Talk: "If your client has the same AI subscription, why do they need you?"

Agencias Argentinas and Interact Argentina invited me to close their Lab IA series, and I split the hour into three parts.

The first was the question in the title. If a client can buy the same AI subscription I can, what are they still paying an agency for? My answer is that everybody has the same models. What nobody can buy is the context, the systems and the judgment around them. So I showed ten things running at Tombras rather than what demos well: synthetic audiences that leave a step-by-step trail of what they saw and where they gave up, campaign briefs in minutes, review replies where a person approves every negative one before it posts, UX audits in minutes, internal tools one technologist built in days, and agents grounded in our own client data. Two cases make the point from opposite sides. For PODS, the AI was the idea: about 6,000 hyperlocal headlines across 299 New York neighborhoods. For Quili.ai, taking the AI away was the idea: people without safe water answered every prompt by hand.

The second part asked how the role of the engineer changed in 2026. After twenty-five years of writing code, I mostly direct it now. I plan before building, let the expensive model reason and a cheaper one execute, and have two other models review every change. Some nights a loop keeps committing while I sleep. It has let me build well past websites, from a Photoshop plugin to a native desktop app to an MCP server. The deck itself was built that way: one source that renders to video, Google Slides, Keynote and a design canvas, in English and Spanish.

The third part was the risks for an agency using AI: bias, sameness, hallucination, copyright, the energy and water behind every prompt, client data privacy, and the jobs people are afraid of losing. The judgment did not disappear. It moved, which is why I still say to test every model that comes out and be selective about what you put into production.