Home/Learn/AI marketing

Get customers with AI doing the legwork

AI marketing can handle research, drafts and campaign production for a founder without a marketer. Give agents those jobs, with people choosing the experiments and reading what buyers will see. Someone also needs to adjust the instructions each week, using the results to decide what the agents should do next.

Updated November 2026

If you already use AI for code, asking it to help with marketing feels like a reasonable next step. You can get a draft or campaign idea while the day fills with product and sales work. The first output arrives before you have had to learn much about marketing.

The question is what happens after that first answer. Someone has to choose the buyer question the draft answers and where it will appear. A recurring campaign also needs someone to notice when its instructions stop producing useful work. The useful split starts with those responsibilities.

In AI marketing, agents draft and people pick the experiments

An agent can research keywords and draft a buyer guide before anyone has decided whether that guide is the right next job. A person needs to make that choice.

Research and production contain plenty of work to hand over. An agent can gather search terms, assemble information about potential buyers and prepare a page draft. Campaign work can include draft ad creative and research for a contact list. These jobs produce something a person can inspect and use.

Choosing among those jobs asks a different question: what should this company try next? A technically sound page may address a topic buyers rarely search. A well-researched list may contain companies with little reason to consider the product now. Producing either item faster leaves that choice unanswered.

We found that a model asked for a growth plan tended to reach for technical SEO. That is a plausible suggestion, but its familiarity does not make it the right starting point for every company. The choice needs more candidates than the first answer supplies.

Their planning method explores about a hundred ideas, scores them against agreed criteria and connects the selected experiments in a growth plan. A person steers that comparison. Generating many ideas gives the person choosing the work more options to compare.

Once the choice is made, the agent has a clearer job. Keyword research supports a particular buyer question. The guide draft answers that question. The campaign draft speaks to the people the experiment is meant to reach. Each output belongs to a decision someone has made about the company.

This also gives a founder a useful way to assess an AI marketing assistant. Look past the number of formats it can produce. Find out who selects the work those formats serve and who changes the selection when the results suggest another direction. An assistant that drafts well can carry substantial production without owning that decision.

Marketing agents left alone still turn out sloppy work

We spent about four months building an autonomous marketing agent. By August 2026, only a few of its workflows produced work we considered usable.

That experience matters when the promise is an agent that runs all marketing alone. Getting a task to execute is different from getting a campaign or article ready for buyers. A system can return an output, complete its steps and still leave a person with substantial editing to do.

We also reconsidered which jobs were worth rebuilding. For keyword data, we returned to an existing SEO tool. The agent could use established information without reproducing the whole job of collecting it. That made the division of work more practical: use a dependable source for the data, then direct the agent toward research and production around it.

The useful workflows were retained with defined inputs, a schedule and a wait for approval. That is a more specific unit of work than an instruction to grow a company. It makes clear what the agent receives, what it should return and when the next run should happen.

A fixed input also gives the person maintaining the agent something to inspect. If a draft is poor, they can trace it back to the assignment and source material. If the job is simply to run marketing, it is harder to tell whether the problem began with the topic, the evidence or the writing.

The useful parts can remain in production at the scope where they work. A keyword pull, a research brief and a page draft can each remain valuable even when the larger autonomous promise has failed our own trial.

For a founder evaluating AI marketing automation, ask to see the actual output of the recurring jobs. A list of completed runs tells you that the agent did something. Reading the research and draft tells you whether that work can support the campaign you intended to run.

People write the posts and approve what buyers see

A post published in a founder's name carries that founder's judgment. Agents can find topics and prepare outlines, while a person writes the words the audience will read.

This is especially relevant on social media, where buyers encounter the person alongside the company. An outline can identify a useful topic and collect the details behind it. The writer still has to decide which point the founder means to make and how that person would explain it.

The same responsibility extends beyond social posts. A page, email or ad may be written with substantial agent help. Before it reaches a buyer, a person reads the finished artifact. That read concerns the version someone will actually see, including its claims and invitation to act.

Reviewing a final page differs from approving the idea for it. The topic may be right while the draft says something the company cannot support. Equally, a sound paragraph can sit beside a confusing call to action. The final read keeps the buyer's whole encounter in view.

We treat human review as part of the work that allows agents to take on more production. Buyers and founders can know who stands behind the output. Responsibility stays with people even when much of the research and drafting comes from software.

For a founder, that need not mean becoming the editor of every sentence. People running the marketing can review the work, with the founder deciding points that need their own judgment. The useful distinction is between directing the company and operating every production step yourself.

When considering AI marketing services, make the handoff concrete. Identify who writes the founder's posts and who reads pages, emails and ads before they go out. That tells you more about the buyer's likely experience than a broad promise that a person remains involved.

Someone who understands the agents must adjust them every week

When an agent returns an unhelpful draft, someone has to understand why it happened and change the instructions. Repeating the same request next week leaves the same problem unresolved.

Running marketing through Codex or Claude can work, but we found that continued fine-tuning takes considerable effort. The initial setup is only the beginning. Someone has to read the results, relate them to the intended task and decide what the next run should do differently.

Our notes compare this to training a new employee. The agent knows a great deal, yet its default assumptions can pull it away from the instructions. A person who understands both the marketing task and the agent can see where that happened. Without that understanding, the company can lose the value of work it has already put into the system.

This is where the choice of help becomes practical. A platform can give a technical founder the means to run and tune the agents directly. The founder gains control over the instructions and takes responsibility for reading their output. That arrangement fits someone who wants to keep doing the work themselves.

For example, the person responsible for a buyer guide needs to understand the question it answers as well as the instructions that generated it. They can then tell whether a poor draft needs a different assignment or a change to the writing guidance. Familiarity with the agent alone does not settle that marketing judgment.

A freelancer can run AI tools within a chosen channel. That gives the founder a person responsible for that area, such as search or LinkedIn. The scope should say who adjusts the instructions and how their output is judged within the channel.

People running experiments across channels can take responsibility for the wider choice as well as the weekly changes. They select the work, direct the agents and read the results before adjusting the next assignments. The founder discusses direction while the people operating the agents keep the instructions connected to it.

Compare those arrangements on the same questions: who tunes the agents, who reads the output and which channels are covered? Those answers identify the hours and judgment still needed after setup. They also show whether the help ends with a working tool or includes the continued marketing work.

The person tuning the agent can compare its request with what it returned. A buyer-guide assignment, for example, should produce a draft about the chosen buyer question. If it returns a general description of the category, the next instruction needs to address that gap. This gives the person maintaining it a specific writing problem to fix.

A weekly review should lead to a concrete change when one is needed. That might be a narrower research request or a clearer writing instruction. These changes keep the agent useful as the company learns from its marketing.

Run the agents yourself if you already use a coding agent

Self-running fits a technical founder who already works in a coding agent and has time each week to maintain the marketing jobs. Familiarity with the software makes that option easier to assess.

An open-source marketing system can run on your machine with your keys and data. That gives a builder a way to inspect and change the work directly. It is a reasonable choice when operating the system is something the founder wants to keep doing.

The time question deserves the same attention as the technical one. Knowing how to connect an account does not mean there is room to connect it this week. A founder who is already postponing marketing may have little time for the setup, even before recurring work begins.

We heard that distinction from founders who wanted the growth work done. They did not want to set up an agent or build another piece of software. For those founders, comparing platforms alone leaves the main need unanswered: someone must run the work.

There is still a role for the founder when other people operate the agents. They help decide the direction and approve matters that need their knowledge of the company. That is a different commitment from maintaining instructions, preparing production requests and managing each run.

Choose self-running when the coding environment is familiar and the recurring responsibility fits your week. Choose help with the full sequence when your priority is to have marketing progress alongside product and sales. In either case, decide who owns the ongoing job before treating the first successful draft as a finished setup.

Hand research and first drafts to agents

  • Separate the recurring jobs

    List the research and production marketing needs, then mark each item that a buyer will eventually see or read.

  • Assign the first drafts

    Give agents keyword pulls, research and drafts tied to a chosen experiment, with a defined output for each job.

  • Keep people responsible

    Put experiment selection and the final reading of buyer-facing work with people who understand the company and its audience.

  • Bound the paid test

    Set the spending cap and stopping condition before a campaign runs, so continued spending depends on the agreed test.

  • Adjust the next run

    Read the week's results and change the agent's instructions where its output has missed the purpose of the assignment.

Have people run the agents on your marketing

Talk to Tin about which marketing jobs agents could take on, which experiments to try first and how people would review the work. We read every agent draft ourselves before it goes out.

Book a call