Meta learned the hard way that automating everything is not the same as actually growing.
When Mark Zuckerberg announced his internal vision of turning the company into an organization fully centered on artificial intelligence, the plan seemed ambitious enough to become an industry benchmark. And it did — just not exactly for the reason he was hoping.
The initiative, dubbed OT — Organization Transformation, proposed replacing traditional teams with lean structures where AI agents would handle a large share of day-to-day tasks. According to a special investigation by Reuters, based on internal documents, recordings, and conversations with more than 20 people familiar with the process, Meta even evaluated scenarios where much smaller human teams would simply oversee operations run by automated agents.
The practical result was a paradox nobody at the company had properly anticipated: more automation generated more work, not more value. Technical incidents spiked, the time to resolve crises increased, and the second phase of restructuring ended up being canceled.
But the story doesn’t end there. While internal workflows were stumbling, Meta’s commercial side kept growing at a rapid pace — and that reveals something very important about where AI actually performs well and where it still needs experienced people in the cockpit.
If you work in marketing, tech, or manage teams that have already adopted some level of automation, this case study will make you rethink a few metrics you might be looking at with a bit too much optimism. 👇
What Project OT was actually proposing
The core idea behind Organization Transformation was simple in theory: reduce human decision-making layers, embed artificial intelligence agents into operational processes, and create a far more agile corporate structure. The most aggressive scenarios envisioned cuts of up to 60% in the size of certain teams, although Meta clarified to Reuters that this figure was never applied to the total workforce and that the exercises included layoffs, reassignments, and the elimination of open positions.
The organizational model was also different. Instead of traditional product teams of 10 to 20 people, the so-called AI-Native Playbook proposed pods of three to five members, with much more fluid roles and specialists shared across teams. In some cases, classic titles like product manager, designer, or engineer would simply vanish, replaced by a generic role: builder.
In practice, what happened was that workflows relying on human judgment to function well started breaking down quietly. The AI agents were efficient when the context was predictable and the data was well-organized, but they hit serious walls when non-standard situations came up — which, let’s be honest, are exactly the kind of situations that pop up most often at large companies operating in multiple markets around the world. This created an invisible pressure on the teams that were still around: they had to both supervise the agents and fix what automation couldn’t deliver.
When producing more stopped meaning being more productive
The most critical takeaway from Meta’s experience isn’t the headcount reduction — it’s what happened when the company started measuring the actual results of all that automation.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
According to internal documents reviewed by Reuters, the use of AI tools generated a 220% year-over-year increase in code changes across internal infrastructure and platforms. But the changes that actually translated into new or improved features for end users grew by only 36%.
In short: AI was multiplying output much faster than it was multiplying real results.
The problem got even worse when it came to system reliability. Reuters reported that significant technical and security incidents — including service outages and potential data leaks — jumped 40% year over year, while the time teams spent resolving those emergencies surged 70%.
This scenario explains why Zuckerberg ended up adjusting course. Meta carried out a workforce reduction of about 10% in May, but canceled plans for a second wave of restructuring that had been scheduled for November. Weeks later, he acknowledged in private conversations that the development of autonomous AI agents wasn’t progressing at the pace he had imagined.
Here’s the most important lesson: an organization can dramatically increase the amount of work it produces without proportionally improving the work that actually generates value. And this confusion is now showing up across modern marketing operations.
Why Meta’s ad business tells a different story
There’s a fundamental distinction that can’t be ignored. The reported stumbles mainly involve autonomous AI agents tasked with executing complex, multi-layered internal engineering at Meta. They are not the same as the ranking, optimization, and creative generation algorithms that power the company’s advertising engine.
In fact, the commercial side paints an almost paradoxical picture. In the second quarter of 2026, Meta generated $59.36 billion in ad revenue, a 27% year-over-year increase. Ad impressions rose 14% and the average price per ad went up 12%. The company attributes much of this commercial momentum directly to its AI systems.
Meta also reports that more than 8 million advertisers already use at least one of its generative AI creative tools, while continuing to integrate models like Muse Image directly into Advantage+ Creative.
The operational difference is clear:
- Selecting an audience, generating creative variations, or adjusting a bid are well-defined problems supported by massive data volumes and explicit success metrics like CTR, conversions, CPA, and ROAS.
- Coordinating people, prioritizing product roadmaps, managing edge cases, safeguarding security, and making high-level strategic decisions are far less defined problems.
AI tends to perform very well in contexts where variables are measurable, patterns are recognizable, and the feedback loop is short. Organizational risk spikes when executives mistake localized algorithmic efficiency for proof that automated agents can handle entire functions end-to-end without oversight.
Where AI performed well and where it stumbled
Here’s what makes this case so valuable for anyone thinking seriously about automation: not everything failed. Meta’s commercial side — especially the artificial intelligence-powered ad tools — kept growing consistently, in an environment rich with data, clear objectives, and constant performance feedback.
In internal workflows, the context is completely different: variables are often qualitative, problems are unique, and the consequences of a mistake can take weeks to surface — more than enough time for the damage to already be done.
What Meta’s experience makes clear is that the right question isn’t whether we should automate, but what makes sense to automate right now, given the resources and maturity we have today. Companies that skip this diagnostic step tend to discover, too late, that they replaced human complexity with technical complexity — and the latter is usually more expensive to manage than the former, especially when there’s no team trained to understand what’s happening under the hood.
The real risk for brands and agencies: confusing speed with efficiency
The advertising industry as a whole is hitting the same inflection point. A report from the IAB published in January found that 83% of ad executives surveyed already use AI in their creative workflows — up from 60% in the previous survey — with 64% citing cost reduction as the primary driver.
But a far less comfortable data point also emerged: more than 70% of marketing professionals using AI have already faced an AI-related operational incident, while fewer than 35% planned to increase investments in governance or brand safety oversight.
An agency can now generate 200 creative variations where it used to produce 20. A performance team can launch campaigns faster than ever. An advertiser can fully automate reports, analyses, and bid management. None of these capabilities automatically guarantee better business outcomes.
The essential question becomes: how many of those additional automated actions actually generate incremental value once you account for errors, rework, human oversight, brand integrity risks, and technology costs?
Governance: the missing ingredient
When you look at what went wrong with Meta’s Project OT, one word keeps coming up in the most careful analyses of the situation: governance. There was no shortage of technology, no shortage of budget, and certainly no shortage of ambition. What was missing was a clear framework for accountability, monitoring, and decision-making around the artificial intelligence agents that were embedded in the processes.
Governance in automation contexts isn’t bureaucracy. It’s actually the opposite: it’s what allows you to scale safely, because you know who’s responsible for each part of the system, how to measure whether it’s working as expected, and what to do when something goes off the rails. Without that framework, any AI-driven transformation is vulnerable to the same kind of paradox Meta ran into.
It’s worth noting that the governance challenge in AI isn’t unique to Meta. Companies of all sizes are dealing with smaller versions of the same problem as they adopt automation tools in their workflows. The difference is that at a company Meta’s size, the effects amplify very quickly. In a smaller business, you still have time to spot the issue and adjust before it becomes a crisis — but only if someone is watching the right indicators, with the right criteria, from the very start of implementation.
Meta is still betting big on advertising agents
Despite the internal reassessments, Meta is far from backing down on its broader AI ambitions. In June, the company introduced the Meta Business Agent, designed to answer user questions, recommend products, qualify leads, schedule appointments, and close sales transactions. Meta says more than one million businesses already use versions of these messaging agents on WhatsApp and Messenger.
On top of that, on August 19, Meta AI began connecting directly to business accounts on Facebook and Instagram, integrating with Meta Ads and Google Workspace to analyze campaign performance, generate operational insights, and optimize ad delivery in real time.
That’s why Project OT doesn’t mark the end of corporate automation. It signals the beginning of a much more demanding phase: a period where companies will need to prove that their AI agents drive real business impact, not just empty activity.
What this case changes in practice for anyone working with AI
If you’re implementing — or thinking about implementing — automation with artificial intelligence in your workflows, Meta’s case offers some concrete reflections that are worth more than any market trends list. The first is that speed of adoption without operational maturity is a dangerous combination. AI systems integrated into critical processes need time to be calibrated, monitored, and fine-tuned — and that time rarely shows up in the planning of projects under pressure to show results fast.
The second reflection is about the role of people within automated environments. Project OT operated on the assumption that smaller teams with more AI would be equivalent to larger teams with less AI. But what reality showed is that the quality of human judgment inside the system matters just as much as the quality of the algorithm. People with experience and context are what turn data into good decisions — and when you cut that human capital without properly replacing it, you’re not saving money, you’re removing exactly what made the system work.
The third — and perhaps most important — reflection is about metrics. Many teams look at productivity and efficiency as indicators of automation success. But Meta’s case shows that those metrics can be climbing while operational resilience is dropping — and you only realize it when a serious incident hits that the system can’t absorb on its own. Measuring the health of an automated workflow requires going beyond output and understanding the system’s response capacity, recovery time, and the quality of decisions being made within it.
At the end of the day, the real competitive advantage won’t belong to whoever generates the most ads or deploys the most agents. It will belong to whoever builds robust governance models, precise measurement systems, and workflows with well-designed human oversight around all of it. 🎯
