Hightouch

Agents take authenticated action in systems Hightouch does not own: syncing audiences into destinations, activating campaigns through the customer's existing ESPs with Braze, Iterable and Salesforce Marketing Cloud named, and writing campaign identifiers back so outcomes reconcile with the customer's own analytics.

The pricing page states more than 300 destination integrations spanning advertising, email, SMS, push and web, sources including Snowflake, Databricks and BigQuery, dbt for modeling, and DAMs, product catalogs and Figma read by the studios. The MCP server and Personalization API point the other way, as routes into Hightouch rather than out of it.

Sourcehightouch.com/pricing and hightouch.com/docs/lifecycle-studio/overviewread 2026-09-01

Handles the full loop of audience building, journey orchestration, cross channel launch, measurement, and feeding learnings back into the next decision across advertising, email, SMS, and web.

Sourcehightouch.com/docs/ai-decisioning/overviewread 2026-10-01

Grounding runs on a maintained structure rather than per-request assembly, and the vendor is specific about what feeds it. Agents connect directly to the customer's data warehouse for campaign and customer history without copying it out, and Context Hub holds business context, brand guidelines and guardrails applied to every campaign the agent runs.

Lifecycle Studio additionally reads connected DAMs, the product catalog, Figma files, legal policies and past ESP templates, and AI Decisioning is prepared with event models reflecting the outcomes to optimize and engineered per-user features. Selecting products by title or SKU makes generated emails use real catalog imagery rather than invented products, which is grounding doing visible work.

Sourcehightouch.com/docs/lifecycle-studio/overview, hightouch.com/docs/ai-decisioning/prepare-data and hightouch.com/docsread 2026-09-01

A hard approval gate is documented in the vendor's own words: agents never create or change anything without approval, and during chats an agent may draft audiences or reports but users must review and explicitly save those outputs. Around it sit several shipped mechanisms. Approval flows are a named purchasable feature on the pricing tiers.

In AI Decisioning the agent operates entirely within the audience, goals, approved messages, scheduling rules and guardrails the marketer configures, with Smart Suppression able to withhold sends, and the documentation states agents do not create new content or act outside that configuration. In Lifecycle Studio marketers review and refine every generated message and approve it before export, with campaigns sent for final approval before agents activate. Every agent query is additionally scoped to the data the current user is authorized to see, so agents never bypass per-user access controls.

Sourcehightouch.com/docs/agents/overview, hightouch.com/docs/lifecycle-studio/overview and hightouch.com/pricingread 2026-09-01

The security documentation states that Hightouch maintains SOC 2 Type 2 compliance and is ISO 27001 certified by Schellman, a named accredited certification body, supports customer compliance with HIPAA, GDPR and CCPA, and runs regular third party penetration testing with reports available on request.

SOC 3 is offered alongside SOC 2, and a custom security review covering BAA, DPA and the full SOC 2 report is a named purchasable item.

Controls: role-based access control and label-based access control, SAML SSO with Okta, Azure AD and OneLogin named plus Google and Azure SSO, audit logs, approval flows and Git version control, with data encrypted at rest and TLS in transit and the metadata database under AES-256.

Sourcehightouch.com/docs/security/overview and hightouch.com/pricingread 2026-09-01

Three distinct records rather than one dashboard. Platform: audit logs are a named feature on the enterprise tier, alongside Git version control over resource changes. Agent operation: agent alerting proactively notifies on run failures, stalled agents, slow runs and drops in recommendation volume, and image edit history is browsable in Lifecycle Studio.

Decision visibility: the agent Overview tab shows live performance and experiment results including how messages were distributed, which goals are performing, and treatment against holdout, with statistical confidence levels appearing once enough event data is collected, and AI Insights breaks that down by message, goal impact and timing. Read together a customer can see what the agent chose, whether it worked and with what confidence.

Sourcehightouch.com/docs/ai-decisioning/agents, hightouch.com/docs/changelog and hightouch.com/pricingread 2026-09-01

Agents read real cross-session state: campaign history, per-user response history and engagement events read at every decision, group assignment held deterministically so a user keeps their treatment or holdout allocation across runs, and Context Hub carrying business context, brand guidelines and guardrails applied to every campaign.

No memory feature, scope or lifetime is documented, and the state lives in the customer's own warehouse, so nothing could be deleted for one brand separately from that brand's commerce data. The learning the vendor describes is model update, its engineering blog stating the reinforcement learning model learns from results and the model updates.

Sourcehightouch.com/docs/ai-decisioning/agents and hightouch.com/blog/rl-agents-for-marketingread 2026-09-01

Two independent forms of customer control over where data sits. Region: the pricing feature list names deployment in the EU region and deployment in multiple regions as purchasable capabilities, alongside SSH tunneling for private connectivity.

Architecture: the platform is warehouse-native and does not hold a duplicate profile store, so activation data stays in the customer's own warehouse, and the security documentation states that where AI and real-time features must cache feature data, outcomes or logs, those artifacts are single-tenant and can be stored in the customer's own cloud bucket. That is region selection plus data resident in customer-controlled infrastructure. The control plane remains cloud SaaS, with no self-hosted or on-premises option.

Sourcehightouch.com/docs/security/overview and hightouch.com/pricingread 2026-09-01

A complete template system that ships empty. Customer Studio carries three template types, audience, sync and journey, each with its own browsable list in the product, previewable before use with size calculation and destination preview, applied to create a new object from a configured starting point, with syncs staying linked to their template so locked-field changes propagate, and a workspace setting that can require sync templates for every audience activation.

The library is stocked by the buyer, not the vendor; the documentation states platform admins typically create and govern templates while marketers apply them, and the named jobs it gives, a churn-risk audience, a standard Braze sync, a new-customer welcome series, and the cart abandonment and welcome series journeys, are examples of what a customer would build. Suggested reports are generated output, and the email-component and image libraries hold the customer's imported material.

Sourcehightouch.com/docs/customer-studio/templates and hightouch.com/docs/customer-studio/journeysread 2026-09-02

Always on agents proactively find opportunities twenty four seven and execute across advertising, email, SMS, push, in app, and web channels.

Sourcehightouch.com/docs/ai-decisioning/overviewread 2026-10-01

The customer does not choose the model and no provider is named. AI Decisioning is the vendor's own reinforcement learning system, described in its documentation as ingesting audience, goals, creative and guardrails and using performance data to pick message, channel and timing, with the vendor's engineering blog detailing contextual multi-armed bandits and an LLM used only to generate semantic tags over message content.

No model selector, per-agent model choice, bring-your-own-key or routing policy appears in the documentation, the pricing feature list or the product pages, and no third-party provider is identified anywhere; the security overview refers to third-party models only obliquely, stating artifacts are never used to train shared or third-party models.

The vendor's line about a warehouse plus an AI of your choice describes the customer's wider stack rather than model selection inside the product, and the MCP server lets a customer's own assistant call Hightouch, which is not the same as choosing the model inside it.

Sourcehightouch.com/docs/ai-decisioning/overview and hightouch.com/blog/rl-agents-for-marketingread 2026-09-01

The platform is callable from outside through several documented surfaces.

A hosted MCP server connects a customer's own AI assistant, with Claude and ChatGPT named, and exposes real capability rather than a data pipe: understanding the customer schema, building audiences, creating and analyzing journeys, managing syncs and creating syncs from templates for all customers, building and iterating on email and SMS content for Lifecycle Studio customers, and generating and refining ad creatives for Ad Studio customers.

Alongside it sit a Personalization API, Hightouch Events for inbound event collection, and a Resource API named on the enterprise pricing tier. The destination catalog, by contrast, points outward to other systems.

Sourcehightouch.com/docs/ai-integrations/mcp and hightouch.com/pricingread 2026-09-01

A controlled experiment on the agent, with a criterion and a published result.

Holdout groups are configured per agent, users are assigned to treatment, holdout or customer-managed by a deterministic uniform hash of a stable identifier plus a static experiment key so allocation stays stable and uncontaminated across runs, incrementality is measured against that counterfactual to establish whether personalized messaging drove conversions beyond what would have happened anyway, and statistical confidence levels are surfaced once enough event data has been collected.

AI Insights then breaks performance down by message, goal impact and timing. That answers directly whether the agent got better or worse after a change. The counterfactual is about the agent itself: the holdout withholds the agent's treatment from a deterministically assigned group, rather than isolating the performance of the content or the media spend the agent influences.

Sourcehightouch.com/docs/ai-decisioning/group-assignment and hightouch.com/docs/ai-decisioning/agentsread 2026-09-01

Executes through warehouse queries and destination integrations rather than driving a browser or operating a computer interface.

Sourcehightouch.com/docs/ai-decisioning/overviewread 2026-10-01