Why is AI ROI falling as adoption doubles?

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 was the week that was...

This was the week the tools got sharper and the advice got more careful - inside the same seven days.

Two frontier labs converged on the same move this week: agents that pick up a task and carry it through to done. On 7 July, Anthropic's Claude Cowork - the mode where you hand Claude a task and it works across your files, calendar, email and the web until it's done - expanded from desktop-only to mobile and web. Two days later, on 9 July, OpenAI shipped GPT-5.6 and relaunched ChatGPT Work the same day, describing it as a tool that gathers context, plans the approach and takes action across your tools, files and desktop apps. Two labs, two "do the actual work" products, 48 hours apart. For scale on how real that appetite already is: OpenAI's Codex has gone from a niche coding tool to more than 5 million weekly users, a sixfold rise since February.

Then came the more interesting story. On 14 July, OpenAI published "How to manage AI investments in the agentic era" - a frontier lab telling its own customers to get a proper grip on their AI spend. That's OpenAI itself, in the same month it shipped its most capable and most agentic model yet, saying that capability and cost discipline need to move together.

That advice lands on ground already well prepared for it. Research published in March by the British Chambers of Commerce and Atos found UK SME AI adoption at 54%, up from 35% the year before - genuinely mainstream now. And TEKsystems' State of Digital Transformation 2026 report found enterprise-wide implementation had doubled year on year, from 12% to 24% - while confidence that AI spend will pay off within six months fell from 42% to 27% over the same period, and nearly a third of organisations reported costs quietly overrunning budget. I see the same pattern everywhere, across clients of every size and industry: the tool goes in after a lot of hard work deciding on platforms and then getting people to use it is ever harder. On the ground right now, two clients: both enabled AI tooling in January, both switched it on properly, and both are asking me some version of "how do we get people to use it?" and “Where’s the return”.

Enablement solved one problem. This week, both the tools and the labs that build them told us which one it opened. AI adoption.

Let's get into it.

🔥 Urgent Priorities

✅ No fires to fight this week
✅ The adoption number everyone's quoting measures enablement - tools bought and switched on. Whether people actually changed how they work is a separate, much slower question
✅ Worth asking now: is your AI rollout horizontal (everyone, find your own use) or vertical (one process, one owner, one metric) - it decides whether you'll ever be able to see the payback

No panic needed this week. What it does call for is a proper look at whether what you enabled this year actually changed how anyone works, or just how fast they did the same thing.

🎯 Strategic Insight

Tension: 54% of UK SMEs now use AI, up from 35% a year ago, and enterprise-wide adoption has doubled too - both frontier labs even shipped agentic tools within 48 hours of each other this week. On paper, the adoption question is solved. In the same data, ROI confidence is falling and spend is quietly overrunning budget. A 54% figure that comfortable should have settled the question. It hasn't - the number counts something narrower than everyone assumes.

Optimistic insight: Here's the split I keep seeing, across clients of every size and industry: what gets reported as AI adoption is almost always AI enablement - the tool bought, the licence issued, the assistant switched on. Adoption is a different, much harder job: people actually changing how they work. Enablement is a purchasing decision, done in an afternoon. Adoption is a people-change project, and people-change is always slower than procurement. That gap is exactly why ROI confidence is falling in the same data that shows adoption "doubling": enablement is easy to buy and easy to count, adoption is neither, and most of what's being measured right now is the easy half.

There's a mechanical reason this gap matters: as AI pushes the cost of doing toward zero, doing stops being what makes you different, because your competitor's doing got cheaper at exactly the same rate as yours. Enablement buys everyone that same falling price. Adoption is the only lever left that's still yours to pull.

What's shifting: How the rollout happens explains why adoption is harder to see than enablement. In what I see across clients, roll AI out horizontally - Copilot or Claude switched on for everyone, find your own use - and the benefit is real but gets absorbed into the business: hard to trace at the task level, more a 3-5% movement on function KPIs than a headline win, which is exactly what "we can't see the payback" looks like from the CFO's chair. Roll it out vertically - one named use case at a time, run as a product with an owner, a metric and a cost line - and the benefit is designed in at the start, so it can be measured at the end. Most of what's been enabled this year went in horizontally, because horizontal is the easy purchase. Vertical is the slower, people-change work - which is exactly why it's the one still undone.

Takeaway: You already enabled the tool - that part's done. What's still undone is the room adoption needs. This week, give it to someone else: pick one team, hand them an afternoon with no task list, the AI you've already switched on, and one question - what would you build with this if speeding up the same report wasn't the brief? That's the adoption job: the people-change work the numbers don't measure, and it only moves when you give people room to explore what you've already enabled and imagine something more ambitious with it.

🤓 Geek Out

1️⃣ A £40 computer just proved you don't need the biggest model

Cambridge's Frugal AI Hub, working with the nonprofit Saving Voices Project, built a working speech-to-text system for the Soliga people of southern India from just five hours of recorded voice data. It runs offline on Raspberry Pi hardware costing about £40, needs no cloud connection, and is helping preserve a language that was never going to get frontier-model attention or a training-data budget from anyone in Silicon Valley. The wider project has similar ambitions for close to 500 million indigenous people across 90 countries, all built on the same principle: small, task-specific, and cheap enough to actually deploy. Rest of World has the full story.

Why it matters: This is proof that "good enough, cheap, and built for one specific job" beats "biggest, priciest, does everything" for most of what businesses actually need AI to do. If you're paying for headline capability you never touch, this is the nudge to check.

Action: Before your next AI tool renewal, list the three tasks you genuinely use it for, and check whether a smaller, cheaper, task-specific tool would do those three jobs just as well.

2️⃣ The whale-spotting AI that changed a shipping lane

Happywhale, a computer-vision platform that identifies individual whales and dolphins from photos submitted by anyone with a camera, won the AI for Planet category at this year's AI for Good Impact Awards, announced during the AI for Good Global Summit in Geneva earlier this month. It's cut identification time from over an hour per photo to seconds, catalogued 1.5 million photos with help from 53,000 citizen scientists in over 100 countries, and its data has directly influenced a new 40,000 square kilometre slow-navigation zone near Antarctica - a policy outcome that started with holiday snapshots.

Why it matters: This is what AI is actually good at - doing the tedious pattern-matching so humans can spend their time on the judgement calls and the decisions that follow. Most businesses have a version of this exact task sitting in their operations right now.

Action: Find the one task in your business that's essentially pattern-matching with photos or documents - invoice checking, enquiry sorting, compliance review - and ask whether just that piece, not the whole job, could be handed to AI this month.

3️⃣ The password problem nobody had solved until this week

Hand an AI agent your password so it can log into a site and finish the job, or stop and do the task yourself - until this week, that was the whole choice, and it's why most people won't let an agent near anything with a login screen. On 16 July, 1Password shipped 1Password for Claude, which lets Claude log in and complete tasks that need passwords or two-factor codes without the credential ever entering Claude's context or memory. When Claude reaches a login page, 1Password shows you exactly which credential is being asked for and why, you approve with Touch ID, and 1Password injects it straight into the page - Claude never sees the password or the code, and access is scoped to that one task and expires the moment it ends. A companion "Agentic Mode" locks the extension down the instant any AI agent takes the wheel of a browser, hiding its interface and limiting it to only the logins you approved - and it isn't Claude-only. 1Password calls this "zero-exposure architecture"; in plain English, the password never actually passes through the AI. As 1Password's CTO Nancy Wang put it: "The answer isn't handing agents your secrets. It is to let a user give an agent permission to use a credential without letting the agent see it."

Why it matters: Enabling an agent takes an afternoon; trusting it with something that matters is the slower part, and security has been the blocker. This closes that gap: the scaffolding for trusting an agent with real access arrived the same week both labs shipped agentic tools that need exactly this access to be useful.

Action: It's available now, not a beta - but Mac only at launch, and you'll need the 1Password desktop app and browser extension alongside the Claude desktop app and Claude-in-Chrome extension, all running on your existing 1Password plan. If you've got that stack, try it on one low-stakes login this week before you trust it with anything that matters. On Windows, this is one to watch for, not one to chase yet.

🎨 Weekend Playground

This week's Takeaway asked what your people would build with the room you gave them. Here's an hour to try the same question on yourself first, away from a screen for most of it - Hugo calls this a Wonder Ponder.

Why this matters: it's the smallest possible version of the adoption job - taking something you've already enabled and actually giving yourself the room to imagine what you'd build with it. (Quick disclosure: I'm a cofounder of Sherpas AI, and this exact exercise - walk, no screen, envision something bigger - is what we ask teenagers to do on our own programmes. If you want the fuller version, it's at www.sherpas-ai.com.)

👉 Mission:

• Go for a 20-minute walk with no phone in hand and no task list - just the question of what you'd build with an extra afternoon a week, if you let yourself be properly ambitious about it

• On Android: when you're back, record a two-minute voice note in the free Google Recorder app - it transcribes on-device as you talk

• On iPhone: record it in the built-in Voice Memos app, then use the dictation button in whatever free AI chat tool you already use to turn it into text. (I use Wisprflow)

• Read it back once tomorrow morning. Keep whatever's still true, bin the rest.

📢 Share the Optimism

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

And here's the question I'm genuinely curious about this week: which process in your business have you actually redesigned with AI this year, versus just sped up? That question has a cost dimension too: the bill balloons because you're paying frontier prices to run the old job faster. I've just finished a briefing on exactly this, covering both ways I see AI roll out - horizontal and vertical - and the people side nobody budgets for, called "AI Adoption Without the Ballooning Bill." Reply and tell me what you're working on. I'll send you the briefing alongside a proper answer - I read every message and come back to everyone personally.

Stay strategic, stay generous.

Hugo & Ben