What changes for brands when customers’ AI runs on their own devices?
Flavia Barbat and Thomas Marzano in conversation. Legible and Lovable, episode 3. Published by Brandingmag, 2026-09-21. 71 minutes.
Episode notes by Thomas Marzano.
Your customer is becoming a superhuman. Their AI is moving onto their own devices, where it works for them alone. No brand can probe it, and I predict none will advertise in it. The levers brands use on cloud assistants do not reach it. Integrity does: what the brand publishes, what it codifies, and what it proves.
The AI that works only for its owner
The AI your customer relies on is moving off the cloud and onto their own devices: a computer, a phone, a box beside the home router that serves the whole household. It handles everyday work locally, calls on a cloud model only for frontier tasks, and answers to one party, its owner. It browses, compares, books and renews on their behalf, and it acts first to do right by that person and to protect them.
Our manifesto named edge AI as a necessity in November 2025, and when Flavia and I recorded this conversation, new hardware built for it had just been announced. A launch is not adoption, but the direction is set.
The cloud is not going away, and what brands have learned there still works. A brand can probe how each major assistant describes it, and in the free tiers it can buy sponsored ads next to the answers. Keep tracking there. But the edge adds a world none of that reaches: thousands of private models, many tuned by their owners, with no central service to probe and, by my prediction, no ads at all.
Picture two instructions. In the first, a customer asks their AI for the latest products from a sportswear brand they already prefer. Trust was given before the agent set out; it lives in the prompt. In the second, they tell it to track every subscription they hold and cut their spending each month. Nobody named you. At renewal the agent asks three questions on its owner’s behalf: is this the best product, at the best price, and, from everything it can infer, the most trustworthy choice? In the open mandate, trust is not handed over in the prompt. It has to be found in the evidence.
So the decisive work sits in the one place every AI reads from, cloud or edge: the brand itself. Publish owned content that represents the company, the product and the brand accurately. Codify more than the rational data: codify the value, so an AI can infer why someone would choose you. And curate the proof the market holds, which starts with delivering on the promise at every touchpoint, so that the evidence the customer’s AI finds is evidence you earned.
Sovereignty is moving to the customer. Govern what you can still govern: whether the brand keeps its word.
Questions this conversation answers
What is edge AI, in plain terms?
Edge AI is artificial intelligence that runs on your own equipment instead of in a cloud data centre. For a company, that means its own data centres; for a person, their computer, their phone, or a dedicated box at home. In the episode I describe where this is heading: a local AI next to the home router that serves every device in the household and stays reachable when you leave the house. It handles everyday tasks itself, because those don’t need a frontier model, and hands off to a cloud model only when a task demands it. The reasons it will spread are practical: it is cheaper, faster, more secure and more private. For Europeans with sensitive data, such as health, finances or family, there is also the question of where that data is processed at all.
If AI runs on the customer’s device, does their data stay at home?
Mostly, but not entirely. Flavia raised the common misreading: that on-device AI means nothing ever leaves the house. In principle a local model processes everything locally. But the moment you ask it to book a hotel and fill in the form, data goes out with that request. And when a task needs a frontier model, the AI sends the data that task requires. The difference is control. Only what is needed leaves, and the owner holds the decision rights: authorising health data for one specific query, say, and keeping it locked otherwise. I compared it to how companies already manage access, with circles of who can read, who can write and who can revoke. That logic can be designed for personal data, and people will find it familiar quickly.
Is AI visibility tracking still worth it when AI moves onto customers’ devices?
Yes, and it stays necessary. The major cloud assistants are not going away, so probing how each of them describes your brand remains part of the job, and part of how I assess a brand’s readiness. What tracking cannot do is follow the AI onto the customer’s own devices. There it becomes one of hundreds or thousands of local models, many tuned by their owners, and monitoring every one of them is a battle I would not advise anyone to fight. The two worlds are additive: some intelligence stays in the cloud, and more and more moves to the edge. So track where you can, and invest at the source for where you cannot. Flavia named the consequence on air: the work turns inside out, into the information and the conduct every model will read.
Listen from 17:38 · The Shortlist Effect · The Agentic Brand Readiness Assessment
What are the no-regret moves to prepare a brand for AI agents, whatever happens next?
Four, and most companies haven’t made them yet. First, act and communicate with integrity, so that everything first-party, information and action alike, is consistent with who you say you are. Second, make it legible and retrievable, so an agent never has to choose between what one page says and what another says. Third, don’t overshoot your promises: put proof behind every claim and make sure lived experience echoes it. If it doesn’t, go fix it. Fourth, give agents agency on your digital estate, so they can complete what their user asked for. None of this depends on which AI wins or where it runs. It is also why the Brand Constitution holds when AI moves to the edge: the work comes down to governing your first-party data so that autonomous agents keep the brand accurate.
Why does the AI “harness” matter more than which model you use?
Because the model is only the intelligence. What you actually work with is the harness, the application around the model that organises everything else. It has five parts: the model; memory and context; skills, the standard operating procedures you can customise and build yourself; tools, the connectors and plugins that let it act; and interface elements such as voice. In the episode I describe a voice interface acting as an orchestrator across every project I run: briefing sub-agents, working the browser and the applications on my computer while I talk to it, and picking up on my phone when I leave the house. Flavia called it a little chief of staff. It is why I judge these tools by their harness, and by their model only second.
What does it mean to delegate to AI rather than prompt it?
It means specifying the goal and the standard, then letting the system work until it meets them. The skill used to be prompting: learning exactly how to ask. What has changed is that these systems now pursue long, goal-oriented tasks on their own. My practice is to have the agent build an evaluation rubric first, then loop, rebriefing itself and spinning up further agents, until it averages 80 per cent across roughly ten dimensions. The decisive move is the benchmark. Set it with something human-made, a photograph, a layout, a page of copy, and the system keeps going until it gets there. It burns tokens. On my own website it produced, in under an hour, what would have taken me weeks. I call it the move from the age of prompting to the age of true delegation: a team delegating its own work, which is a different thing from a customer’s agent acting on their behalf.
Has AGI arrived?
For anything done on a computer, I would argue yes. Combine a strong harness, a natural voice interface, control of the whole machine and the ability to pursue a goal over a long stretch, and AI can now execute what a person can execute on a computer. Where I stop short is the physical world. General intelligence includes spatial intelligence: moving through a house, acting in the world. AI is not there yet, which is why I hesitate to use the word. The consequence for daily life is already visible. The need to sit at a screen clicking buttons and filling in forms is disappearing. What remains is a rendering device: in your earphones, on a screen, in the car.
Which work should a brand team automate, and which should stay human?
Automate what keeps you from the things that matter, and protect what you are valued for. For a person I use three filters: automate whatever keeps you from doing what you love, from the people you love, and from adding value to the customers you serve. Then check in. If you find you have automated the thing you love most, you took a wrong turn. For a brand the rule is the same at scale. The core of why customers value you is the piece you don’t automate, because that is where your humanity lives. A lifestyle brand has no business showing synthetic people; a home-utilities brand can generate its rooms. Machines already outperform us on raw intelligence. What they cannot do is feel, and the human instinct that says this colour is off still sets the bar.
Calls made on air
Published 21 September 2026; windows added in October 2026. Each call as I made it, with the window in which I expect it to land: the year it could first happen, and the year by which I think it will have happened.
- A local AI box beside the home router, serving the whole household and reachable remotely. (07:26) From 2027, by 2028.
- Edge-capable hardware becomes affordable for most people in Europe. (13:19) From 2027, by 2028.
- Local AI becomes the default for everyday tasks, with cloud models reserved for frontier capability. (09:24) From 2028, by 2030.
- You cannot run ads inside a local edge AI. (10:24) Holds wherever AI runs locally.
- The age of prompting gives way to the age of true delegation. (45:04) Under way in 2026.
- AGI, as I would argue it, has been reached for cognitive, computer-based tasks, and not yet for the physical world. (48:06) Computer-based tasks: 2026.
In their words
“Just by doing that, you will leave the evidence behind, and that evidence is what then ultimately is going to be picked up.”
“It all becomes like an inside out thing.”
“Make sure you’ve got proof behind the claims, and that lived experience is echoing what it is that you’re saying yourself. If it’s not, go fix it.”
“That’s the piece where your humanity lives.”
“The one thing AI cannot do, it cannot feel.”
Chapters
- 00:00 A new format: field check-ins
- 01:40 What edge AI means
- 05:17 New edge AI hardware, and signal versus adoption
- 12:00 Where your data really goes: the privacy myth and access circles
- 17:38 Will brands still need constitutions?
- 22:00 Agent swarms and weaponised brands
- 29:02 What responsible brands must do now
- 33:36 New frontier models and the “harness”
- 40:12 From prompting to delegation
- 52:07 What should you automate?
- 56:58 The automation trilogy and signing off
Sources
- NVIDIA, “Sparks Fly: NVIDIA Accelerates Local AI at IFA 2026”, 3 September 2026
- OpenAI, “ChatGPT Ads expands across Europe”, 18 August 2026
- Brandingmag, “Brand Constitutions: The Legible-Lovable Standard for Building Equity in an Agentic Economy”, November 2025
Where to start
A readiness assessment has to cover both worlds: how the major cloud assistants describe your brand today, and whether what you publish, codify and prove will hold up with the AI you cannot track. That is where the Agentic Brand Readiness Assessment begins.
Explore the engagements · Start a conversation
About the series
Legible and Lovable is Brandingmag’s podcast on agentic branding, hosted by Flavia Barbat, editor-in-chief of Brandingmag, and Thomas Marzano. In Brandingmag’s words: “Thomas identified the field’s dynamics and wrote the initial thesis; Flavia named the field and edited the content.” Episode 3 opens the series’ field check-ins, which test the manifesto’s horizons against what has happened since.
Where this sits. Agentic Branding is the discipline of making brands intelligible, governable, and desirable in an AI-mediated world. Its governing principle is the Legible-Lovable Law: brands must be legible to machines and lovable to people, with the machine now carrying both halves.
All episodes of Legible and Lovable. Episode notes by Thomas Marzano, updated 2026-10-06.