👋 Greetings {{first_name|earthlings}},

Emma here, fresh back from Cannes. Jules and I were there all week repping for Spark – we met with clients including Neverland, Everland (this has caused some confusion in the office!) and ACC. We ran AI Audit Sessions - which was a wonderful way to meet loads of new people. Jules was on a panel at thenetworkone's indie gathering, we were interviewed on 300 Consultancy's podcast, and we grabbed some exclusive interviews for the next episode of our own What's New in Creative AI podcast (coming soon). The rest of the time was spent squeezing in as many talks as possible to soak up what everyone else is thinking about AI. Hot spots for us were LBB, Monks and BIMA.

Every session this year was an AI session – it didn't matter what it said on the billing. And this year the AI word on the street has shifted to the people side of AI: change management, workflow redesign, new roles, reorganisation. It's finally dawned on everyone that subscribing to the tools isn't enough. As we've always said - the competitive advantage up for grabs with AI lives in the human, leadership and organisational part of the equation.

I find that what gets argued out on a Cannes stage tends to land on our clients' desks three to six months later. So treat this as your read on what's coming, and your headstart on being ready for it.

After three years of "which tool should I use?", the room had landed in the same place: the tools aren't the problem. We are. And I mean that as good news.

This was the year the industry woke up to a truth we've been harping on about for two years: humans are the bottleneck in harnessing the tech that's already available. Not the models. Not the licences. Us. The way we've learned to work, the org charts we've built, the habits we can't quite put down – none of it is optimised for a world with AI.

The clearest illustration I heard all week: a team of developers at Iterable saw their efficiency rise 20% by adding an AI tool into how they already worked. But when they redesigned the actual business process for AI, rather than plugging it into the old one, the productivity gain jumped to 6x. Same technology. The difference was entirely in whether they were brave enough to change the work itself.

The stat that's been quoted to death is that 95% of AI pilots fail. What doesn't get said often enough is why. It's rarely the technology falling over. It's that the people around it were never set up to think differently about the work. The change management piece in the way we work, about how we now approach tasks in a world with AI is wildly underestimated, and it's the whole game if you want real competitive advantage out of the moment we're in. 

I've been thinking about this line since I heard it on Tuesday. It came up in a conversation about agency economics featuring Justin Thomas-Copeland, CEO of the 4As, and Floriane Tripolino, CEO of WPP Open X.

The numbers Thomas-Copeland shared are sobering. 61% of agencies still see AI as a cost of doing business, not an investment. Only 9% say they're actually monetising it, up a sliver from 6% last year. That data is likely US-based, but it matches exactly what we're seeing in the UK (The Spark Report has the UK numbers if you want them). The needle is moving, but slowly, and in the wrong gear. Which means if you are doing this well, you're taking real advantage in the market while most of your competitors stall.

The trap is subtle, and Jules and I gave it a name on our podcast interview with 300 Consultancy (you can watch it here). If your only conversation with a client is about how AI makes your process more efficient, you've set the trap yourself. Efficiency, in a procurement meeting, translates to "so you're cheaper now." Tripolino put it more bluntly: don't pass token costs through as a line item on the invoice, because that's just quietly billing the client for your own inefficiency. Price on the outcome instead.

The same logic applies internally. Don't treat building your team's AI capability as a cost of doing business. Reframe it as an investment in skills that adds more value to your clients. It scales up to the whole organisation, too. Everyone wants to ask "what's your AI strategy?" when the right thing to ask is "what's your company's strategy, and how does AI enable it?" Bolt AI onto the old way of working and all you get is cost. Connect it to what the business is actually trying to do, and it becomes something else entirely.

Strong contender for the killer line of the week: expertise is becoming software.

Every month, something we used to bill for quietly turns into a button. Versioning, localisation, first-draft copy, mood boards, research synthesis. None of it is going to zero overnight, but the floor is rising, and the things that used to be the deliverable are becoming the easy part.

So where does the value go? Upstream. Into insight, judgement, accountability, taste, and a genuine human relationship with the client. You can't charge on outcomes while selling execution, so position around the bit AI can't do, and let the tools handle the bit it can.

Don't sell speed. Sell better thinking. The moment you try to hand the AI the thinking, you've given away the exact thing your client is paying for.

The word that came up a surprising amount this year was decision. Including by Dr Laura Weiss from WPP, she said: before you redesign how you work, map every decision in a process, then choose where you're happy to hand the decision to an agent and where a human has to keep it. Agents get the calls where the territory is known or the stakes are reversible.

It's a wonderfully practical test. We talked about optimising your processes for AI earlier, part of this is moving the decisions to the left of the workflow and the activity to the right. Decide up front what's known territory and let the agent run there. Keep the irreversible, high-stakes, taste-dependent calls firmly in human hands. It heads off both the panicky "automate everything" and the equally unhelpful "automate nothing."

Weiss also named a truth that all leaders should consider the impact of: AI doesn't take the hard work - it takes the easy work. The big tech and the press promised it would clear the boring stuff and free people up for the good stuff, but in practice the low-stakes tasks go first – the ones that used to be a breather between the hard bits. So the working day becomes all peaks and no troughs, people do ambiguous, high-judgement work from morning to night, and the risk of burnout is high.

So we have to start designing jobs, teams and ways of working with that in mind. Get it wrong and you don't get a team with the headspace to be creative and add more value, you get an exhausted one. Which is why leadership itself is shifting, from managing tasks to creating the conditions for people to do their best thinking.

When people search now, around 80% don't click through, because the AI summary is enough. And if the answer engine is the destination, optimising for clicks starts to fall apart.

Two things follow.

This connects to a phrase I scribbled down that I love: brand is now code (this my actually take the top spot as my favourite line of the week).

People are finding your brand in ways you can't see or track, mediated by systems choosing on their behalf (i.e. agents). Your brand now has a second kind of reader, one that has no feeling about your work but absolutely makes choices about it. Showing up well for that agent reader, without losing the humans, is one of the core marketing puzzles of the next year.

Clients are entering a phase where they want complete transparency about how much AI went into their work, and they need to report it themselves, both for carbon intensity and to tell their own stakeholders exactly which AI was used where. Brandtech described already running this through their Pencil Pro platform, splitting it into visible watermarking, the bit that tells a consumer AI was used in a piece of content, and invisible watermarking, the traceability of exactly what was used, where and when, so they can produce full reporting on demand.

And it's being legislated, so this isn't a nice-to-have you can defer. Under the EU AI Act, the Article 50 transparency obligations apply from 2 August 2026: AI-generated or manipulated content aimed at EU audiences will need to be marked as artificial in a machine-readable way, and deployers will need to disclose it. There's a limited grace period for systems already on the market before that date, but 2 August 2026 is the one for the diary. There's parallel movement in the US too, so expect the direction of travel to hold even where the detail differs.

The reactive version of this story – clients asking which tools you use, and a few starting to shortlist the ones you're allowed to touch on their account – is already arriving. But playing defence on that is the wrong posture. Instead treat transparency as something you lead with, not something to brace for. A clear point of view on how you use AI, what you disclose, and why it makes the work better is fast becoming a trust signal in its own right – the kind of thing that wins the pitch rather than survives the audit. AI has taken governance out of the boring compliance corner to become part of the value proposition.

Which brings it back to where we started: if the bottleneck is human, so is the solution, and it's within reach to all of us. When speaking at thenetworkone's event Jules laid out the three skills that matter most now:

Notice what's missing: anything technical. These are judgement skills, not tooling skills. You don't get them from a licence or a lunch-and-learn, you build them deliberately, across a whole team, over time. The good news, if you're a leader reading this, is that this is the most controllable part of the whole picture. You can start building it on Monday.

Pull the AI themes from Cannes together and a clear picture emerges of what leaders now need to be good at. If you want to get ahead, these are the skills to build in yourself and your team:

I came home optimistic. The AI panels moved on hugely from last year, featuring more people who really know what they're talking about with real experience of AI, and a much stronger emphasis on the people, culture and organisation-change side of all this. All the stuff we love at Spark. We've been banging this drum for a couple of years, so it was strangely satisfying to see the panel conversations reflect our messaging. 

The companies that pull ahead will be the ones who did the harder, human work. Mindset over money, as ever.

That's exactly what Spark AI is built to help with. Our AI Accelerator takes a team from scattered experiments to structured capability in 90 days, across client services, strategy, creative and leadership. Our AI for Leaders programme is part of that, where we help you connect AI to your company strategy rather than bolting it on and hoping.

Even though I think it's all about leadership - you still need to know the latest tools available to you, so here goes:

The biggest story this fortnight is how governments (well, the US so far) have started switching AI models off. Here's what happened.

Earlier this month, Anthropic launched its most powerful public model yet, Fable 5. Three days later, the US government told the company to pull it, citing worries that they could be misused for hacking. The order said the models couldn't be used by anyone who isn’t a US citizen, and since Anthropic has no way of telling what nationality you are, it had no practical choice but to switch both off for everyone, worldwide. If you were using Fable from the UK, you lost access overnight. Opus 4.8 carried on working fine, and Anthropic has been clear it thinks the decision is a misunderstanding, pointing out that the same hacking trick the government was worried about already works on other widely available tools.

Then this week it happened again, to OpenAI, just handled differently. The US government asked them to hold back their newest model, GPT-5.6, and only let a small group of approved companies and government departments use it at first. So not a full public shutdown like Fable 5, but the same move: sovereign decisions on who gets the most powerful tools, and when.

Two things to take from this. The first is a strategic one: the era of assuming the most capable models will always be available to everyone, everywhere, is over. Access is becoming something governments decide, which means the gap between who can use the frontier and who can't may widen sharply by geography. If your business is building anything mission-critical on a single frontier model, this is the moment to think about resilience – knowing your workflows can move between models, and not betting the agency on one provider's continued availability in your market.

The second is: you do not need the newest, fastest, most powerful model to be a genuine power user. The advantage has never lived in raw model capability – it lives in how well you've built your context, your skills and your workflows around whatever model you've got (everything I’ve said in this newsletter!). Almost everything we teach in our programmes works just as well on the model you're already paying for. So don't let a headline about a suspended frontier model convince you that you're somehow out of the race. You're not.

Microsoft’s partnership with Anthropic is really paying off, and Cowork It is now available right within Copilot. Watch next week's podcast for an interview with Tracy Pilon, Microsoft's head of AI for Advertising. Rather than just answering questions, it does longer pieces of work for you across your connected files, apps and data, just lik Claude Cowork or OpenAI Codex now. The thing to watch is how it's priced. Instead of one flat monthly fee, you pay based on how much you use it, so the more work it does, the more it costs. That makes budgets harder to predict, and it stays switched off until an admin turns it on.

Cowork-style tools are a genuine step change for productivity, so skilling your team up on how to use them well is a strong return on investment, and it offsets the cost of clumsy use quietly inflating your usage fees. A competent Cowork user is a cheaper one. As you'd expect, reaction has focused heavily on cost, with IT and finance flagging that pay-as-you-go spending is hard to plan around.

Adobe has rolled Firefly across Photoshop, Premiere, Illustrator, InDesign and Frame.io (announced 18 June). Instead of clicking through menus, you describe the outcome you want and the assistant orchestrates the multi-step work, while you keep the creative calls. In Premiere it can pull together a rough first cut from your clips, in Illustrator it can spin out 50 versions of a file from a spreadsheet, and the new "Elements" feature lets it remember and reuse characters, objects and brand assets across projects, which goes straight at the consistency and hand-off problems that usually trip up AI work. Adobe's own framing echoes the whole theme of this edition: the agent handles the grind so creatives keep the craft, taste and judgment that make the work distinctly theirs. Most of it is public beta, so still early. I was trying to get a look at this in Cannes, but Adobe were being strangely tough to get access to.(Adobe's announcement; a hands-on walkthrough is on Thurrott.)

The more interesting move is where Adobe is showing up. It's bringing these tools to ChatGPT, Claude, Copilot and Gemini, so you may see Firefly as a connector inside the assistant you already use rather than only inside Creative Cloud. Worth watching the challengers too: Claude working inside Affinity (now part of Canva) is an exciting free alternative, and a hint of where Adobe will have to keep pushing. If you're a designer or creative on your team, this is the one to get hands-on with this fortnight. Have you been using it? Let me know what you think.

Anthropic has overhauled Claude Design (find it bottom left in the Claude app), the part of Claude that turns a prompt into designs and prototypes. 

The big change is you can now import your design system, and Claude builds against your actual components, colours and type, checking its own output and correcting drift before you see it. There's also an admin "brand lock" so everything it produces stays on-guidelines. You can now edit directly on the canvas rather than only through chat, and export or push work into Adobe, Canva, PowerPoint, Figma-adjacent tools and more. It also fixes the token-burn problem early reviewers hit, by sharing one usage pool across chat, Cowork and Design. 

Interesting. Generative design tools are growing up, from "make me something pretty" to "make me something that's unmistakably ours." Worth a proper look. Our own Rob Crow, who helps deliver our creative programme, is busy exploring exactly what it can do.

Our monthly Spark Sessions series, where we invite agencies to walk us through exactly how they've built AI into the way they work, is back and coming to you live tomorrow afternoon (Tuesday 30th at 2:15pm).

In this edition, I'll be joined by Jason Bradwell, founder of B2B Better, a B2B podcast and thought leadership agency. Expect a live demo of B2B Hub, the internal operating system his team built to run the agency end to end. He'll share what it takes for a small, focused team to build its own AI tools, and why it was worth doing.

👉 Save your place

One last thing. I can’t let the moment slip by without celebrating that this is the 50th edition of Spark Intelligence! 🥳 Each one takes a day or two of thought and love to write, so that's getting on for 100 days of writing poured into your inbox over the life of this thing. I've loved writing every one (mostly!), and I really hope you're still finding them useful. Here's to the next 50.

See you then,

Emma

Co-founder, Spark AI

Spark AI helps you lead your team through the biggest shift since digital, with AI training, transformation and tools. We’ve worked with 70+ agencies, published the #1 bestselling book on AI for agencies, and teach at Oxford University.

👉 Find out more