Skip to main content

​
Integrations, Data, and Information Security

Aampe works with your event and attribute data. Agents don’t need a perfectly unified table—just connect your existing data streams (e.g., analytics, CDP, app events), and the agents will organize them for learning.
Yes. Aampe can ingest blog posts, customer reviews, or recommendations and repurpose them into messages. This expands the types of content agents can deliver while enriching personalization.
No. Aampe uses hashed user IDs for targeting. No personally identifiable information is required.
Each customer’s data is stored in its own secure tenancy, co-located at Aampe’s GCP BigQuery in the customer’s chosen region. Data never leaves your tenancy.
Aampe can work without history, but providing 2–3 months of event data helps warm-start agents. Historical data is mainly used for analysis and priors; agents still learn best from live interactions.

​
Content Creation

Components are the building blocks of a message (e.g., Greeting, Value Proposition, Offering, Incentive, Call to Action, Tone). Each can have multiple alternates for agents to mix and match.
Labels are semantic tags (e.g., Convenience, Trust, Urgency) applied to alternates. They help agents learn which strategies work for different users and contexts.
No. You need coverage, not volume. Start with 10–15 message groups across your key use cases, each with 3–5 alternates and labels. This gives agents enough variation to learn effectively.
Aampe works across push notifications, email, SMS, in-app screens, and web. Each surface has its own formatting rules, but agents can learn across all of them.
Yes. Alternates can be adapted across surfaces (e.g., a blog post reformatted as a push notification). Agents optimize for context, so content can be reused flexibly.

​
Tools for Working with Aampe

It’s a design tool that helps you prototype and visualize how your Aampe messages will look on different surfaces (push, email, in-app, etc.) before going live.
Yes. You can create component-based templates in Figma and sync them into Aampe for testing, making the design-to-deployment workflow faster.
AamPT is your assistant for generating modular content and labels. It helps expand your library, suggest new alternates, and create message structures in Aampe format.
Use AamPT when you need to expand message groups, generate new topics, or quickly draft alternates with consistent labeling.

​
How Does Aampe Work?

Each user receives their own dedicated Agent. Agents share information, especially for new or low-data users, borrowing from “nearest neighbors.” They also respect global limits like budgets and messaging frequency.
Yes. While each Agent makes decisions for one user, they share aggregated learnings to improve performance across the system.
Cold start depends on engagement. Highly active users can be understood in 1–2 weeks, while less active users take longer. Historical data helps warm-start but is not required.
Agents optimize content, timing, and channel—but always within guardrails you set. They only choose from approved content and respect frequency, hours of operation, and campaign rules.
Agents use Thompson Sampling to select labels (e.g., greeting style, offering, CTA) and assemble them into a final message. Variations are drawn independently, but large-scale experimentation helps discover which combinations work best.
Each user event has a reward value (e.g., conversion = 1, search = 0.2). Agents forecast expected outcomes, compare them with actual outcomes, and attribute the difference to the message. Over time, this tunes strategies to maximize long-term value.
Behavioral signals (clicks, searches, features used) typically outweigh static attributes like demographics. Recency is also weighted more heavily than older events. This ensures agents adapt to current behavior rather than historical averages.
Agents continuously update based on observed vs. expected outcomes. They re-enter exploration when rewards plateau, shifting strategies without discarding past learning.
Thompson Sampling ensures that agents try new things even when they’ve found successful strategies, preventing “local maxima” and keeping learning active.
Yes. You can assign different weights to events (e.g., prioritize conversions but guard against unsubscribes). Agents optimize for the sum of rewards across all goals.
Yes. You can set a conversion window (e.g., 7d, 30d, 90d). Shorter windows provide clearer attribution; longer windows make it harder to isolate cause and effect. Agents update rewards daily, even within long windows, so learning is ongoing.
Since each user has a dedicated Agent, competing campaigns aren’t a problem. Agents simply have more options to choose from, and randomization ensures learnings are shared across users.
Aampe provides message-level metrics in the tool and richer exports through Data Share. This includes user-level profiles, sampling details, and reward tracking.
By default, Aampe exports chosen decisions. Full counterfactuals can be exported but are usually cost-prohibitive. Data Share enables deeper offline analysis using inverse propensity scoring.
You can run holdout tests to validate performance. While personalization makes clean campaign holdouts difficult, extended holdouts or offline causal inference methods (like synthetic control) can provide validation.
Feature importance is available in Data Share at the label level (e.g., value prop, offering, tone). Because agent orchestration is emergent, per-decision explanations are hard—but aggregated insights reveal which factors drive outcomes.
Promos can be treated as context events in the agent’s data stream. This prevents them from confounding results and allows the agent to adapt strategies around promo activity.