Midgrow Logo

Transform Your Business

Will AI Replace Marketing Agencies? The Honest Answer, With the Evidence

The best evidence says premium creative talent was hit hardest, not protected. What AI is actually displacing in agency work, and what it provably is not.

16 min read
AI and human collaboration in marketing with a robot hand and business professional working together

MIDGROW graphic about whether AI will replace marketing agencies, featuring a robotic hand and a human hand working together to represent artificial intelligence and human expertise. The visual highlights the role of AI in marketing, creative strategy, business growth, and teamwork, using navy, orange, and white branding.

The comfortable answer to this question is that AI replaces production work while strategy stays safe. We are an agency, so that answer suits us, and you should treat it with suspicion when you hear it from us or anyone like us.

Here is the problem with it. The single best piece of evidence on what happens to paid creative work when generative AI arrives says the opposite: the most experienced, highest-rated, best-paid people were hit hardest, not protected. That finding is the reason this article exists, and it changed how we think about what we actually sell.

So the honest answer is more specific and less reassuring than the version you will read elsewhere.

What does the best evidence actually show?

The closest thing to a controlled experiment on this question comes from Upwork. Researchers Xiang Hui and Oren Reshef of Washington University and Luofeng Zhou of NYU used platform data from January 2022 to April 2023 to compare freelancers in occupations exposed to generative AI against those in less-exposed ones, before and after ChatGPT's release in November 2022. Their working paper is the evidence everyone should be citing and almost nobody quotes correctly.

The measured effects:

  • Writing, proofreading and copy-editing: monthly jobs down about 2%, monthly earnings down about 5.2%
  • Image work — design, image editing, illustration: monthly jobs down about 3.7%, monthly earnings down about 9.4%

Those are real but modest numbers, and they matter precisely because they are modest. The figure circulating as a "21% collapse in creative freelance demand" does not appear in this paper. The market did not vanish.

But the distribution is the finding. The researchers expected experience and reputation to act as protection. The data pointed the other way. Freelancers rated higher on past job count, skill level, hourly rate and Upwork's "top-rated" badge saw larger drops in new work and income than the average. In the image analysis, freelancers with above-median prior earnings saw jobs fall 7% and income fall 14%, both statistically significant at the 1% level.

Two caveats that honesty requires. In the writing analysis the top-freelancer effect is weaker and often not statistically significant, especially for income — the authors call it suggestive rather than definitive. And this is a working paper under review, not a peer-reviewed publication, measuring short-run effects on one platform over seventeen months. Anyone describing it as undeniable peer-reviewed proof has not checked.

With those caveats, it still points somewhere uncomfortable, and it fits other research. Brynjolfsson, Li and Raymond's study of 5,179 customer support agents found a 14% average productivity gain, 34% for novices, and minimal effect on experienced staff. AI raises the floor. It does not raise the ceiling.

What is the Good-Enough Collapse?

Put those two findings together and you get the mechanism, which has nothing to do with AI being better than a skilled professional.

AI makes the output of an unskilled person acceptable. Once acceptable output is cheap and instant, the market stops paying a premium for excellent output — not because excellent stopped being better, but because the gap stopped being worth the price difference to the buyer. The skilled professional's advantage was never just quality. It was quality that nobody else could produce at all.

That is the Good-Enough Collapse, and it explains why the top-rated freelancer lost more than the average one. They had more premium to lose.

This is also why "move upmarket" is weaker advice than it sounds. The premium tier is not a shelter. It is the tier with the most exposed pricing, because it was charging the largest multiple over good-enough.

Is AI actually causing agency job losses?

Here the evidence pushes back on the alarming version, and the correction matters.

US advertising agency employment, in the Bureau of Labor Statistics series, ran as follows: 219,100 jobs in 2021, 247,100 in 2022, 250,600 in 2023, a peak of 251,400 in 2024, then 233,700 in 2025. That is a sharp fall of roughly 7% in a single year — and it comes after three years of growth, with the peak landing in 2024, two full years after ChatGPT launched. You can check the series yourself.

Whatever is happening to agency employment did not begin when the tools arrived. Claims of a sustained three-year AI-driven collapse do not survive contact with the data.

And when you read what agencies themselves disclosed, AI is rarely the stated cause. WPP's 2024 results reported like-for-like revenue less pass-through costs down 1.0%, with the decline attributed to client losses and a 21.2% drop in China — alongside restructuring driven by merging legacy agency brands. Those are lost accounts and consolidation synergies, not efficiency gains from automation. Publicis grew organically at 5.8% in the same period, which is hard to reconcile with a technology destroying the category.

The pattern worth noticing: agencies would much rather attribute redundancies to AI efficiency than to losing business, because the first reads as strategy and the second as decline. Treat every AI-attributed restructuring announcement as a claim to check against the same company's client wins and losses.

Then what actually is taking agency work?

In-housing, and it started long before any of this.

Association of National Advertisers data shows the share of members operating an in-house agency rising from 42% in 2008 to 78% in 2018 to 82% in 2023. The curve was almost fully travelled before generative AI was commercially available. AI did not start this. It removed the last capacity constraint on volume production for teams that had already decided to bring work inside.

The nuance that gets dropped: 92% of brands with in-house agencies still retain external agencies. In-housing moved volume, not everything. Where the ANA's longitudinal work traced what moved first, it was consistently production and adaptation — the artefact work — while strategic and specialist assignments stayed outside.

This sequence matters more than the headline percentage. Clients did not begin in-housing because production got cheap; they in-housed production and then production got cheap, which validated a decision they had already taken. An agency reading the last three years as an AI story is reading the second half of a fifteen-year story.

The second force is budget arithmetic. Gartner's 2026 CMO Spend Survey put marketing budgets flat at around 7.8% of company revenue, with 15.3% of marketing budget now allocated to AI and only about 30% of organisations ready to scale AI capabilities. New AI spending inside a frozen budget has to come from somewhere, and external suppliers are the softest line.

We should flag something here. Widely circulated figures claiming that 39% of CMOs plan to cut agency budgets and 22% have already reduced agency reliance because of AI are attributed to Gartner but do not appear in Gartner's release, which does not address agency spend. We checked this when writing our piece on agency pricing and found the same thing. The budget squeeze is real; that specific number is not Gartner's.

What has not been replaced?

Three things, each with a reason rather than an assertion.

Accountability for a decision. A model will generate forty campaign concepts. It will not be answerable for choosing the wrong one, and the cost of a wrong allocation decision scales with the budget behind it. This is the part clients keep buying, and it is why in-housing moved production rather than strategy.

Ownership of what gets produced. The United States Copyright Office has stated that copyright protects only material that is the product of human creativity, and that where a human supplies only a prompt and the technology executes the expressive work, the output is not protected by copyright — and must be disclaimed in a registration application. A human can copyright a sufficiently creative arrangement of AI-generated material, but not the generated material itself. The guidance is public. For anyone selling generative output as a proprietary brand asset, this is a live commercial problem, and most contracts have not caught up with it.

Distribution judgement. Google's position is that it rewards high-quality content however it is produced, while automation used primarily to manipulate rankings remains a spam violation. Production volume is therefore not an advantage in search, which is the thing everyone assumed AI would hand them. We covered what actually works now in our guide to SEO in 2026.

There is a fourth item we would add with less confidence, because the evidence is thinner: knowing when the model is wrong. The research on knowledge work found that people using generative tools on tasks just outside the model's competence performed meaningfully worse than people without the tools at all, and the boundary is not visible from inside the task. Someone has to know where it sits. That is a real skill and a real service, but we have not seen it measured as a market, so we are describing it rather than claiming it.

The Decision Test

If you are buying from an agency, or running one, this is the question that settles it.

For each line in the scope, ask: is someone answerable for a decision here, or is this the production of an artefact?

Writing forty ad variants is artefact production. Deciding which three to fund, at what budget, against which audience, and being accountable when it fails, is a decision. Building a report is artefact production. Interpreting it and changing the plan is a decision. Resizing creative is artefact production. Choosing the positioning the creative expresses is a decision.

Every artefact line in your scope is repriceable toward zero. Every decision line is not. If your scope is mostly artefacts, you are paying agency rates for something the Good-Enough Collapse has already repriced — and if you run an agency whose scope is mostly artefacts, that is the honest reading of your position.

What this means for Indian agencies specifically

This is where we should be most direct, because it is our own market.

A large share of Indian agency revenue is execution, and a large share of Indian agency competitiveness has been cost. Both of those are exactly what the Good-Enough Collapse prices down. A business built on producing volume more cheaply than someone else is competing against a technology whose marginal cost is approaching zero. Labour arbitrage does not help when the alternative is not labour.

We should also be honest about the evidence gap. There is no credible data on survival rates, revenue contraction or client churn among small and mid-market Indian agencies since 2023. The macro ad-spend numbers are growing, which tells you nothing about whether execution shops are surviving inside that growth. Anyone telling you confidently what is happening to Indian boutique agencies is inferring it, as are we.

What follows from the evidence we do have is uncomfortable and probably right: the Indian agency most at risk is not the cheapest one. It is the one whose entire pitch was being cheaper.

There is a second-order effect worth naming for Indian businesses on the buying side. For years the sensible move for a small Indian company was to outsource production to an agency because doing it internally was slower and worse. That calculation has genuinely changed for a narrow band of work — social captions, basic ad variants, first-draft copy, simple resizing — and a founder who now does that in-house with tools is making a correct decision, not a cheap one. The agency conversation should start after that band, not inside it. Any agency still pitching inside it is selling you something you no longer need to buy.

The strongest case that agencies are finished

We owe you the version that disagrees with us, properly made.

The premium tier is not protected — that is what the Upwork data shows, and it is the exact opposite of the industry's comfort narrative. In-housing has already reached saturation among large advertisers. Budgets are frozen while AI allocation rises, so agency spend is structurally the funding source. Agency employment fell 7% in a year. The reassuring material is published almost entirely by agencies, agency trade bodies and the consultancies that sell to them, delivered in annual reports and earnings calls where admitting obsolescence would be financially catastrophic. Meanwhile the pessimistic case rests on platform data, official labour statistics and anonymous buyer surveys.

On methodology, the pessimistic case is the better-sourced one. We think it is right about production and wrong about the end state, but we are not going to pretend it is the weaker argument.

The counter-evidence to the replacement thesis is thinner than you would hope. The AI-native agency category that was supposed to displace traditional agencies has produced very little verifiable scale — mostly launch announcements and funding rounds rather than disclosed client retention. The most heavily funded content-generation company of the last cycle, valued at $1.5 billion in 2022, ended up cutting staff repeatedly and repositioning as foundation models absorbed the function it was built on. We are not publishing its revenue figures because every version we found traces to secondary teardowns rather than disclosure. The pattern, though, is the durable lesson: the software layer selling generated output got commoditised faster than the agencies it was meant to replace.

Note also that the UK's Institute of Practitioners in Advertising published a Pricing Playbook urging agencies off full-time-equivalent and hourly billing. A trade body telling its own members that their commercial model is obsolete is not reassurance. It is a sector acknowledging the problem.

So what should you do?

If you buy from an agency. Run the Decision Test on your current scope. Pay for decisions and accountability; stop paying agency rates for artefact volume you can now produce or have produced far cheaper. Ask who owns the copyright in what you are being delivered, and get the answer in writing. Judge the relationship on allocation and outcome, not on output counts — which is the argument in how to measure marketing honestly.

If you run one. Your exposure is the proportion of revenue that is artefact production, and you already know the number. The move is not upmarket, because the premium tier is where the collapse bites hardest. It is toward work where someone has to be answerable — allocation, positioning, measurement, and the systems that connect them, which is what we set out in our AI marketing guide and our analytics piece.

Where we would decline work. If what you want is cheap volume production, we are the wrong choice and you should use tools directly. If you want someone accountable for what gets spent and why, that is a different engagement and it should be priced as one.

Want an honest read on your own position? Send us your current scope or your service list, what you pay or charge, and what the deliverables actually are. We will run the Decision Test against it and tell you which lines are exposed, which are defensible, and what we would change. If the answer is that you do not need an agency for most of it, we will say so — it is a worse sales outcome and a better answer. Talk to Midgrow, or message +91 9632398896 with your sector. Our engagement structures are on the services page, lead handling is covered in why lead generation fails, and content production sits with our social work.

Frequently asked questions

Will AI replace marketing agencies?

Not entirely, but it is repricing a large part of what agencies sold. The evidence shows displacement concentrated in artefact production — drafting, variant generation, resizing, reporting assembly — while work involving accountability for a decision has held. Agencies whose revenue was mostly production are materially exposed; those selling allocation and judgement are less so.

Does AI protect experienced marketers or hurt them?

The best available evidence says experience was not protection. In Upwork platform data covering ChatGPT's release, freelancers with higher ratings, higher rates and top-rated badges saw larger drops in work and income than the average, with above-median earners in image work down 7% in jobs and 14% in income. The premium tier had the most premium to lose.

Is AI causing agency layoffs?

Less than claimed. US advertising agency employment actually grew from 2021 through 2024, peaking two years after ChatGPT launched, then fell about 7% in 2025. Where agencies have disclosed reasons for restructuring, they more often cite lost client accounts, regional revenue declines and merger consolidation than AI efficiency.

What is actually taking work away from agencies?

In-housing and frozen budgets, both of which predate generative AI. Association of National Advertisers data shows in-house agency adoption rising from 42% in 2008 to 82% by 2023. Separately, marketing budgets have stayed near 7.8% of revenue while AI allocation has risen to 15.3%, so AI investment is funded by cutting elsewhere.

Can you copyright AI-generated marketing content?

Not the generated material itself, under United States Copyright Office guidance. Copyright protects human creative expression, so output produced by a model from a human prompt is not protected and must be disclaimed in a registration application. A sufficiently creative human arrangement of AI-generated material can be protected, but the underlying assets are not.

Does publishing AI content hurt your search rankings?

Not by itself. Google's stated position is that it rewards high-quality content however it is produced, while automation used primarily to manipulate rankings remains a spam violation. The practical consequence is that production volume is not a search advantage, which removes the main reason agencies expected AI to win them ranking work.

Should marketing agencies move upmarket to survive AI?

It is weaker advice than it sounds. The premium tier charged the largest multiple over acceptable output, so it has the most exposed pricing when acceptable output becomes cheap. The defensible direction is toward work where someone is answerable for a decision, which is not the same as simply charging more.

Are AI-native agencies replacing traditional agencies?

There is little verifiable evidence of scale. The category is heavily publicised through launches and funding announcements but has produced almost no disclosed client retention or revenue data. Notably, the most heavily funded generative content platform of the last cycle contracted sharply as foundation models absorbed its core function.

Share this article

Share this article

Midgrow

Midgrow

Contributing Author

Midgrow is a futuristic digital solutions and services studio based in Indore, Madhya Pradesh. We specialize in helping local businesses, startups, and industries grow online through high-performance websites, mobile apps, SEO, and creative digital marketing. With a passion for design, performance, and results, Midgrow is committed to transforming your business into a strong digital brand. From strategy to execution — we deliver premium experiences backed by data and creativity.

Stay Updated

Get the latest insights and tips delivered to your inbox weekly