The $75M playbook hiding in 1.5 years of campaign data
A global CPG had its strongest moves on record and didn't know it. Here's what happened when the system could finally read them.
For CMOs & marketing leaders. Companion piece for architecture leaders: $47M in validated decision patterns from 1.5 years of marketing archive →
Every marketing organization has a version of the same problem. The team makes a call, the call works, someone writes it down, and a year later no one remembers it happened. The brand manager who made the call has rotated. The agency has changed. The campaign is closed. The learning, theoretically captured in a readout or a dashboard, lives somewhere no one looks before the next planning sheet starts.
This is the quietest, most expensive form of the insight-to-action gap. Slow reporting gets fixed. Bad data gets cleaned. Institutional forgetting persists, silently, across hundreds of decisions per quarter on a portfolio where any one of those decisions can be worth tens of millions of dollars in sell-through.
Here's what that looked like at one global CPG.
What the campaign archive actually looked like
The client runs a multi-hundred-brand portfolio and thousands of campaigns a year. Millions of decisions sit logged across quarterly readouts, agency reports, performance dashboards, and brand manager handovers. Every campaign generates signals about what worked and what didn't. Almost none of it transfers.
A brand manager pivots an audience mid-campaign on Brand A. Sales climb 58% year-over-year, a top-5 outcome in the brand's entire history. The decision gets documented. The team celebrates. Six months later, that brand manager rotates. The next campaign on Brand A starts from a clean planning sheet. A year after that, Brand B faces the same setup: healthy YoY, running campaigns, custom audiences. The new team makes a different call. No one in the room remembers the +58% pattern.
The team's instincts about which moves worked were good. They didn't have a system that could find the patterns at scale, codify them with confidence, and tell them where to lean in and where to pull back.
What 1.5 years of campaign data revealed
A.Team unified 1.5 years of the brand's campaign data into a single decision log: context, action, outcome, all connected to the brand's own business meaning. Six patterns surfaced with enough evidence to codify into a playbook. Three "do this" patterns worth $47M in cumulative lift. Three "avoid this" patterns worth $28M in damage averted. A $75M signal sitting in the archive, invisible until the system could read it.
The same record, distilled into a playbook every downstream agent can read: three moves to lean into, three to pull back from, each weighted by dollar impact and confidence.
The strongest "do this" pattern: when the brand was already in a healthy stretch with campaigns running on custom audiences, shifting targeting on one running campaign drove the biggest lifts. Same creative, same channels, sharper audience. One such pivot delivered the +58% YoY outcome. Across seven episodes in the data, the same pattern held. The system codified it as a play at 92% confidence, worth $11.7M in cumulative lift. The takeaway: when momentum is already on your side, sharpen targeting on what's already running.
The corresponding "avoid this" pattern was the same move in the wrong context. Audience pivots applied during declining brand momentum, without a creative refresh, did the opposite. -62% YoY in one stretch. -60% two months later. -57% two months after that. The same lever, applied in the wrong context, was destroying value.
Four other patterns rounded out the playbook: leaning into Commerce & Search during healthy stretches (+$25.5M), custom audiences over broad audiences when momentum exists (+$10M), heavy campaign activity during declining periods (-$9.2M), and frequent pivots without creative refreshes (-$7.3M).
What we built
A.Team builds learning systems for marketing teams on Assemble, the platform that makes enterprise data accessible to AI agents and gets sharper with every cycle. For this client, the system runs in five steps and shows the Continuous Learning capability of Marketing & Media Performance in its most direct form.
It unifies the data. Assemble connects to the brand's existing systems (sales data, campaign performance, audience definitions, market context) and brings them into one structured layer where AI agents can read them the way a member of the marketing team would.
It captures every campaign decision. Past and active campaigns get structured into a decision log: context, action, outcome, all tied to the brand's own business meaning. History that had been scattered across readouts, reports, and dashboards now lives in one place.
A slice of Brand A's decision log: every move, its measured outcome, and the pattern the system anchored to each.
It codifies the patterns. When the same context-action-outcome sequence repeats often enough, the system turns it into a play with a confidence score, a dollar value, and trigger conditions for when it applies.
It surfaces recommendations where brand managers work. "Do this" plays appear with their dollar values and confidence scores. "Avoid this" plays appear with the cost of getting it wrong. Both embedded inside the workflow tools the team already uses to plan campaigns.
It transfers proven plays across brands. Patterns that mature on one brand become available across the portfolio. An AI plus human curator reviews whether the trigger conditions are brand-specific or generalizable. If they're brand-agnostic, the pattern's scope expands from one brand to the whole portfolio.
What compounded next
The first cross-brand transfer came months later. Brand B hit the same conditions Brand A's winning pivot had been codified against: healthy YoY, active campaigns, custom audiences. The system surfaced the play (sharpen targeting on running campaigns, defer creative refresh) with $9M of projected lift based on the pattern's track record. A piece of brand-specific institutional knowledge had become a portfolio-level asset.
The system today runs at 67% predictive accuracy on the original brand. Two out of three times, it correctly anticipates whether a planned campaign decision will hit or miss before it runs. That's the baseline.
What that playbook is worth over three years: a projected 2.7× lift over baseline as learnings accumulate, mature, and fully integrate.
The forward case is where the model gets interesting. A learning system's value compounds across a curve, not a snapshot. Projected three-year cumulative incremental sales reach $5.8M at the current baseline. If the system matures during the first 12 to 18 months of running, the projection reaches $10.3M. At full maturity (validated across the portfolio and integrated into planning workflows), the projection reaches $15.6M. A 2.7x lift over the baseline, earned entirely from the system getting smarter at the work.
Most marketing tech depreciates. Learning systems accumulate.
What this unlocks
Two things change once a system like this is running.
Good decisions get cheaper. A pattern with 92% confidence and $11.7M of historical value doesn't need a six-week analysis cycle to justify. It needs a thirty-second check that the trigger conditions hold.
Mistakes get visible earlier. The avoid pattern from the decision log becomes a high-confidence "do not" in the playbook, coded with the cost of getting it wrong. The next time a brand manager faces those conditions, the play surfaces before the decision gets made.
Both moves are versions of what a great marketing team does already. The system does them at the scale of the portfolio, indefinitely.
The diagnostic question for marketing learning
When your strongest brand manager rotates next quarter, how much of what they figured out walks with them? And when the next manager faces the same conditions on a different brand, is there anything in the system that tells them what's already been learned?
If the answer to the first question is "most of it" and the second is "no," the architecture is the constraint. The distance between knowing something and being able to find it on the day it matters is where most marketing organizations are losing the most money.
The $75M playbook wasn't unique to this client. Any brand portfolio with a few years of decisions has its own patterns hidden in the same way. The question is whether a system that can find them is running.
See how Marketing & Media Performance works → or Request your Strategic Intelligence Gap Assessment
This is the fourth essay in A.Team's series on how enterprise marketing and planning organizations are closing the insight-to-action gap. It describes a client engagement delivered through A.Team's AI Solutions practice. The client name and brand identifiers are withheld under NDA. Dollar figures reflect the engagement as documented in the client's own performance systems and Assemble's modeled forward case. Realized cumulative pattern value (+$47M / -$28M) reflects historical performance attributed to the codified plays across the 1.5-year dataset. Three-year cumulative projections ($5.8M / $10.3M / $15.6M) reflect Assemble's compounding-intelligence model under three maturity scenarios.
Frequently asked questions
How marketing learning systems codify patterns from historical campaign decisions, transfer plays across brands, and compound value as they mature.
A marketing learning system codifies the patterns hidden in a brand's historical campaign decisions into a playbook of plays, each scored by confidence, dollar value, and the conditions that trigger it. The system surfaces "do this" recommendations when those conditions repeat and "avoid this" warnings when they don't. Where dashboards report on the past, a learning system applies the past to the next decision.
The bottleneck is in how the data is structured, not in how much analysis the team does. Most CPG brands have years of campaign decisions logged across readouts, dashboards, and agency reports, but the decisions aren't connected to outcomes in a way machines can read. Unifying campaign performance, sales data, audience definitions, and market context into a single decision log makes the patterns findable. With enough volume, the system can codify recurring plays at high confidence and surface them when the conditions repeat.
Compounding intelligence is the property of a marketing system that gets sharper with every decision the team makes, rather than resetting between campaigns. Where static dashboards or one-time AI implementations stay the same the day after they're built, a compounding system retains context, learns from outcomes, and creates a competitive advantage that widens over time. A.Team's litmus test: if you turned off your AI, would you lose intelligence or just a tool?
Cross-brand learning transfer is when a pattern codified on one brand becomes applicable across a portfolio. A play with high confidence on Brand A (for example, sharpen audience targeting on running campaigns when brand momentum is healthy) gets reviewed for whether its trigger conditions are brand-specific or generalizable. If they're brand-agnostic, the play surfaces automatically for any brand in the portfolio that hits the same conditions.
By structuring every campaign as a context-action-outcome record, then watching for sequences that repeat with similar results. When the same pattern surfaces enough times with statistically significant lift, the system codifies it as a play, assigning a confidence score, a historical dollar value, and trigger conditions that tell brand managers when the play applies. Plays mature as they accumulate more evidence across the portfolio.
For one global CPG, A.Team's system unified 1.5 years of campaign data and surfaced a $75M signal: $47M in cumulative lift from three "do this" patterns and $28M in damage averted from three "avoid this" patterns. Forward modeling shows three-year cumulative incremental sales of $5.8M at baseline, $10.3M when the system matures during the first 12 to 18 months, and $15.6M at full portfolio maturity. A 2.7x lift over the baseline, earned from the system getting smarter at the work.
A dashboard reports what happened. A learning system applies what happened to the next decision. Dashboards aggregate performance metrics across campaigns; learning systems codify the patterns inside those metrics into actionable plays with confidence scores and dollar values. The system surfaces recommendations inside the workflow tools brand managers already use to plan campaigns, not in a separate analytics interface.
A learning system reaches initial codification within months as enough decision-outcome data accumulates to identify high-confidence patterns. Full maturity (patterns validated across the portfolio, embedded in planning workflows, and accumulating new learnings continuously) typically takes 12 to 18 months. The system's value compounds over that maturity curve: from baseline cumulative lift to a multiple-x gain as patterns prove out across brands.
$180M in 90 days
A global beverage company's media team was optimizing against the wrong signal. Here's what happened when the system connected 16 platforms and showed them what the data actually said.
Beyond market-mix modeling
MMM tells you what worked last quarter. Here's what changes when attribution runs continuously instead of annually.
Beyond metrics theater: Measuring AI impact that actually matters
A July 2025 MIT study found that 95% of companies saw zero measurable return on their in-house AI investments. Most enterprises are engaged in metrics theater: tracking what's easy to measure while ignoring what drives actual business value.