Skip to content
Back to Guides
Forecasting9 min read5 chapters

Compound Learning: Why AI Gets Smarter Over Time

How every marketing decision and its measured result feeds back into the next estimate.

Shubham Raghav

Chapter 1The Problem with Static AI

Most AI marketing tools work like a snapshot. They ingest your data, run a model, and spit out recommendations. The problem? That model doesn't learn from what happens next. It doesn't know if you followed the recommendation, what the outcome was, or how the market shifted since.

Static AI is essentially a fancy calculator. It's useful the first time, marginally useful the second time, and actively misleading by the tenth, because the market has moved and the model hasn't.

A model that doesn't learn from outcomes is just an opinion with math. It degrades in value over time as the market diverges from its training data.
DimensionStatic AICompound Learning AI
Training dataHistorical snapshotEach change's measured result, recorded as it lands
Over timeDrifts as the market movesHas more of your own results to draw on
PersonalizationGeneric benchmarksYour specific patterns
Outcome awarenessNoneEvery decision tracked
After months of useThe same modelA record of what each change did on your account

The difference matters most at scale. A brand spending heavily on ads can't afford recommendations based on stale data.

Ask how a tool learns

Some marketing AI tools never train on your specific data at all: they use generic benchmarks and call it “intelligence.” Ask whether a tool records what each recommendation predicted and what then happened on your account. If it does not, you're paying for the label, not the learning.

Chapter 2What Is Compound Learning

Compound learning is what happens when every marketing decision, and its outcome, feeds back into the model that made the recommendation. In Cresva, every change that is carried out is measured over 1, 7, 14 and 28 days, and each result is recorded as that window ends.

The name borrows from compound interest deliberately. Just as $1 invested grows because returns generate their own returns, a model that learns from every decision gets better because each improvement makes the next insight more accurate.

Compound learning isn't a feature, it's an architecture. The system is designed so that every output becomes an input for the next cycle. The model doesn't just predict; it watches what happened and adjusts.

Here's how it works in practice. Say the forecasting model predicts that shifting $10K from Google to Meta will improve blended ROAS by 12%. You make the shift. Three things happen:

  1. Prediction recorded

    The system logs the exact prediction: $10K shift, expected +12% ROAS, expected CPA change, expected revenue impact.

  2. Outcome observed

    Over 1, 7, 14 and 28 days after the change, the system measures what happened. Did ROAS improve? By how much?

  3. Model updates

    The measured result is recorded against the prediction. It sharpens this brand's next estimate for that kind of move, and is added to the anonymised pool only if the brand chose to share.

This cycle runs on every change that is carried out and measured. The velocity of learning is what separates compound learning from traditional model retraining.

Chapter 3The Feedback Loop Architecture

Compound learning requires a specific architecture. It's not something you bolt onto an existing system, the entire data pipeline has to be built around the concept of closed-loop feedback.

In Cresva, agents read one brand memory: the decisions on your account, the results measured after them, and what you told any agent. A change Sam proposes and the result measured after it are there for every other agent when it answers you.

Why shared memory matters

Most multi-model systems are siloed. The attribution model doesn't talk to the forecasting model. The creative analysis doesn't inform budget allocation. Every insight dies in the silo that generated it. Shared memory means one discovery improves every downstream decision.

The architecture has four layers:

Data Ingestion Layer

Unified pipeline from Meta, Google, TikTok, Shopify, GA4. Every event, every conversion, every impression, timestamped and normalized.

Agent Processing Layer

Specialized agents analyze data through their domain lens. Each agent produces insights and recommendations tagged with confidence levels.

Outcome Tracking Layer

Every change that is carried out is tracked against actual results. Predictions become scored data points: was the agent right, and by how much?

Model Update Layer

Measured results shift the estimate for the next similar move, and a forecast that keeps missing its range brakes the agent back to proposing.

The feedback loop isn't optional, it's the core product. Remove it and you have a static dashboard with AI branding. Keep it and you have a system that keeps a record of what each change did.

Chapter 4Shared Learning, Opt-In

Cresva learns mainly from your brand's own history: every decision on your account, and how it turned out. Beyond that there is one narrow shared starting point for the moves agents propose, built only from Cresva's own agent commerce records, and nothing a brand does is added to it unless an owner or admin of that brand chooses to share.

Who opts in. Every brand starts at “Keep it to this brand”. An owner or admin can choose “Share anonymised patterns” in Settings, under Data sharing. The choice is recorded with who made it and when, and it can be switched back at any time. The pool is rebuilt every night, so a brand that switches back is out of it by the next run.

What is pooled. Only the measured results of decisions a brand made while it was sharing, and only results built entirely from figures whose platform allows them to be used for other customers. Meta, Google, TikTok, Microsoft, Shopify and Klaviyo do not allow their data to be used for other customers, or their terms do not clearly allow it, so nothing from them is used. Today that leaves changes measured on Cresva's own records, such as agent requests to a brand's storefront and orders placed through Cresva's checkout. They are kept as ratios and counts, with no brand name, id, campaign or amount. A pooled figure exists only when at least 5 brands from at least 3 companies are in it. Brands that Cresva runs itself, and demo brands, are never pooled.

What every brand reads. When an agent proposes a move, its expected effect starts from the pool for that kind of move, preferring brands of a similar category, with the brand's own results taken out. The brand's own measured results then count for more as they build up. Any brand's proposals can start from the pool; only brands that share add to it. Below the floor there is no pooled figure, and the proposal says it has no comparable history instead of guessing.

What it is not

It is not a model trained on every brand's data, it holds nothing from an ad or store platform that does not allow it, and no brand's numbers are shown to another. It is one starting figure per kind of move, built from the results of brands that chose to share.

Who adds to the pool

Brands whose owner or admin chose to share anonymised patterns

What is pooled

Measured results of their decisions built only from Cresva's own records, as ratios and counts, from at least 5 brands and 3 companies

Who reads it

Every brand, whether or not it shares: a starting estimate for a proposed move, outweighed by the brand's own results as they build

Adding to the pool is off until a brand chooses it, the pool carries no names or amounts, and it only sets the starting point: a brand's own results decide.

Chapter 5How to Tell If a System Actually Compounds

Every vendor in this category now says their system learns. The claim is cheap and mostly unfalsifiable as stated, so here is how to test it during an evaluation. These questions work on us as well as on anyone else, which is the point of writing them down.

Ask what it records when a recommendation is rejected. This is the single most diagnostic question. A system that only learns from actions you took is learning from a sample selected by your own judgment, which teaches it your habits rather than the market. If rejected recommendations leave no trace, the loop is open, not closed. Cresva records a declined recommendation and who declined it, but it measures outcomes only for changes that were carried out, because a change nobody made has no result of its own.

Ask thisCompoundingRetraining, or neither
What happens to a recommendation you declined?Stored with the outcome that followedDiscarded
Can it show you a prediction it got wrong?Yes, with the recordNo prediction log exists
What improves when your data grows?Its calibration on your accountOnly its coverage
What transfers if you switch vendors?Your recorded decisions and outcomesA data export with no decision history

That last row is worth pressing on, and it cuts against every vendor including this one. Accumulated learning is a real switching cost, and a vendor benefits from it whether or not you do. The protection is to insist that your decision-and-outcome history is exportable in a form another system could read. If it is not, what the system has accumulated is leverage rather than an asset you own.

Compounding is a property of the feedback loop, not of the model. The test is not how capable the model is; it is whether the system records what it recommended, what you did, and what happened, and whether you can take that record with you.

Every change a Cresva agent carries out is recorded with what was expected and measured against what happened. That record sets the starting estimate for the next similar move, and the measured changes export as a CSV.

Written by Shubham Raghav, Founder, Cresva. Questions? Email us.