Meta Said It Would Fire Every Ad Agency by 2026. Here's What Actually Happened.

Eight months ago, we ran a piece with a deliberately provocative headline: Meta wants to fire every ad agency by 2026. The thesis was simple. Meta had publicly committed to fully automated advertising by the end of this year. Upload a product image, set a budget, walk away. The AI handles creative, targeting, optimisation, and delivery across 3.4 billion users. No humans required.

We are now well into 2026. The deadline is eight months away.

So: did the agencies get fired?

The short answer: Meta sent the termination letter but forgot to file the paperwork.

The Scorecard Nobody Asked For (But Everyone Needs)

Credit where it's due — Meta has moved the needle. Advantage+ campaigns now account for roughly 60–70% of agency spending on Meta's platforms. The Andromeda ad retrieval engine has rolled out globally, and it's fundamentally changed how ads get matched to users. The system no longer targets a fixed campaign at a predefined audience segment. Instead, it evaluates creative assets against a broad user pool and determines, in real time, which message is most likely to convert for each individual impression.

The practical upshot: creative quality has become the dominant performance lever. A 2025 AppsFlyer study found that 70–80% of Meta ad performance is now driven by the creative itself, not budget or targeting configuration. Which is a polite way of saying that your painstakingly assembled lookalike audiences matter less than whether your ad looks like it was designed by a human or generated by a particularly ambitious microwave.

That's a meaningful shift. But it's not full automation. It is not the "one button" vision Zuckerberg described. And the gap between where Meta is and where it said it would be tells us something important about what happens when Silicon Valley confidence meets the messy reality of people who sell things for a living.

What the People Who Actually Press the Buttons Are Saying

The most useful data on Meta's progress doesn't come from earnings calls. It comes from the media buyers using the tools every day, and their reviews are... mixed. (In the way that "the restaurant had great ambiance but gave everyone food poisoning" is mixed.)

Morning Brew spoke to several agency leads in early April. Aaron Edwards, CEO of The Charles Group, described a platform systematically removing manual controls — broader audiences, fewer targeting levers, fewer ad sets per campaign — while cheerfully insisting this is all for advertisers' benefit. The system favours larger data sets and less human intervention. On those terms, it's working as designed.

But the creative automation tools — the part Meta actually needs for its "one button" vision — are meeting resistance that would make a French labour union proud. Hayley Owen, SVP at Deutsch, said she has yet to find a single client willing to hand creative over to Advantage+. The tools she's tested are limited to what she diplomatically described as "lower risk" adjustments: overlays, contrast tweaks, minor formatting changes. Not exactly the death of the advertising industry. More like the death of manually adjusting brightness.

Daniel Johnson, founding partner at We Scale Startups, was blunter. He said Meta's own creative AI tools consistently underperform the third-party systems he uses, and that Meta's reported performance numbers are "directionally useful but not specifically reliable." (Which is a very polite way of saying "we believe the vibes, not the decimals.")

And here's the irony that really ties this together: Edwards described a recent client brief that required 300 unique creative assets under a single concept. Once mapped across four target personas with five creative approaches, they needed 1,000 individual assets. Andromeda's appetite for creative volume is enormous — and generating that volume at brand-quality standards still falls to human teams, not Meta's AI. The system designed to replace agencies is currently generating more work for agencies. Zuckerberg automated the middleman into a bigger middleman.

The Black Box Problem (Or: Trust Us, We're Meta)

There's a structural tension running through Meta's automation push that never gets adequately addressed: the more it automates, the less advertisers can figure out what's working and why.

Multiple agencies reported performance instability in March — fewer leads, higher acquisition costs, inconsistent delivery — with limited ability to diagnose the cause. When a system operates as a black box, the platform's gains in aggregate efficiency come at the expense of individual advertiser transparency. Meta can claim a 22% ROAS improvement from Advantage+ campaigns versus manual setups. But when your cost per acquisition spikes 30% overnight with no explanation, "average efficiency gains" is not a phrase that soothes anyone signing the cheques.

Hawke Media's SVP of services noted ongoing issues with Advantage+ steering spend toward low-quality placements, low-engagement demographics, and poorly configured ad-level setups. His conclusion: human interaction will remain necessary "for quite some time." Which, in agency-speak, means "please don't fire us yet."

Meta's own response here is quietly telling. As of March 2026, the company introduced a feature to save opt-out preferences for Advantage+ creative across future campaigns. This is Meta tacitly admitting that its previous approach — silently re-enabling features advertisers had explicitly turned off — was eroding trust. (Imagine a restaurant that keeps putting anchovies back on your pizza after you send it back. Eventually you stop ordering pizza.)

Enter Muse Spark — And Why You Should Actually Pay Attention to This One

Last week, Meta launched Muse Spark, the first model from Meta Superintelligence Labs — the division assembled under former Scale AI CEO Alexandr Wang after a $14.3 billion investment. The model placed fourth on the Artificial Analysis Intelligence Index, behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. The benchmarks are not the story.

Three things about Muse Spark matter for anyone selling online.

First: it's closed-source. This is Meta — the company that built its entire AI credibility on the open-weight Llama family, which racked up 1.2 billion downloads by early 2026. Walking away from open-source isn't a product decision. It's a strategic declaration. Open models commoditise the infrastructure layer. Closed models protect the application layer — specifically, the advertising and commerce applications that generate 98% of Meta's $200 billion in annual revenue. Meta looked at its open-source AI reputation, looked at its ad revenue, and did some quick maths. (The maths won.)

Second: it has a dedicated Shopping mode. Muse Spark draws on creator content within Meta's ecosystem alongside individual user behaviour signals to power product recommendations. This is not a generic assistant bolted onto Instagram. It's a commerce-intent engine trained on the behavioural data of three billion users, sitting inside the apps where those users already scroll, message, and impulse-buy ceramics at 11pm. For ecommerce sellers, this is the feature buried in the product announcement that deserves the most attention.

Third — and this is the one that should genuinely change how you think about Meta advertising: the company is now feeding anonymised conversational data from over one billion monthly Meta AI users into its Advantage+ delivery targeting. Purchase intent, product research, life-event signals expressed in natural language conversations — all of it is becoming input to the ad targeting system. When someone tells Meta AI they're looking for running shoes for a half-marathon in September, that's not an inferred interest from browsing behaviour. That's stated intent. It's the difference between watching someone window-shop and having them walk up to you and say "I need shoes."

What This Actually Means If You Sell Things

Meta will not achieve fully autonomous advertising by December. The creative tools aren't good enough. Brand resistance is too entrenched. The transparency problems remain real. Regulatory friction across multiple jurisdictions isn't going anywhere.

But dismissing Meta's progress because it'll miss its own headline deadline would be a mistake. The direction is clear, and the practical implications are already here.

Creative volume is now a competitive input. Andromeda rewards diverse, genuinely distinct creative concepts across formats and personas. Not 50 variations of the same ad — five fundamentally different approaches, each adapted across multiple formats. The sellers who can produce that volume (through internal teams, agencies, or their own AI workflows) will outperform those who can't.

Your product data is your new targeting strategy. As manual targeting controls continue to disappear, the quality of your catalogue data — images, structured attributes, descriptions, reviews — becomes the primary signal the system uses to match your products to the right users. Machine-readable product information is what determines whether AI systems surface your products or your competitor's. We've been saying this for months. It remains true.

Conversational commerce signals are arriving faster than expected. The integration of Meta AI chat data into ad targeting is significant. Sellers in research-heavy categories — health and wellness, home goods, electronics, fitness — should expect targeting to improve materially as these signals mature. But the corollary is important: the system is learning from what consumers say, not just what they click. Brand presence in those conversations — through content, creator partnerships, product visibility — will increasingly influence ad delivery.

The differentiation question hasn't gone away. If Advantage+ absorbs the mechanical work of campaign management, the remaining competitive advantage is everything it can't automate: product quality, brand positioning, customer experience, and the strategic thinking that determines what story you tell and why. Sellers who treat AI ad tools as a replacement for strategy will find themselves competing on identical terms with every other seller feeding the same machine. Those who treat them as execution infrastructure while investing in real differentiation will find the tools amplify their advantages rather than flatten them.

The Bottom Line for Sellers

Meta didn't fire every ad agency. It won't fire them next year either. But it is steadily removing the work those agencies used to bill for — the targeting, the bid management, the placement optimisation, the basic creative iteration — and replacing it with systems that are good enough for most advertisers, most of the time.

Muse Spark's Shopping mode, the conversational targeting signals, the Andromeda creative-matching engine — these aren't abstract product announcements. They're the infrastructure of a system that will increasingly decide which products get shown to which consumers, and on what terms.

The question for sellers isn't whether this automation is coming. It's whether you're building the assets — the creative depth, the data quality, the brand clarity — that determine whether you direct the machine or the machine directs you.

The ones who wait for the "one button" to arrive may find they've already been pressed.

P.S. — Yes, Meta built a feature specifically to stop itself from re-enabling things you turned off. That's the 2026 version of a trust fall, except Meta kept dropping people until someone filed a complaint.

P.P.S. — If your reaction to "1,000 creative assets per brief" was mild panic, same. If your reaction was "finally, job security," you might be an agency strategist.

• GPT-Image-2 may have quietly surfaced under cryptic Arena aliases, and if early testers are right, OpenAI’s next image model could raise the bar on text rendering and real world knowledge.

• Reddit is becoming the internet’s highest intent discovery layer, with founders spotting buyers at the exact moment they ask for tool recommendations and are ready to act.

• A simple folder based knowledge base is having a moment, turning raw notes, scraped content, and AI summaries into a personal wiki without the usual app stack or setup chaos.

• Amazon’s Rufus ads are still tiny in scale, but cheaper clicks, richer prompt data, and clearer shopper signals are giving marketers an early playbook for conversational commerce.

• Google’s AI Overviews may now answer correctly about 90 percent of the time, but at Google scale that still means millions of wrong answers slipping into search every day.

• OpenAI reportedly sees ads becoming a massive revenue engine, with projections pointing from billions this year to a potential $100 billion business by 2030.

• Anthropic’s new Managed Agents aims to turn cloud hosted agent building into a much faster, cleaner process, giving teams production infrastructure without months of engineering lift.

🎯 Dart - A tool to manage tasks and workflows using AI assistance.

🔎 Outrank finds untapped keywords and automatically publishes daily SEO-optimized blog posts to grow your website's traffic

🧬 Parsio - Extract structured data from your PDFs, emails and other documents, automatically.

🎥 Pictory GPT - Effortlessly create captivating videos

🌐 Wonderish - Describe what you want to build. Wonderish makes beautiful websites, landing pages, funnels and more.

😍It was great, all I needed

🙂Good, not great

🙄It sucked