Scale is the most overused word in enterprise technology. Every platform claims to scale. What that claim usually means is: our system did not break at the volume of data in our reference customer's deployment, and we have a white paper explaining why it will not break at ten times that volume. What it rarely means is: we have operated continuously, in production, at extreme scale, for multiple years, across one of the most demanding infrastructure environments in the world, and here is precisely what we learned.
Intent's scale credentials come from managing one of the most complex customer intelligence deployments anywhere, with over 100m customers. Telco scale generates a qualitatively different set of engineering challenges from a mid-market retailer or a consumer app. The data volumes are in a different order of magnitude. The regulatory requirements are stricter. The latency tolerances are lower. The consequence of system failure — in customer experience terms, in revenue terms, in regulatory terms — is higher. And the diversity of data types and sources is greater than almost any other industry context.
The numbers that define the current platform are not aspirational benchmarks. They are operational realities: 320 million active profiles managed simultaneously; 250 billion events processed per day; 25 terabytes of new data ingested in every 24-hour cycle; 8 petabytes of data held in a compliant lakehouse; 60 million profile inferences executed per hour. Each of these figures represents an engineering decision made under production constraints, not a laboratory result.
The 60 million inferences per hour figure is worth examining in detail. An inference, in this context, is the process of taking a customer's current feature set — their DNA vector, their active intent signals, the outputs of the enrichment layer — and producing a ranked set of opportunities with associated confidence scores and best-moment predictions. At 60 million per hour, the system is completing roughly 16,000 inferences per second. Each inference draws on a profile that contains more than 850 enrichments, updated continuously from the Deep Signal pipeline. The inference engine is not batch-processing these overnight. It is running continuously, prioritising which profiles to refresh based on the freshness of their signal inputs and the proximity of their predicted engagement windows.
The engineering architecture that makes this possible is event-driven throughout. The system does not poll data stores. It reacts to events. When a new signal arrives from the Edge SDK — a behavioural observation that changes a customer's micro-intention state — that signal flows into the Deep Signal pipeline, updates the relevant feature vectors, triggers a re-ranking of that customer's opportunity set, and, if a best-moment condition is met, initiates an activation sequence. The latency between signal arrival and activation is measured in seconds, not hours.
The 15% cost optimisation figure in the platform statistics is frequently underappreciated as a proof point. It reflects what happens when intelligence improves the efficiency of downstream operations at scale. When marketing communications are sent to customers at the right moment rather than at a scheduled time, conversion rates improve — and the number of communications required to achieve a given commercial outcome falls. When fraud detection is based on individual behavioural baselines rather than population-level rules, false positive rates decrease — and the operational cost of investigating flagged transactions falls. Scale, applied correctly, does not just improve commercial outcomes. It reduces operational waste.
We have learned lessons that are now built into the platform's default behaviour. The suppression logic — which ensures that a customer who has just converted is not immediately re-entered into the active opportunity queue — came from observed patterns in large-scale deployments where the absence of this logic produced measurable degradation in customer trust. The notification limiter — which enforces a minimum interval between best-moment triggers for the same customer — emerged from deployment data that showed campaign collision reducing engagement by a factor that more than offset the volume gain. These are not theoretical design choices. They are empirically grounded operational decisions.
The final point about scale that is rarely made directly: large-scale deployments are where the privacy architecture is most tested. At 320 million profiles, the consequences of an anonymisation failure — a de-anonymisation attack, a compliance breach, a regulatory investigation — are proportionally large. The privacy built into the architecture — differential privacy applied to all aggregations, anonymisation at the point of ingestion, identity data never persisting beyond the pipeline's privacy layer — were designed with this scale in mind. Privacy by design at full Telco scale is not the same problem as privacy by design at pilot scale. Intent has solved the harder version.