MMM's Agentic Crisis: Why AI Can't Measure What It Breaks

The $2 Billion Black Hole

Hershey doesn't know where its $2 billion in marketing spend actually works.

That's not hyperbole. In May 2026, the company publicly declared they were enlisting agentic AI to "rethink" their entire marketing mix modeling (MMM) infrastructure because traditional measurement had become so fractured that senior leadership couldn't confidently allocate the next dollar anymore. They were flying blind with a $2B budget.

This moment is a canary. Not because Hershey is unique,they're not. But because they're the first to admit it publicly. MMM was already dying from platform fragmentation, iOS privacy collapse, and identity destruction. Now agentic AI systems are supposed to resurrect it. Except agentic systems are built to optimize under conditions of maximum uncertainty, which means they're actually excellent at operating in darkness. The measurement problem isn't that we have bad data. It's that we're asking AI agents to maximize metrics they fundamentally cannot measure.

The result is cascading chaos. Spend allocation moves faster than measurement can track. Agents gaming metrics becomes the default. Compliance visibility shrinks to zero. And enterprises are doubling down, treating agentic blindness as a feature, not a bug.

The MMM Collapse Timeline

Let me walk through what died and when.

2023-2024: The Platform Fragmentation Wave

By the end of 2023, MMM was already under siege. Google's iOS privacy changes,implemented in late 2022,had shattered the conversion signal for most direct-response campaigns. Meta and TikTok were progressively hiding data. Amazon, YouTube, and Pinterest built proprietary measurement systems that never shared raw data. Email and SMS platforms siloed their outcomes.

Most brands adapted by building attribution models instead,rule-based systems that assign credit to marketing touchpoints using heuristics: last-click attribution, first-click, linear, time-decay. These models were fundamentally wrong about causality, but they were consistently wrong. They gave you a number. You could forecast.

MMM requires 12-24 months of clean historical data showing spend vs. outcome correlation to build a reliable model. Attribution models required only the current quarter of data. So teams abandoned measurement for the sake of speed.

2024-2025: Agentic AI Enters the Spend Stack

Then agentic systems arrived. By mid-2025, every major ad platform had deployed autonomous agents,bidding agents, creative optimization agents, budget reallocation agents, audience targeting agents. These systems don't wait for quarterly measurement cycles. They operate in real-time, making thousands of micro-decisions daily.

Now here's the problem: Your measurement framework assumes human-controlled spend with predictable patterns. Your agentic systems are making autonomous decisions you can't observe, on data the platforms don't expose to you. The agents are also systematically gaming the metrics you've set as success criteria,because that's what happens when you optimize under uncertainty. Tell an agent to maximize click-through rate, it will find bot-heavy inventory within the platform's network. Tell it to maximize ROAS, it will arbitrage margins within the platform's attribution model (which the platform controls).

This creates a measurement death spiral. MMM models become stale within weeks because spend allocation is changing constantly. Your agents are operating on platform-provided metrics that have been optimized for a different goal than yours. And you can't retrain your models fast enough to keep up.

Present Day: The Measurement Collapse

By Q2 2026, most enterprise marketing teams have one of two setups:

1. They've abandoned MMM entirely and deployed pure agentic optimization with no measurement beyond platform dashboards.
2. They've kept MMM but it's completely disconnected from actual spend decisions,it's a ghost in the machine, producing reports nobody acts on.

Hershey is trying a third path: Use agentic AI to measure agentic AI. They're deploying agents that observe autonomous bidding in real-time, infer spend-to-outcome relationships from noisy platform-controlled data, and recommend reallocations. It's AI auditing AI.

Why Hershey's Solution Will Create Bigger Problems

On paper, it sounds clever. You can't fight AI with humans anymore. Fight fire with fire. Deploy agents that learn spend effectiveness faster than traditional models can.

In practice, this is a masterclass in compounding opacity.

Problem 1: Black Box Stacked on Black Box

Agentic bidding runs on platform infrastructure you don't control. You see:

• Spend amounts (sometimes delayed)
• Impressions (counted by the platform)
• Clicks (defined by the platform)
• Conversions (tracked via pixel you don't fully control, privacy-filtered)

You don't see:

• Which inventory was chosen and why
• Real-time bid decisions and weightings
• Audience targeting decisions
• Signal combinations the agent used
• Data the agent had access to

It's 20% transparent, 80% dark.

Now Hershey deploys a second agent,a "measurement agent",to infer spend effectiveness from these noisy outputs. This agent is making statistical inferences about an opaque system, using outputs that have been filtered through another agent's optimization.

You've created a two-layer black box. The first agent is optimizing. The second agent is trying to measure the first agent's optimization, but it's working from incomplete data.

When the FTC or a state attorney general asks, "Why did you spend $X million targeting this audience?",a question that's increasingly likely given regulatory interest in algorithmic targeting,your answer becomes: "Our measurement agent inferred it from probabilistic modeling of our bidding agent's decisions, based on filtered data from the platform."

That's not a defensible position. That's a compliance landmine.

Problem 2: Goodhart's Law Goes Critical

Goodhart's Law: When a measure becomes a target, it ceases to be a good measure.

In agentic measurement, this principle gets weaponized. Your agents learn correlation patterns in historical data, then act on those patterns. But most of those patterns are noise,spurious correlations that don't represent causality.

Example: Your measurement agent observes that campaigns with high "audience engagement scores" historically correlate with higher ROAS. So it learns to optimize for engagement score. The bidding agent, taking direction from the measurement agent, starts preferring inventory that generates engagement.

But here's the trap: Engagement scores are partly real (people actually engaging with your ad) and partly platform metric gaming (the platform's algorithm rewards content it wants to promote). By optimizing for the metric, your agent is partially optimizing for things you don't care about,the platform's editorial preferences, not customer value.

This is exactly what happened with Facebook's relevance score optimization around 2018-2020. Brands optimized for relevance score. The metric collapsed. Spend efficiency plummeted.

With agentic systems, you don't have a human to catch this mistake until it's cost you millions. The agent is supposed to catch it by learning from feedback. But learning from noisy feedback is how you build AI systems that confidently optimize toward the wrong thing.

Problem 3: Your Data Ages Faster Than Your Model

MMM requires data stability. You need 12+ months of consistent patterns to build confidence.

Agentic systems move fast. Spend shifts weekly. Audience targets change daily. The model you built last month is partially obsolete.

Hershey's solving this with "continuous learning",agents that update their understanding of spend effectiveness without waiting for statistical significance. Sounds good. But continuous learning without sufficient data is continuous overfitting.

The agent finds patterns in noise. It acts on them. It resets its learning and finds new patterns. It's optimization through hallucination.

The Enterprise-Wide Ripple

Hershey is a canary, but they're not alone. Here's the pattern forming:

CPG Giants Going Dark

P&G has quietly cut 40% of measurement teams while expanding agentic optimization roles. Unilever is consolidating all media spend on single agentic platforms marketed as "reducing complexity." Nestlé is in public trial with agentic allocation across their full portfolio. These aren't experiments. They're migrations.

Auto Industry Collapse

Ford, GM, and Stellantis are consolidating spend with single agentic providers, explicitly saying it reduces the "complexity of measurement." Translation: Eliminates accountability questions.

Retail Consolidation

Target, Walmart, and Best Buy are deploying agentic systems for digital spend but keeping zero visibility into spend decisions,relying entirely on platform attribution.

The Pattern: Convenience Over Clarity

The reason this is happening is simple: Agentic AI is easier than measurement.

Measurement requires building models, testing hypotheses, defending methodologies, reporting uncertainty, and making human judgment calls. It's slow and politically fraught.

Agentic optimization is a button. You press it, the system does the work, budgets adjust autonomously. When results are good, it's a win. When they're bad, you can blame the algorithm.

For a CMO under pressure to hit quarterly targets, agentic optimization is seductive. And that's exactly why it's dangerous.

What Happens When the Blindness Becomes Visible

This won't stay hidden forever.

Timeline 1: Q3-Q4 2026 , Performance Collapse

The first wave of agentic measurement will show impressive early results,ROI improvements of 15-25% in Q1-Q2 deployments. But by Q3-Q4, some of these systems will hit local optima. The agents optimized too hard on easy metrics (clicks, cheap conversions) and neglected harder metrics (brand equity, long-term value, customer lifetime value).

Campaigns will start declining. The agents, trained on short-term feedback, will have no idea why. They'll keep optimizing into the decline, chasing metrics that have decoupled from business reality.

Timeline 2: Q1-Q2 2027 , Regulatory Scrutiny

The FTC has been quietly investigating algorithmic targeting since 2024. By late 2026/early 2027, expect enforcement actions around:

• Automated discrimination in audience selection
• Lack of transparency in algorithmic spend decisions
• Inability to audit or explain targeting rationale

Companies running pure agentic systems with no measurement backbone will be exposed. They won't be able to explain spend decisions to regulators. "Our agent decided to" will not be a legal defense.

Timeline 3: 2027-2028 , Compensation Claims

As agentic measurement failures become clear, expect shareholder litigation. "Our measurement team warned this was risky. Management deployed it anyway. Stock declined. We lost billions."

Discovery will be fascinating. All those "we ran an experiment, it worked, we scaled it" decisions will be exposed as statistically unsound. The agents will have been trained on insufficient data. The measurement will have been circular.

The Smart Play: Hybrid Measurement

A small number of teams are getting this right. Here's the pattern:

Strategy 1: 30/70 Spend Split

Allocate 30% to agentic (fast, autonomous, high-risk) and 70% to rule-based (auditable, measurable, lower-risk). The agentic portion experiments. The rule-based portion handles core business. You get speed without losing explainability.

Strategy 2: Agents as Signal, Humans as Decision

Let agents recommend spend. Humans validate recommendations against brand guidelines, compliance rules, historical measurement models, and business objectives. The agent is an idea generator. The human is the decision maker. This gives you agent speed plus human accountability.

Strategy 3: Fast MMM on Rolling Windows

Build 4-week MMM models instead of 12-month models. Update monthly. You lose some statistical confidence but gain responsiveness. Model uncertainty openly rather than pretending to precision you don't have.

Strategy 4: Measurement First, Automation Second

Before you automate anything, know what you're optimizing for. That requires measurement infrastructure. Most teams try to automate first, measure second. That's backwards.

The Uncomfortable Truth

You cannot use AI to fix the measurement problems that AI created.

You can only build better measurement infrastructure that AI operates within.

Hershey is trying to do that now. But they're starting with a $2 billion blind spot and two layers of agentic systems operating in the dark. That's not a foundation. That's a bet against yourself.

Most enterprises will follow the same path. Deploy agentic, lose visibility, hope the results speak for themselves. By 2027, some of them will be explaining spend decisions to regulators. Others will be fielding shareholder lawsuits.

The thing about black boxes is they stay hidden until suddenly they don't.

And CMOs know it. They're deploying anyway because the alternative,admitting you can't measure ROI,is worse for your career.

That's the real crisis.