How proprietary transaction data becomes an enterprise AI advantage

Any company that can afford it can now license a frontier AI model, often the same one its competitors use. But regardless of the model, there’s data only your business provides: the transactions it has processed, the customers and partners behind them, and what happened afterward. A public model knows what fraud or churn looks like, but it doesn't know which of your customers switched phones this morning or which offer a shopper turned down yesterday.

"A general-purpose model can understand patterns, but it cannot know the continuously changing reality of your enterprise,” says Don Dama, Aerospike CEO. “That advantage lives in the proprietary transactions, relationships, and outcomes only your business can see."

Proprietary transaction data becomes an advantage when it helps the AI system reach a decision while that decision is still being made. That requires infrastructure that keeps the data current and delivers it in milliseconds. “Keeping intelligence connected to the continuously changing reality of the enterprise is the infrastructure challenge of the AI era,” Dama says.

Why does proprietary data matter more as AI models get easier to buy?

When every company can license a capable model, the model stops setting any one company apart. Instead, the proprietary data a business generates becomes the differentiating factor. 

The boardroom question has moved from whether a company is moving fast enough to whether its AI approach protects what makes it distinctive. Microsoft CEO Satya Nadella's warning is that companies risk giving away proprietary knowledge^[1], the kind "a competitor could never buy."

Investors have reached the same view. When TechCrunch surveyed 20 enterprise-focused VCs^[2], more than half named the quality or rarity of proprietary data as what gives an AI company its edge.

What counts as proprietary transaction data?

Proprietary transaction data is the record a business creates each time it does business with someone: a payment, an order, an ad bid, or a claim. Over time, those records add up to each customer's history; the relationships between accounts, devices, and partners; and the outcomes that followed. Every industry that processes transactions at volume has it.

Why isn't owning the data enough?

Transaction data becomes an advantage only when it reaches a decision in time to change it. A record waiting for tonight's batch job informs next quarter's plan, but it can't stop a fraudulent payment being authorized now or win an ad auction that closes in milliseconds.

A company's transaction history is valuable only when the right data is current and available to the system making the decision. "Proprietary data becomes an operating advantage only when it can influence the decision that is being made now," Dama says. "That requires an authoritative representation of what is true, the relevant context for the decision, and an architecture capable of delivering both within the time the decision allows."

Not every fact needs millisecond delivery. A customer's usual spending changes slowly, so an overnight refresh keeps it accurate. Facts, such as a card frozen a minute ago or an item just added to a cart, go stale in seconds, and a model working from yesterday's snapshot doesn’t see them.

The infrastructure behind a transaction-data advantage 

The advantage needs an infrastructure that gives each decision the right history and the latest facts within its time budget, at peak volume, even when part of the system fails. Warehouses and lakehouses are built to analyze history in bulk, and caches in front of slower systems can fall behind the system of record.

AI agents raise the stakes. 53% of enterprises already run AI agents in production, and IDC expects 80% of agentic AI use cases to need real-time, contextual data^[3], according to research director Devin Pratt. “As AI moves from analysis into action,” says Dama, “the value of intelligence increasingly depends on whether it begins from an authoritative understanding of what is true and can remain connected to that reality as its own actions change it.” 

Four requirements matter most:

• Current facts in milliseconds. Payment authorizations, ad bids, and checkout recommendations run on millisecond budgets. The data layer has to return the required records within that budget at p99 and higher, including p99.9 where the workload demands it.

• Availability when something breaks. When a lookup times out, the decision falls back to a default rule or a generic offer, and the company's proprietary data is wasted.

• A record of each decision and what followed. Chargebacks, returns, and cancellations often arrive weeks later. Learning from them requires a record of what the model saw when it decided.

• Years of history at a sustainable cost. Keeping years of records available in milliseconds gets expensive if it’s all stored in memory.

Why Aerospike is the real-time state, context, and decisioning platform for operational AI

Aerospike has spent more than a decade running this workload: decisions that read and update transaction data in milliseconds, at very large scale, across industries. Half of revenue came from large-scale operational AI and machine learning use cases^[4]. 

Current operational data at decision time

Barclays consolidated its fraud rules, models, and customer profiles onto one shared platform with Aerospike as the real-time data layer. Latency fell 80%, and the fraud dataset grew from 3 TB to more than 30 TB over three years without a redesign. A bank's history helps its fraud models only if they can read it inside the authorization window.

Availability when something breaks

Consider PhonePe, a digital payments platform that processes over 100 million daily transactions across India. To keep its fraud detection and feature store lookups running without interruption, PhonePe relies on Aerospike's active-active replication and high-availability architecture. The result is 99.999% uptime, meaning the system is down for less than five and a half minutes per year. By surviving hardware failures and routing traffic seamlessly across data centers, Aerospike ensures that PhonePe's models always have the current facts they need to make the right call, without ever falling back to generic rules.  

Recording each decision as it happens

AppsFlyer's mobile attribution platform, which credits app installs to the campaigns behind them, handles 2.5 million writes and 4.5 million reads per second with sub-millisecond latency. At that write rate, a company can record the context of every decision as it’s made and hold it until the outcome comes in (weeks later, for a chargeback or return), so the model learns from what it saw at the time.

AMD's grid computing monitoring platform tracks more than 20 million jobs a day on Aerospike, which shows that this need isn’t limited to customer transactions.

Extensive history at a sustainable cost

Aerospike's patented Hybrid Memory Architecture keeps the primary index in memory and the records on flash, so years of history stay available in milliseconds without holding it all in RAM. That meant PayPal needed 120 servers for its fraud platform instead of the 1,024 it had projected with its legacy in-memory system. That result is credited to the combination of Aerospike's Hybrid Memory Architecture, Intel Xeon processors, and Intel Optane persistent memory.

Aerospike stores the operational records each decision relies on and serves them within the decision's time budget. The application and surrounding data pipelines capture each decision and its later outcome, and Aerospike stores those records and serves them to the next decision. The model and application still decide and act, and exploratory analysis still belongs in the warehouse or lakehouse.

What this means for the next phase of enterprise AI

Competitors can license the same models, so the difference comes down to what each company’s AI knows when it decides. As AI takes over more enterprise business decisions, that becomes an infrastructure question. Each decision needs current facts and the relevant history inside its time budget, from a data layer that stays up when part of the system fails and records what the decision saw and what followed. Once agents are making those decisions at machine speed and volume, the business can’t operate without that layer. 

Aerospike already runs this workload for Barclays, PayPal, and AppsFlyer, the platform they depend on to put their proprietary transaction data to work.