AI actually pays off is no longer a guess - the answer now has real data behind it

• Compliance, residency and sovereignty all moved this week - worth five minutes checking whether the objection you'd been holding onto still holds

No panic needed this week. What it calls for is those same five minutes, spent on what you'd actually design once that objection is out of the way.

🎯 Strategic Insight

Agentic AI got reliable enough this year to run real production work - and that's the hinge. The moment an agent can carry work that actually matters, the design of the system around it, and of the organisation holding it, matters more than it ever did while this was still pilots.

Tension: most leaders hold that the returns arrive with the rollout - that once the licences are out, the value follows.

Optimistic insight: the returns follow the design. An agent is only the interface. What pays is the system built behind it, on purpose, holding AI, code, data and people in the right proportions - and design is a job most organisations already know how to do. Get that part right and the adoption that follows - the piece leaders worry costs the most - turns out to be affordable and entirely achievable, because the groundwork that makes it stick is already done.

Simon Willison made the same point from the technical end this week: natural language can never be rewritten without losing something of the original. If a restatement always loses something, precision in how you say what you want carries the entire message - which is the same thing someone on LinkedIn put more bluntly this week: the real content of an AI rollout is the work redesign underneath it.

What's shifting: the reasons to wait just ran out. Compliance, residency and sovereignty were the last real objections, and all three moved this week - a capable agentic model now runs on hardware a business owns outright. The constraint has moved from what's permitted to what's actually been designed properly.

Why this matters now: 42% of companies abandoned most of their AI initiatives last year, up from 17% the year before. That's real, and it's recoverable, because most of them were missing a design. The reason it's recoverable is almost mechanical: the cost of being properly careful - of thinking through what a process needs before you build it - used to run higher than most people believed the benefit was worth. AI is collapsing that cost close to zero, for anyone who sets the process up properly first. Diligence stopped being expensive right when it started being decisive.

In this AI-driven world, the best use of AI is as little use of AI as possible.

Design the system first - what's AI, what's plain code, what's data, what stays with a person - and let the model take only the part that earns its place. That's frugality as a design discipline, and it's also exactly where compliance, residency and sovereignty actually live.

Action: pick one process you're tempted to hand straight to an agent, and draw the system before you build anything - what's AI, what's code, what's data, what stays with a person. Then brief only the part that goes to a model, in specifics: the situation, the action wanted, the role the model's playing, the goal, and what "good" looks like - SARGE, if you want the shorthand.

I miss this myself, more often than I'd like to admit. Once you're actually doing the work, it's easy to forget you were meant to be redesigning how the work gets done in the first place - the doing swallows the redesigning. The check I use on myself: am I talking to Claude, or am I typing a document? If I'm typing the document myself, I've slipped back into doing the human thinking I was meant to hand over. Do the one process properly this month, before you do the next ten badly.

🤓 Geek Out

1️⃣ Meta put a serious AI model on hardware you can actually own

Meta released Muse Glimmer, an AI agent model large and capable enough to rival what you'd normally only reach through someone else's cloud service - and small enough to run on the graphics card inside a high-spec gaming PC, not a data centre. (For scale: it's built from 30 billion "parameters", the adjustable settings that make a model good at its job - roughly the same range as the models most businesses currently rent access to, not a scaled-down toy.) It landed alongside two smaller but telling pieces of research from the same week: a way of shrinking large models down to a fraction of their size without an expensive retraining step, and a plain-English walkthrough of taking a model from around 140GB down to about 4GB while keeping it useful. The direction is the same in all three: capability you currently rent by the request may soon sit on a machine you own outright.

Why it matters: the sovereignty argument in this week's Strategic Insight is about to get very concrete. Capability that used to exist only behind someone else's servers can now run on hardware in your own building, which changes who sets the terms and where your data actually sits.

👉 Action: ask whoever manages your AI systems which of your current workloads could run locally within twelve months, and what that would cost against what you're paying to rent it today.

👉 Meta's announcement

2️⃣ AI reading your blood sugar, and a caution worth reading alongside it

Abbott and Google announced a partnership putting Google's AI on top of continuous glucose readings and other measurements taken directly from the body, aiming to make everyday health data actionable in real time. The same week, a study in the American Journal of Epidemiology found that machine-learning methods for estimating how a treatment affects a specific person aren't yet reliable enough to act on with confidence. Both things are true at once: the technology can help someone manage a condition day to day, and the algorithms behind a "personalised" recommendation aren't yet proven enough to bet on without checking.

Why it matters: whatever "personalised" means in a product you're evaluating - health, HR, finance, anything - this is the shape of the question worth asking before you rely on it: what's the real upside, and what hasn't been proven yet.

👉 Action: ask that same question of any AI-powered system you don't own but rely on - what evidence sits behind what it claims to do. Building that question into how you vet suppliers is what makes your procurement AI-ready, not only your product.

👉 The Abbott-Google partnership · the reliability study

3️⃣ A brake that watches your coding agent while it works

Apollo Research made Watcher self-serve this week - a program that runs alongside an AI coding assistant (a "coding agent": AI that writes and runs code with only light supervision) and watches everything it does in real time, only stepping in when it matters. In its own words, it blocks things like a leaked password or a system being wiped out the moment either starts to happen, and stays quiet while the agent gets on with reading files and running tests.

Why it matters: once an agent moves from suggesting edits for a person to approve to actually reading and changing the code and files that run your business, it needs a brake that works at the moment something goes wrong - and a team can now switch that on themselves, without waiting on a specialist vendor to build it for them.

👉 Action: ask whoever runs your coding agents what would stop a destructive action mid-flight, and what it would cost to find out.

👉 Watcher

🎨 Weekend Playground

Clone your own voice this weekend with ElevenLabs - free to start, and done in well under an hour.

Why this matters: voice cloning arrives with one story already attached: deepfakes, fraud, someone's voice used against them. That story is real - and there's another one just as real: the same capability can restore a voice to someone who's losing theirs, read a service into the language a user actually speaks, or let a small charity make audio it could never have afforded to record. Which story you get depends entirely on who does the designing - and this weekend, that's you.

👉 Mission:

• Go to ElevenLabs and record a couple of minutes of your own voice reading anything you like

• Let it build your voice clone, then type a sentence you've never said out loud and listen to yourself say it

• Think of one use for it that helps someone rather than deceives them - a message in a language you don't speak, a story recorded for a child who won't always have you there to read it, anything at all

• Notice that the choice of what to build was yours the whole time. That's this week's whole argument, in your own hands

A quick note from Hugo: AI Savvy Leaders is a free, government-funded course for leaders, built with Working Knowledge - and it goes deep on exactly the gap this edition keeps circling: getting AI enabled in an organisation is one job, getting people to actually adopt it is a harder one. If you want to find out whether your organisation qualifies, email [email protected] and Oscar will talk you through it.

(Disclosure: AI Night School is my own venture.)

📢 Share the Optimism

If The AI Optimist helps you think more clearly, forward it to someone else handling the shift.

And here's the question I'm curious about this week: if you designed the system behind one of your processes properly, how much AI would it actually need? Reply and tell me - I read every message and I'll come back to you personally.

Stay strategic, stay generous.

Hugo & Ben