Decision Engine & Proprietary AI

Two pipelines

• Batch pre-processing, when you are onboarded and whenever your platform changes. inncivio maps your platform, isolates the meaningful parts of each page and turns them into structured page knowledge: what each element does, which terms need explaining, where guidance can appear.
• Real-time engine, on every interaction. Anonymised footprints reach the engine, which detects signals, decides the action, writes the content and chooses the placement. The result returns within the same interaction.

Five proprietary models

inncivio runs five proprietary models: clustering models that find behavioural patterns and cohorts, a Decision Model that makes the many small decisions along the process, a contextual Bandit that makes the main decision of which action to take, a positioning model that chooses where it appears, and inncivio's own language model, which interprets journeys and writes the content. All are owned, trained and hosted by inncivio on its own AWS infrastructure. External AI services are never used in production and never receive client data.

How an action is reached

1. Default for the signal. Friction leads to Unblock, a comprehension gap to Explain, eligibility to Disclose constraint, context divergence to Surface context, a catalyst to Anticipate.
2. Combinations that change it. For example, context divergence together with hesitation surfaces the context in the moment; attrition after earlier friction escalates to support with the failure attached.
3. Your rules. For example: the market is closed, so deliver next session; the user is only browsing, so do not interrupt; at most three in-the-moment actions per session.

Deterministic rules keep the last word

Code, not a model, validates evidence, removes duplicates, sets the compliance class and expiry, and enforces your policies before any action is released. The Bandit only chooses among actions that have already passed these rules.

• A signal needs at least two behaviour events, or one event plus one outside fact.
• A restricted action becomes "Suppress: blocked by policy". The original proposal stays in the trace, so your team sees the opportunity that was refused.
• Every action carries a "must not convey" list that always includes recommendation, forecast, suitability and urgency language.

Continuous learning

The Bandit learns from observed outcomes inside your deployment. Across deployments, only aggregated, anonymised learning signals refine inncivio's models. No client data crosses tenant boundaries.