There is a persistent assumption in the data industry that intelligence requires centralisation. You collect data, send it somewhere large, process it in the cloud, and return a result. This assumption is so embedded in how enterprise technology is built that questioning it feels almost theoretical. But it has a concrete cost: latency, privacy exposure, regulatory risk, and a fundamental distance between where behaviour happens and where it is understood.
Edge AI starts from a different premise. What if the inference happened where the data was generated — on the device, at the moment of behaviour, without anything personal ever leaving the handset?
The Push Problem
The conventional approach to personalisation is a push model. Customer data must reside in the cloud in order to determine which person should receive which message. The data flows from device to cloud, the decision is made centrally, and the message is pushed back. This architecture made sense when it was designed. It no longer makes sense in a world where privacy regulation has fundamentally changed the cost and risk of centralising personal data, and where the intelligence available from first-party behavioural signals on the device itself is richer than anything the cloud can synthesise from the same data after the fact.
Edge AI inverts the model. In the pull architecture, only general campaign logic resides in the cloud. The personalisation decision happens on the device — where the customer's behavioural context is live, where their PII stays, and where no regulatory exposure is created. A fitness enthusiast, detected by their morning run pattern and device activity signals, receives a contextually relevant offer at the moment they open their phone after exercise — not because a cloud system sent them a targeted push, but because the intelligence on their device recognised the moment and pulled the relevant message. The result is indistinguishable to the customer. The difference in privacy architecture is categorical.
What 8,000 Data Points Means
The SDK draws on 50 distinct signal types to process up to 8,000 data points per device, generating 500 structured behavioural insights. The signal taxonomy is broader than most organisations expect. It covers:
- Daily Routine — wake time, sleep patterns, home and work arrival, commute behaviour
- Health and Wellness — physical activity types and frequency
- Travel — international trips, countries visited, time abroad, commuting mode
- Financial — financial app usage, digital payment patterns, affluence indicators
- Device & Connectivity — device age, Wi-Fi vs mobile data usage, battery patterns
- Consumer Preferences — brand interactions, app categories, content behaviour
- Demographics — age prediction, gender prediction, household size inference
- Contextual — point-of-interest data, best moment to engage signals
- Lifestyle — car ownership indicators, fitness segmentation, life stage signals
No PII collection is required at any stage. The SDK operates while the app is closed. Approximately 90% of signals require no OS permissions from the user.
The practical consequence of this signal depth is that the Edge SDK produces ground truth about customer context that no other data source can replicate. A CRM knows what a customer bought last month. The Edge SDK knows what they are doing right now — whether they are commuting, at work, in an exercise session, or sitting at home in the evening. This contextual layer is what makes Best Moment to Engage possible: not just identifying the right message, but identifying the precise window when the customer is actually receptive to receiving it.
The Model Accuracy Argument
One of the less discussed commercial advantages of Edge AI is what it does to the accuracy of models that organisations already have. Churn prediction models, LTV models, propensity scores — these are trained on whatever data a company has available, which is typically transactional and historical. They work. They would work better with more information about the customer's current context.
Intent has demonstrated a 74% improvement in prediction accuracy when Edge behavioural signals are combined with existing customer data. The improvement is visible across all model quality metrics: true positive rate increases, false positive rate decreases. The mechanism is straightforward — Edge data provides ground truth about context that the model was previously inferring from proxies. A churn model that previously had to infer disengagement from declining transaction frequency can now observe it directly from behavioural patterns on the device. The inference becomes recognition, and recognition is more accurate than inference.
This has direct commercial consequences. A 74% improvement in prediction accuracy on a churn model operating across millions of subscribers is not an incremental gain. It is a structural shift in the quality of the commercial decisions the model supports.
Proof at Scale
The case study evidence is concrete. At EVEST, an online investment and trading platform, Edge AI was used to detect and engage non-registered and dormant users and prompt first-time deposits. Against a control group running standard engagement campaigns in parallel, the results were: 46% increase in conversion rate, 27% increase in monthly active users, 84% increase in deposit volume. These are not A/B test margins. They are the kind of uplift that changes how an organisation thinks about the commercial value of its mobile channel.
At Aditya Birla Capital, one of India's largest financial services groups, a mutual fund campaign delivered at the Best Moment to Engage — running for one week, presented to users two to three times — achieved 9x CTR performance against the control group. The underlying mechanism was timing: reaching customers at the moment their behavioural signals indicated readiness to consider a financial decision, rather than broadcasting at a scheduled time to a broad segment.
The Permission-Free Architecture
The permission question is worth addressing directly because it is frequently raised as a concern. Most mobile analytics depend on permissions that users increasingly decline — location, contact access, tracking authorisation. The regulatory trend, particularly Apple's App Tracking Transparency framework, has made this dependency progressively more damaging for systems built on it.
Edge AI was built to operate without these permissions. The 90% permission-free signal architecture is not a fallback position — it is the design. The signals available without OS-level permissions turn out to be extraordinarily rich. Interaction velocity, device context, screen behaviour, app activity patterns, timing — none of these require the user to grant access. They are the ambient data of how a device is used, and they reveal more about a customer's current context than the location data that most personalisation systems have historically depended on.
The remaining 10% of signals — those that do require permissions — are valuable when available but not structurally necessary. The SDK degrades gracefully when permissions are absent, rather than silently failing or producing misleading outputs.
The IP Position
The Edge AI capability is protected by an extensive portfolio covering the specific engineering problems that make on-device intelligence work at production scale — signal extraction without permissions, battery-efficient processing, OTA model delivery, background operation across iOS and Android, privacy-preserving signal handling. These are not broad software patents. They are specific solutions to specific hard problems, built over a decade of production deployments.
Integration
Integration is a single line of code. The SDK is lightweight, App Store compliant, and designed to be adopted without engineering friction. The intelligence, once deployed, operates invisibly — no drain, no lag, no user-facing indication that anything is happening beyond the relevance of the experiences the SDK enables.
The question it poses to any organisation with a mobile application is straightforward: your customers are generating 8,000 data points per day on their devices. Are you using them?