Fundamentally, AI is only as good as the data it has access to. Ask today’s most widely used LLM something it doesn’t have sufficient data on, and it will attempt to fill in the gaps through language prediction. Instead of admitting the shortcoming, the AI will likely answer confidently, convincingly, and often incorrectly. This is referred to as a hallucination.
Brian Mandelbaum, CEO and co-founder of Attain, built his business on high-quality purchase data that predicts sales outcomes with accuracy—helping marketers avoid performance “hallucinations” that come from low-quality data and vanity metrics. The Outcome spoke with Mandelbaum about how AI is reshaping the use of data across the industry and where it’s headed next.
AI is not new to advertising. Ad serving, look-alikes, and optimization have run on machine learning for years. What has changed is latency and learnable surface area. We can now process more context in less time, which allows us to move from broad heuristics to precise, outcome-driven decisions at impression scale.
Compute is abundant, but truth is scarce. Without proprietary, permissioned outcome data, models gravitate to proxies like clicks, views, and engagement. This creates “performance hallucinations,” akin to what we see when AI doesn’t have sufficient information on a subject and then creates its own false reality. The ground truth is not the algorithm, it’s the data that AI is trained on. When we launched Attain, we believed that first-party, deterministic purchase signals would be the strategic asset. That is what transforms AI from a commodity into a compounding advantage.
Prediction without verifiable training data is just storytelling. Deterministic purchase data is the model’s conscience because it tells you what actually happened. It calibrates propensity models so they optimize to sales rather than vanity metrics. It reduces leakage from probabilistic linking. And it enables continuous learning loops where every new conversion improves the next forecast.
The unlock is not simply better targeting. It is pre-training and fine-tuning on outcome data, then using reinforcement from real commerce events to keep models honest. That is how you achieve both precision and scale without having to trade one for the other.
It’s hard to pick just one, there are very interesting capabilities AI unlocks within each phase of a campaign.
For audience and consumer insights, we’re moving from research snapshots to living maps. AI can synthesize billions of weak signals into actionable motivations: what people value, what trade-offs they accept, and the moments that tip them into purchase.
Planning can now be based on simulation instead of guesswork. You can stress-test audiences, channels, and creative before you spend, then allocate dollars like a portfolio manager balancing risk and return.
Execution is now sub-100 millisecond decisioning that is context-aware, price-aware, and outcome-aware. Think of it as autonomous media trading: bid the right inventory, with the right message, at the right time, with guardrails for brand, cost, and incrementality.
We’re shifting from performance reporting to actually understanding causality. Historically, measurement was slow and human-heavy, with batch jobs and post-hoc attribution arguments. Today, models can run always-on experiments, update priors in near real time, and price media on expected incremental outcomes rather than raw conversions.
The practical result is cheaper, faster feedback loops, where MMM and lift testing inform each other, and controls for selection bias are baked into the system. If optimization is not grounded in incrementality and lifetime value, then you are just accelerating the wrong answer.
Advertising will become a conversation between brand systems and personal agents. Instead of pages of search results, you will see agentic recommendations that are sponsored but constrained by your preferences, budget, and past behavior. Ads will be machine-readable offers with guarantees, not just messages.
Persuasion will shift to negotiation. Your agent will know you prefer Sony, the price ceiling for a sound bar, and what integrates with your home. Brands will compete to satisfy the agent’s constraints. Loyalty will become a set of model weights rather than a punch card.
The core risk is incentive misalignment. If an agent auto-buys paper towels, is it because that was best for you, or because an affiliate payout nudged the model? Add opaque data provenance, model drift, and fairness concerns, and you risk eroding trust.
The solution is governance at the protocol layer: auditable logs, spend limits, preference contracts, and clear disclosures when incentives influence recommendations. Think of it as nutrition labels for AI decisions. Autonomy should be earned with performance and transparency, not assumed by default.
The Internet collapsed the cost of information. AI is collapsing the cost of decisioning. That makes trust, not attention, the scarce resource. Brands that align optimization with consumer outcomes will compound. Those that chase shortcuts will burn trust faster than they buy reach.
