AI Just Got Job Titles. Marketing Still Has Prompts.

The latest AI news this week is not another model card.

Companies started writing job descriptions for software.

Farms are handing records to models. Security teams are giving cameras an owner. Salesforce shipped Job-Ready Agentforce agents with a long-horizon runtime. OpenAI opened the Agents API in public beta. That is not a chatbot era. That is a staffing era.

Labs spent the same window asking everyone to slow down. Amodei wanted pacing. Altman parked a 2026 IPO. Asian AI stocks wobbled. Fine. Your Friday delivery date did not move.

What actually shipped

I am not collecting headlines. I am watching work get assigned.

• Applied AI left the demo stage. Teams are giving models one job: watch the cameras, read the books, draft the quote.

• Salesforce put job-ready on agents and added a runtime that can stay on a task instead of dying after one prompt.

• OpenAI opened the Agents API so you can spin a cloud agent on GPT-6 Astra without building a toy orchestration layer first.

• DeepSeek V4.1 Flash keeps undercutting frontier prices. Intelligence is no longer the scarce resource. Taste and ownership are.

If your agency still measures AI by how many captions you generated, you are measuring the wrong output.

The marketing job AI should actually hold

Here is the part most recaps skip.

Marketing does not need a smarter text box. It needs three named seats.

LAYER 1 — Answer-engine presence.

Someone on the team owns whether the brand shows up when a buyer asks ChatGPT, Gemini, or Claude. That is generative engine optimization. It is a weekly job, not a slide.

LAYER 2 — Variation production.

Heroes, hooks, landing blocks, ad cuts. Volume with a brand kit and a human editor. Not forty captions that share the same sentence rhythm.

LAYER 3 — Agent-readable offers.

Pages a buying agent can quote without inventing your price, your proof, or your next step. If Time can sell ads to agents, your media plan cannot stay human-only.

We are running it that way at Dzine Prodigy. One owner per seat. Kill switch on anything that can spend or publish. The model is rented. The judgment is not.

Do this week

Pick one client.

Write a one-page job description for an AI seat: input, tools it can touch, what done looks like, who signs off.

Then ship one artifact from that seat by Friday.

The labs can pace the frontier.

Your clients still need work that lands.