Skip to main content
A music producer seen from behind at a large mixing console, while the instruments in the live room beyond play with no musicians.

How should you structure a marketing team when AI does the work?

Eric Helsel
Eric Helsel·Senior Marketing Executive
12 min read
Share

TL;DR

A $2M to $20M company now needs two or three strong marketers running a library of AI skills, not five people producing assets. The headcount is the easy part. Whether it works depends on getting your best people's judgment out of their heads, and one step most owners skip is what turns that judgment into something AI uses every day.

The right marketing team structure for a $2M to $20M company is now two or three great marketers running a library of AI skills, not five people producing assets. Cutting heads is not the hard part. Getting what your best people know out of their heads and into the system is, and that is the step almost every owner skips.

I have run marketing teams for 25 years, and I have never seen the job change this fast. The tools are not the story. The story is that the work your team spends most of its week on, making things, is turning into something a machine does while your people decide what gets made. That changes who you hire, what they do all day and how you know it is working.

What does a marketing team look like when AI does the production work?

It looks like a recording studio where one producer runs the whole band from the mixing desk. AI agents play the instruments: they write the variations, build the pages, cut the edits and pull the reports. Your marketers sit at the desk and decide what gets recorded, how it should sound and which takes are good enough to release.

An AI-first marketing team is a small group of senior marketers who decide what gets made and judge whether it is good, while AI agents running saved instructions do the making. The old structure was organized around production. You had a writer, a designer, someone on social, someone on paid and someone stitching the reports together, and each of them spent most of the week producing assets for one channel.

Here is how I think about every task on that desk. Each one sits at one of four fader positions. Live take means a person does it start to finish. Producer in the room means AI drafts and a person decides. Automation lane means AI runs it and a person spot-checks. Loop means AI runs it on a schedule and reports back. Most of a modern marketing department now lives in the last two.

The same $10M company, two ways to staff marketing

Five people making things

  • A content writer drafting posts and emails
  • A designer building ads and landing pages by hand
  • A social manager posting and replying
  • A paid media specialist launching and pausing ads
  • A coordinator assembling the weekly report

5 people

each owning a channel

Eric Helsel

Can one marketer with AI really do the work of a whole department?

Yes for most of the production work, and no for the judgment, and that difference is the whole game. Most owners hear "AI can replace your marketing team" and land in one of two places. Either they cut everyone and get a flood of generic content, or they decide it is hype and change nothing. Both miss what is actually happening.

The production half of the job is collapsing. The deciding half is not, and it becomes the bottleneck. When AI can produce two hundred ad variations in an afternoon, the scarce skill is the person who knows which twelve to run and why. That is why the people you keep matter more than ever, and why I would not fire a strong marketer to save money on AI.

A conveyor belt pours out a mountain of identical trophies while one worker lifts a single one to inspect it.

If I have to choose between a brilliant marketer who is new to AI and an AI expert with average marketing instincts, I take the marketer every time. I can teach someone to run skills and agents in a few months. I have never been able to teach taste in a few months. A great marketer with AI behind them can do things that were simply not possible for one person five years ago.

I can teach a great marketer to run AI in a few months. I have never been able to teach a fast AI user taste in a few months.

- Eric Helsel

So the real opportunity for a $10M company is not cutting the team. It is growing to $20M without doubling it. Picture a company with five marketers who each spend about two thirds of the week producing. Hand most of that production to AI and the same five people get roughly a hundred hours a week back for the work that actually moves revenue. Run your own numbers below.

Interactive tool

Where your marketing team's week goes

Put in your own team. The example is a $10M company with five marketers, and the hours that come back are hours for judgment, not hours to cut.

The team's week

225

team hours a week

Spent producing

225

team hours a week

146

hours producing

Handed to AI

146

hours producing

102

hours handed to AI

Cost of an hour

50

$42

cost of one team hour

Time back a year

102

hours handed to AI

50

$42

cost of one team hour

$216,125

of time back a year

Team hours a week on judgment work

Today
79
With AI on production
181

hours a week your team gets back for judgment work

102

Example: a $10M company with five marketers and a 50-week year. Producing means making assets, building reports and moving files. Judgment means deciding what to make and whether it's good.

Eric Helsel

How do you bring your marketing team along without scaring them?

Tell them the truth early, show each person where they fit, and give one strong marketer the job of going first. Picture the Monday meeting where the CEO announces that AI will handle content from now on. Everyone in the room hears the same thing, which is that their job is gone, and they spend the next month polishing their resumes instead of learning the tools.

The conversation that works sounds different. You tell the team that this shift is coming whether anyone likes it or not, that you would rather lead it than watch competitors do it first, and that you are going to invest in making each of them the person who runs the system. Then you mean it. In my experience, the people who hear that offer and take it seriously become far more valuable, and they know it.

Be honest about the condition attached to it. The new role rewards people who are genuinely good at marketing, because the job becomes deciding and judging. Someone who was mostly moving files will find it hard, and you owe them that clarity now rather than a surprise later. You can also point out the obvious: someone who can run a marketing system with AI behind it is ahead of almost everyone else applying for their next job.

Do not try to turn the whole team into AI experts at once. That eats months and distracts everyone from the work that pays the bills. An AI champion is the one strong marketer you free up to build the system first, absorb the frustration of early mistakes and hand working skills to everyone else, while the rest of the team keeps shipping.

💡 Tip

Pick your champion by marketing judgment first and curiosity second. Technical skill is the easiest of the three to teach.

Which marketing work should you hand to AI, and which should stay human?

Hand AI everything that follows a repeatable process, keep a person on anything that decides what your company says, and be careful with the work in between. Positioning, choosing which ideas get made and the final call on what ships stay as live takes. Those are the decisions your customers feel, and they are the hardest thing to write down for a machine.

An airliner flies itself above the clouds at dawn while the captain keeps a hand on the throttle.

Research belongs with the producer in the room. AI is excellent at reading everything, ranking what it finds and handing you ten ideas. It is not good at replacing your critical thinking, so a person picks the three worth making. Writing a script or a campaign concept from a blank page also stays mostly human for now. AI is reliable at rewriting a proven script into new versions, and much weaker at inventing a strong one from nothing.

Almost everything downstream of that decision can move to the automation lane or the loop. Static ads built from proven layouts, first drafts of landing pages, outreach to creators within budget rules you set, simple video edits, weekly competitor monitoring and performance reports all run well with a person checking the output rather than making it. Tap through the board below and switch the business type to see how the ratings move.

Every marketing task, by fader position

Switch the business type, then tap any task for the reason behind its rating.

Live take

A person does it start to finish

Producer in the room

AI drafts, a person decides

Automation lane

AI runs it, a person spot-checks

Loop

AI runs on a schedule and reports

Tap any channel for the one-line reason behind its rating.
Every rating, as a list
  • Positioning and the core message: Live take. This is what customers feel. Write it yourself, then let AI test how it lands.
  • Choosing which ideas get made: Live take. AI can rank ideas. Picking the few worth money is the job you are paying people for.
  • Final call on what ships: Live take. One person with taste and accountability signs off on anything that goes live.
  • Customer and market research: Producer in the room. AI reads everything and ranks it. A person decides what it means.
  • Writing a script from a blank page: Producer in the room. Still weak from scratch. Strong at rewriting a proven script into new versions.
  • Rewriting proven ads into new versions: Automation lane. Give it the winner and the customer language. Review the batch, keep the best.
  • Static ads from proven layouts: Automation lane. Runs well from a library of layouts that already work. B2B brands often want a closer look.
  • Landing page first drafts: Automation lane. AI builds the page from your research. A person approves the headline and the copy.
  • Finding and negotiating with creators: Automation lane. Give it budgets and preferences and it handles outreach and replies. Less common in B2B.
  • Simple video edits: Automation lane. Cutting footage, captions and pacing now run in code. High-production edits stay with editors.
  • Checking delivered work against the brief: Loop. AI compares what came in with what was asked for and flags the gaps.
  • Competitor and ad library monitoring: Loop. A weekly digest of what competitors launched and what is spreading in your niche.
  • Performance reporting: Loop. Pulled, summarized and sent every week, with the questions worth asking.
  • Sales call and review summaries: Loop. Turns every call or review into customer language your skills can use.
Eric Helsel

Wherever AI hands work back, build in a stop. Human in the loop means a person approves the work at set points before it goes out. I like a stop at the headline and copy before anything gets produced, because it is cheap to fix a sentence and expensive to fix two hundred finished ads. You can build skills that run straight through without stopping, but in my experience the control is worth the extra minute.

Can AI edit your video ads now?

Yes, for most of the ads a $2M to $20M company runs. If your ads are footage cut together with captions and a social-style edit on top, an AI agent working in code can now edit them from the brief and the raw files, with no editing timeline open. I build product videos this way now, and the shift surprised me more than anything else this year.

That changes your editor's job rather than ending it. Several edits can run at the same time while the editor does something else, so the editor's real work becomes building and training the editing skills, and handling the complex, high-production pieces that still need a human hand. One quick win: pull every comment your team has ever left on past edits and turn it into a list of what your editors look for. That list is a ready-made starting point for an editing skill.

How do you scale marketing output with AI without producing AI slop?

Start from what your customers actually say, not from a prompt that asks for more. AI slop is content made in bulk with no real idea behind it, and it is what you get when someone types "write me fifty ads" and hits enter. It looks busy, it costs almost nothing, and it rarely earns a single dollar of spend once it is live.

A fishing trawler hauls up a huge net full of plastic bottles instead of fish.

Most marketers believe the format is what is working. A new style of video takes off, everyone copies it, and the copies flop. The reality is that the winners were carried by a strong message that would have worked in almost any format. When you scale with AI, the same rule decides everything: more volume only helps if every piece starts from a message your customers already gave you.

Picture a $6M company that sells ergonomic office chairs. Its post-purchase surveys show that a surprising share of buyers are nurses who chart for hours after twelve-hour shifts. Nobody on the team ever wrote an ad for them. That single insight, turned into fifty variations built on the nurses' own words, is worth more than five hundred variations of a generic "work in comfort" message.

The most reliable way to build those variations is from a library of layouts you already know work. The skill reads your customer language, applies it to each proven layout, and makes several versions of each. Even if only half of them are good enough to run, the math is lopsided once you count review time honestly.

Interactive tool

What a usable asset really costs

Count the review time and the ones you throw away. The example is a batch of static ads.

Generate

$100

to generate the batch

Review

60

$300

to review the batch

Keep

50

assets you'd run

Real cost

$100

to generate the batch

$300

to review the batch

50

assets you'd run

$8

per usable asset

Cost of one asset you'd run

Made the old way
$250
Made with AI, reviewed by a person
$8

real cost of each asset you'd actually ship

$8

Example numbers for static ads. Drop the usable share to 20% and the math still holds, which is why the review step matters more than the generation price.

Eric Helsel

What should you ask your customer data?

Ask for the exact words, the emotions and the gaps, because a summary hides all three. Your best material sits in reviews, post-purchase surveys, support tickets, sales calls and the comments under your ads. The organic feed and your competitors are useful too, but your own customers' words are where the winning messages hide.

Ask AI for the phrases customers use in the moment they decided to buy. Ask for the words only insiders would use. Ask for the ways people use your product that you have never advertised, and for the customer types who show up in reviews far more often than they show up in your ads. Ask what people complain about in your competitors. Each answer is a message you can build on.

Same customer data, two requests

The request that produces slop

Summarize our customer reviews.

A tidy list of themes every competitor already uses. Nothing to build a new ad on.

The request that finds new angles

Read our last 500 reviews and support tickets. Pull the exact phrases customers use when they describe the moment they decided to buy, group them by the problem they were trying to solve, and flag any reason that shows up often but never appears in our ads.

Customer language you can test this week, plus one or two angles nobody on the team had considered.

Eric Helsel

How do you get your team's tribal knowledge into AI?

Have your best people record themselves explaining why they make each call, then feed those transcripts into your skills. Tribal knowledge is the know-how that lives in people's heads and never makes it into a document. It is also the thing that separates your marketing from a competitor using the exact same AI tools you are.

An employee walks out of an office building at dusk carrying a cardboard box glowing with light.

Picture your best strategist going through ten new ad ideas. In fifteen minutes she keeps three and kills seven. Ask her why and you get an answer worth more than any prompt library you could buy: this one is too close to what a competitor ran, that one promises something sales cannot deliver, this headline only works for repeat buyers. None of that is written down anywhere.

The fix is simple and a little uncomfortable. Have your strategists, writers and editors hit record on a meeting tool while they work and narrate their reasoning out loud. Why this idea, why this line, why this cut, why this shot. The transcripts become the raw material your skills learn from. This is also why copying is no longer a moat: what a competitor cannot copy is how your people think.

How a strategist's judgment becomes a skill

  1. 1

    Hit record while you work

    Narrate every decision out loud as you make it.

  2. 2

    Get the transcript

    Your meeting tool does this automatically.

  3. 3

    Pull out the reasons

    AI turns the rambling into rules, examples and exceptions.

  4. 4

    Add them to the skill

    The skill now knows why, not just what.

  5. 5

    Run it and compare

    Check its choices against the strategist's. Repeat until they match.

Eric Helsel

It slows people down for a few weeks, and you will hear about it. Explaining yourself while you work is tiring. But imagine having a year of these recordings from every person on your team, covering every ad, every page and every edit. Skills trained on that would be extraordinarily hard for a competitor to match, because they would be trained on how your people think.

Your competitors can buy the same AI you can. They cannot buy the way your best marketer thinks, unless you let it walk out the door with her.

- Eric Helsel

What is a Claude skill, and why should your marketing team build them?

A Claude skill is a saved set of instructions, examples and files that teaches Claude how to do one job your way, every time the job comes up. On the mixing desk, a skill is the preset you saved after you finally got the sound right. Every workflow your team runs, from the weekly research digest to static ads to landing pages to creator outreach, can become one.

Here is why that matters to an owner: your team becomes exactly as good as its skills. A new hire who opens the shared library on day one produces work that already follows your layouts, your voice and your rules about what you never say. The best person's judgment stops being a bottleneck and becomes a resource the whole team uses. That is the core of what I mean by a marketing engineer.

The part most people miss is the feedback. When a skill gets something wrong, do not just reject the output. Tell it exactly what was wrong and why, in plain words: you moved the headline below the product on this layout, and next time keep it where the template puts it. Then update the skill. Do that every time and the skill learns your taste, the corrections get rarer, and the whole team benefits from every fix.

Five archery targets in a row, with arrows scattered on the first and tightly clustered in the bullseye on the last.

How a skill gets better every week

  1. 1.Run it

    The skill does the job from your instructions and examples.

  2. 2.Review the output

    A person checks it against your standard.

  3. 3.Say exactly why

    Not just no. What was wrong and what right looks like.

  4. 4.Update the skill

    The correction becomes part of the instructions.

  5. 5.Run it again

    Fewer fixes each round, until it sounds like your best person.

  6. and back to 1
Eric Helsel

Is a Claude skill just a prompt?

No. A prompt is a one-time request you type into a chat, and it disappears when the chat ends. A skill is a package that can hold instructions, examples, templates and even small scripts, and Claude loads it only when a task calls for it. It also improves over time, because every correction you make gets written back into it.

How do you share skills across a team?

Put them in one shared library that everyone works from, so a fix one person makes helps everyone the next morning. On Claude's Team and Enterprise plans, an organization owner can provision skills so they appear for every member. Teams working in Claude Code often keep their skills in a shared folder synced through GitHub.

Free marketing skill libraries are a fine place to start, but the skills that pay are the ones built from your own team's recorded judgment. The more people who use and correct them, the faster they improve, which is exactly why one shared library beats ten private copies.

Where should a $2M to $20M company start?

Start with one champion, three skills and ninety days, and measure output and quality before you touch headcount. Nine in ten CMOs are experimenting with AI, yet fewer than one in ten have scaled it or captured value across their marketing workflows. The gap is not the tools. It is the work of redesigning the team around them.

Pick your champion first and give them real time, not leftover Friday afternoons. Have them choose three workflows your team repeats every week, such as the weekly research digest, static ad variations and landing page first drafts. Start the recordings in week one, so the skills have your team's reasoning to learn from from the very beginning, and build an approval stop into every skill before anything goes live.

Put every skill in one shared library from day one, and review them together once a month: what the skills got right, what people keep correcting, and which workflow should move up a fader position next. Leave headcount decisions until after the first ninety days, once you can see real output and real quality. Rushing that call is one of the most common reasons AI projects fail. If you want to see how I think about building these systems, start with the marketing engineer model and the AI marketing agents that do the production work.

Anyone can rent a studio now. The equipment is cheap, the session players never get tired, and every competitor you have can book the same room. The records that sell still come from whoever knows what a hit sounds like, and the companies that win the next five years will be the ones that got that knowledge out of a few people's heads and into the desk while their competitors were still arguing about the gear.

This article is part of the AI Strategy guide.

Frequently Asked Questions

What is the typical hierarchy of a small marketing department?+
Traditionally, a marketing leader sits at the top with specialists underneath, each owning a channel such as content, design, social, paid media or operations. In an AI-first team, the hierarchy flattens: one leader owns the message and the revenue number, one or two senior marketers judge quality, and AI skills do most of the production underneath.
How does Claude know which skill to use?+
Claude checks which skills are available, decides which ones are relevant to what you asked, and loads only the parts it needs. That way a team can keep dozens of skills in its library without slowing anything down. Clear, specific names and descriptions help Claude pick the right one, so write them the way your team talks.
What is human in the loop AI?+
Human in the loop AI means a person reviews and approves the work at set points before it goes out. In marketing, the most useful checkpoint is approving headlines and copy before anything is produced in bulk, plus a final sign-off before anything goes live. It keeps speed high without letting off-brand work reach customers.
What is the difference between Claude skills and Projects?+
A Claude Project holds background knowledge, files and instructions that stay loaded in every chat inside that project. A skill is a reusable procedure that loads only when a task calls for it and works across Claude. For a marketing team, projects hold the context for a body of work, and skills hold the repeatable jobs.
What marketing jobs are safe from AI?+
The safest marketing jobs are built on judgment: positioning, creative direction, choosing which ideas get made, and owning a revenue number. Pure production roles are shrinking fast. The good news for anyone in those roles is that the production experience is exactly what makes someone good at training and correcting the AI that now does it.
Eric Helsel
Eric Helsel

Senior Marketing Executive

26 years of revenue marketing expertise. AI-native Senior Marketing Executive who builds the machine, not just the strategy.

26 years in revenue marketingAI-native executiveCMO and VP of Marketing
revenue marketingAI strategyB2B growthmarketing automationmarketing leadership

More from AI Strategy