Greg Isenberg Says AI Agents Will Be a Commodity. Knowing Where to Point Them Is the New Job.

Marketing has had three eras with a named job attached to each, and Greg Isenberg’s argument is that a fourth just arrived. If he’s right, what reaches a small marketing team isn’t a job title at all. The work moves from producing marketing to building the system that produces it, and the judgment about what that system should make becomes the part you’re paid for.

He published a 35-minute episode on August 31 laying it out. We pulled the captions rather than a summary, so what follows is verbatim with timestamps.

His closing line is the thesis:

“The agents are going to be a commodity at some point. Your judgment about what to point them to is the moat.”

The fourth era

Isenberg has started and sold three venture-backed companies across the web, social and mobile eras, and his frame is that “every time the technology changes, the most valuable kind of marketer changes with it.” He walks the sequence at 00:05:00, and made a shorter version of the same case in an August 18 post:

“traditional marketing was about making people care. Digital marketing was acquiring customers through measurable new channels. Growth hacking was about using product and data to build these loops. And marketing engineering is about using AI, agents, data, code, and taste to build a marketing system that keeps learning.”

The last clause carries the weight. A system that keeps learning without someone rebuilding it each quarter is new, and he’s explicit that it’s “now actually possible in the agentic era.” Agentic here means software that runs a task end to end on its own rather than waiting for instructions at each step.

What he doesn’t say is that the old skills go away, which is where most coverage of AI job titles goes wrong:

“The marketing engineer still needs all that old stuff. It still needs, you know, customer understanding, judgment, positioning, understanding distribution, taste.”

And then the sentence a small team should sit with: “taste matters more now than ever because AI is about to make average marketing just unbelievably cheap.”

That’s the actual economic claim. When competent output stops being scarce, the scarce thing is knowing which output is worth making.

The six systems

The most useful stretch of the episode is the build list, and he grounds all of it in one worked example: a vertical software company selling to commercial HVAC contractors. He picks it deliberately because “the buyer has a lot of money. The workflows are messy and the language is specific.”

1. The customer truth system. Agents read sales calls, support tickets, churn notes, CRM notes and payment movement, and write one file showing what changed. The rule attached to it is the best idea in the episode: every insight needs a quote, a link, or a source. He’s blunt that a vague summary is the failure mode, and the goal is a document that makes the business “harder to lie to.”

2. The founder content engine. Record the founder talking to customers, pull from podcasts, extract the strongest ideas, and let the system watch which hooks people actually keep watching. For the HVAC company, one insight about lost replacement revenue becomes a founder post, a short video on why contractors lose money after the first visit, a landing page line, a cold email angle, and a calculator that estimates the lost revenue. One idea, five surfaces, chosen by what performed.

3. The outbound signal engine. His framing here is sharp: “bad outbound usually starts with a spreadsheet, a spreadsheet full of names. But good outbound starts with timing.” Who just raised, who’s hiring for the exact problem you solve, who posted about the pain publicly, and who fits your profile with a reason to care this week. For the HVAC company that means watching for contractors hiring dispatchers, opening locations, or collecting bad reviews. The agent researches and drafts. A human approves before anything sends.

4. The creative testing engine. One offer becomes 20 hooks and 10 ad angles, tested, with results recorded. His diagnosis of “Facebook ads don’t work for me” is that most teams never tested enough angles to know. The reframe is that creative stops being a treadmill and becomes a learning system.

5. AI search visibility. Whether your company is legible to the systems people now ask instead of searching. He cites Sam Altman’s billion-user figure for ChatGPT alone, before counting AI Overviews, Gemini, Perplexity or Claude. We’ve covered this one from the measurement side repeatedly, including when answer engines started quoting people rather than pages.

6. The growth cockpit. A weekly view of what changed and what to do about it. What content worked, which campaign created real conversations, which objection came up again, what percentage of tests won, what competitors moved, what pain is getting louder. His HVAC version reads: the lost replacement revenue angle drove fewer clicks than the dispatch angle, but twice as many demo requests from owners with more than 20 technicians.

Fewer clicks, better buyers, and a system that noticed. That is the argument in miniature.

How he says to start

The starting point is deliberately unimpressive. Build a folder with five files (customer truth, founder voice, experiments, agent jobs), paste in 20 real customer notes, and give the agent one job: tell me what’s changed, show me the receipts, suggest one test that could create pipeline this week. Then build one thing from the output.

His 30-day version: week one audits one real company and produces a market map of who the customer is, what words they use, and where the funnel leaks. Week two is the repo and the first file. Week three is one system, and he’s firm about scope, because “one working system is going to beat five half-built ones.” Week four is results.

On tools he’s dismissive, which is worth repeating to anyone shopping for a stack: Claude, ChatGPT, Grok Bot, Gemini or local models all work, and “the tools actually matter less than the workflow here.”

The shape of the person

His description of who does this well is a list of halves:

“part marketer, part product person, part revops, part data analyst, part creator, and part engineer”

They can talk to a customer, build the workflow that uses the insight, write the positioning, wire the automation, make the landing page, read the conversion, and know when personalization sounds fake.

For a 5-to-30 person company, that describes a generalist who has stopped doing the work by hand: less a new hire than what your best existing marketer becomes if you give them the room.

What the market is paying, briefly

The title has a coiner. Profound, which sells AI-search monitoring, introduced it and runs marketingengineer.jobs alongside a course and a certification. We read all 12 listings on September 1: ten disclose a band, running from $85,000 at NBCUniversal to $324,000 at Expedia Group, with most clustering between $130,000 and $200,000. Isenberg sizes it at $250,000 to a million and calls that conservative, which fits if you read him as describing the top of a market and consulting work rather than posted salaries. Budget against the board.

The more useful signal in those listings is that Figma, Stripe, Plaid, Expedia and NBCUniversal are hiring some version of this under five different titles at once. The job is real and the market hasn’t agreed on a name.

The pushback that lands

Hanna Huffman, who runs MultiplAI Growth Partners and writes Marketer in the Loop, took the criticism head on in April: “this is just marketing ops with a new label” and “it’s tech bros trying to make marketing sound more masculine.” Her verdict is that the backlash is “one-third right and two-thirds wrong.”

Her defense is about the training rather than the tooling:

“They are not teaching you how to use ChatGPT to write better ad copy. They are teaching you systems thinking.”

She also splits the role in two, one embedded in a pod with other marketers and one operating as a company’s entire marketing system. Those aren’t the same job, and hiring for one while expecting the other is how a title gets a bad reputation.

Since we’ve cited Profound’s data in six previous pieces, the conflict is worth naming: the company that coined the title also sells the course, the certification and the job board. That doesn’t make the role fake. It does mean the enthusiasm and the product share an author.

What to take from it

The systems are free to try and none of them requires the title. Start with the customer truth file, because it costs an afternoon, and the rule about every insight carrying a source is what keeps it from decaying into a summary nobody trusts.

Then watch the language. “One person can do a whole marketing team’s work” is a sentence that will get repeated in headcount conversations by people who never watched the 35 minutes where six systems get built. We’ve written about the experience debt building as AI takes junior marketing work and about job listings dropping “specialist” and adding “head of”. This title will be used the same way.

The claim underneath is the one worth arguing with, and it’s the one Isenberg closes on. If agents become a commodity, the marketer’s value is no longer execution or even orchestration. It’s knowing what deserves to exist.