· 5 min read · Jason Youngjin Koh
What AI agents do in marketing
Most businesses using AI in marketing use it to write captions. The leverage sits elsewhere, in the work that used to need a person watching a screen all day.
The highest return uses of AI in marketing right now are the ones that run continuously without a person: researching the market, tearing down competitors' winning ads, producing product imagery, watching ad accounts and updating the website. Caption writing is the smallest of them. The agents carry the volume. A person with judgement still makes the calls.
What is an AI agent, in marketing terms?
A model with a job, tools and a loop. It gets a goal, access to the places the work happens, such as an ad account, a website, a spreadsheet or a browser, and it keeps working until the job is done or it needs a decision. A chatbot answers when you ask. An agent checks the ad account at 6am because that is its job.
Where do agents pay back first?
| Job | What the agent does | What a person still decides |
|---|---|---|
| Market research | Reads reviews, search demand, trend data and competitor launches every day and files it into one shared knowledge base | Which signals are worth acting on |
| Competitor ad teardowns | Collects the longest running ads in your category and breaks each into hook, offer, format and length | Which patterns fit the brand |
| Product imagery | Produces studio and lifestyle images of the real product in new settings | Whether the image is accurate enough to sell from |
| Ad account monitoring | Watches spend, cost per result and creative fatigue on every campaign, and flags what to scale, cut or refresh | The budget moves themselves |
| Website changes | Builds and ships landing pages, product copy and offer changes in hours | Brand, claims and final sign off |
| Email and Zalo | Chooses products, writes and designs campaigns from sales data | What goes to whom, and when |
Why does the research agent matter most?
Every other decision reads from it. A brand that researches its market once a year for a planning deck makes eleven months of decisions on stale information. An agent that keeps reading reviews, search trends and competitor launches builds what we call the brain: one place holding what customers say, what they search for, which offers competitors are pushing and which of our own ads worked.
Creative briefs, product pages and budget calls all start from the brain. Six months in, it knows the category better than any single person on the team, because it has read everything and forgotten nothing.
How do you reverse engineer a competitor's winning ads?
Start from what they keep paying for. Meta's Ad Library shows every active ad a page runs and how long each has been live. An ad that has run for three months is usually working, because few brands keep funding a loser that long. TikTok's Creative Center publishes top performing ads by category and region.
An agent collects those ads across a whole category and pulls each one apart into testable parts: the first three seconds, the offer, the format, who is on screen, the length, the call to action. The output is a list of patterns with evidence behind them. We then write our own versions for the brand. Copying an ad frame by frame gets you a competitor's positioning and none of their customers' trust.
Can AI produce product photos you can sell from?
Yes, under one rule: the product in the image has to be the product in the box. Working from photos of the real item, our workflow produces a batch of around sixty studio and lifestyle images in about twenty minutes. That speed changes what you can test. One photoshoot a season becomes fresh imagery for every ad set, every week.
Where it fails is fine detail: fabric texture, exact colour, fit on a body, the finish on metal or glass. An image that makes the product look better than it is gets caught by returns and reviews, and by consumer law in both Australia and Vietnam. Every image is checked against the physical product before it runs.
Should AI manage your Meta ads?
It should watch them. A person checking a large account can review a few campaigns properly each day. An agent reviews all of them every hour: cost per result against target, frequency climbing, a creative that has stopped working, budget still sitting on a campaign that stopped converting on Tuesday. It raises the flag with the numbers attached, and a person makes the move.
Meta's own automation already handles bidding and placement well. Our agents sit above it, on the decisions Meta's system is not built to make for you: which products to push, which offer to test next, when a winning ad has run its course.
What about voice agents and live selling?
Both are working now. I have built voice agents that call thousands of people at once and handle each conversation the way it was designed: confirm interest, answer the common questions, book the next step with a person. For live commerce, AI presenters can host product streams on TikTok Shop and similar platforms, and the same monitoring loop applies. Track which presenter, script and product drove sales, and feed that back into the next stream.
What does this change for the team?
The same people do far more. A copywriter who produced ten variations a week reviews fifty. A media buyer who managed three accounts oversees ten, because the watching is automated and only the deciding is left. In the teams we run, agents take the monitoring and the first drafts, and people keep the decisions, which is the part clients were paying for anyway.
Where should a business start?
- Pick the one job that eats the most hours each week and is mostly watching or repeating. For most brands it is ad account monitoring or product imagery.
- Write down what a good result looks like for that job, in numbers, before any agent is built.
- Run the agent beside the person doing the job for two weeks and compare. Keep it where it matches or beats them.
- Build the research brain next. Every agent after it gets better once it can read what the brand already knows.
The agents only pay off inside a budget structure that tells them what to watch for. Scale what works, test the rest