The Croisette Test

In three weeks, fifteen thousand of us will land in the south of France carrying a bingo card. Buzzword bingo. Bullshit bingo, if we're honest. The free square in the middle is already filled. It says "Agentic AI." (You can download your own here!)

You'll mark off the rest before lunch on day one. Outcomes, not outputs. Shoppable everything. Human in the loop. Hyper-personalisation at scale. Every stage, every rooftop, every cabana with a screen and a queue for rosé. By the time you've found the Spotify wristband and queued for a debate you half wanted to see, you'll have a full house.

Last June I wrote that Cannes had a rosé-coloured delusion problem: a festival celebrating an industry whilst that industry quietly cut twice the festival's own headcount. In January I wrote about the 95% truth regarding the way agency announcements add the extra 5% that pushes them just past what's actually been built. This year those two threads meet on the Croisette. 2026 is the agentic Cannes. Overclaim of ability. Underappreciation of complexity. A festival struggling, in the warm evening light, with its own reality.

Here's where I have to stop myself, because the lazy take is sitting right there, and it's wrong.

The lazy take is that agentic AI is overhyped vendor nonsense. It isn't. Gartner's own numbers show the steepest adoption curve they have ever tracked for an emerging technology. Roughly 17% of enterprises have already deployed AI agents, with another 42% planning to within a year. Some of what you see at Cannes will be genuine. Even the people selling hardest will admit, if you press them, that only some of this year's agentic buys are the real thing. Which is itself the tell.

So this is not a piece about AI being fake. The agents are coming, they are coming fast, and a handful of those cabana demos are the future. The problem isn't the technology. The problem is that you cannot tell which handful.

By Gartner's count, of the thousands of companies now claiming agentic AI, roughly 130 deliver genuine autonomous capability. One hundred and thirty. Out of thousands.

When a single word describes both an autonomous system that breaks down a goal, chooses its own tools and iterates and a scripted chatbot with a fresh label, that word has stopped doing its job. It is no longer a description. It is a mood.

In January I drew the line as sharply as I know how, with James Bond. Give Bond the objective (stop the villain) and he disappears. He picks the casino over the yacht, adapts when the plan fails, tries the next thing, and owns the result. That is an agent. Ask a librarian what the files say and they hand you precisely what you asked for, then wait for the next question. That is an assistant.

A real agent decides for itself, pursues a goal rather than answering a prompt, strings the steps together without you, learns from what didn't work, and carries the outcome. Five qualities.

Most of what gets called "agentic" on the Croisette will have one of them, at best, and a confident voice.

And here is what that does to you, the buyer. When you can't read capability through the language, you fall back on the only proxies left: brand and price. The expensive one must be the real one. The famous one must be the capable one. This is how you end up paying agentic prices for rule-based plumbing.

Gartner predicts more than 40% of agentic AI projects will be cancelled by the end of 2027. Read their reason carefully. Not "the technology didn't work." The reasons are escalating cost, unclear value and weak governance of projects launched as top-down mandates and bought on a checklist of features rather than against a real problem. Their own phrasing is the line to carry onto the Croisette: the agent is only as good as the human who bought it.

Which makes the CMO walking into Cannes this June not the victim of agent-washing but its enabler. Buy the buzzword in a cabana and you own the cancellation risk back at the office. Every "agentic" tool you sign is another vendor in the stack, another integration that doesn't quite integrate, another dashboard nobody owns. We have spent two years watching the marketing leader's job swell into Chief Everything Officer. Agent-washing is how it swells further: more tools, more contracts, and less genuine autonomy than you started with.

Staging an agentic demo is trivial. Genuinely. You can make almost anything look autonomous on a stage for ninety seconds. Shipping an agent that survives a real media workflow, a compliance review and a procurement audit is a six-to-nine-month change-management slog and most of that slog is unglamorous plumbing. Who can the agent touch. What the guardrails are. Where a human steps in when it gets something wrong. Governance is not a feature you bolt on at the end. It is the thing the system either has in its bones or doesn't.

When Omnicom built months of anticipation for a platform reveal and then delivered a tidy demo of existing functionality, I called it the Omnothingburger and said the quiet part out loud: the actual plumbing is a lot harder than the PowerPoint. That wasn't a dig at Omnicom in particular. It is the law of the category. The gap between the demo and the workflow is the entire story of enterprise AI, and Cannes is the one week of the year engineered to hide that gap behind better lighting.

So here is the test. Carry it in your pocket. Before you nod along to anything in a beachfront cabana this June, ask the vendor three questions and watch what happens to their eye contact.

One. What does this agent decide on its own, versus execute on fixed rails?

An agent chooses. Automation follows instructions. If everything it does was pre-scripted by a human, you are looking at a very good macro with a marketing budget.

Two. Who owns it when it makes the wrong call at 2am and what is the escalation path?

Real autonomy creates real liability. If nobody can tell you who is accountable and how a human gets pulled in, the autonomy is theatre.

Three. Can you show me this running live in production and not a staged demo for a client who isn't in this room?

The demo is the easy 5%. Production is the 95% that decides whether you have bought a capability or a slide.

Three clean answers and you may have found one of the 130. A flinch on any of them and you've found a bingo square.

That's the whole test. Decision versus execution. Accountability at 2am. Production, not demo. You can run it in the time it takes a waiter to bring the rosé.

Cannes has always sold the industry a flattering picture of itself, and for one week a year there is no harm in that. The harm begins when the flattering picture becomes the basis for a purchase order. This year the festival is selling automation it cannot yet explain to an audience that cannot yet evaluate it and calling the whole thing creativity.

The technology underneath is real, and it is arriving. That is exactly why the discipline matters now rather than later. The winners coming out of this Cannes will be the ones who could tell the 130 from the thousands and who treated "agentic" as a question to be answered rather than a square to be ticked.

I'll be keeping my card. Let's compare notes in July, when the rosé has worn off and the invoices have arrived.

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry's table: senior client, global agency leader, technology founder. First published 2 June 2026 in the Brandflow newsletter on LinkedIn.