MoEngage has built a set of AI agents that can help marketing teams draft and analyse campaigns, and make decisions about what to send individual customers. The system connects those agents to MoEngage’s existing platform tools through MCP, with controls intended to keep customer data and actions within each tenant’s permissions.
AWS AI Watch analysis
What happened
An AWS case study, co-written with MoEngage staff, describes Merlin AI running on Amazon Bedrock and Bedrock AgentCore. It outlines four types of agents: assistants embedded in marketing workflows; conversational agents that work across campaigns and analytics; custom agents that teams can run continuously; and decisioning agents that choose messages, timing and channels for individual customers.
MoEngage says it generated MCP tools from its existing API specifications, rather than rebuilding those APIs. In its described design, the model does not receive tenant and user context as tool arguments. AgentCore validates the user’s token, while existing APIs enforce permissions and data-residency rules. Teams can also require human review before custom agents act, and inspect an activity log of their work.
The case study says one customer, Glance, saw a 38% uplift in push-notification click-through rates. That is a customer result reported by MoEngage, not an independent evaluation.
Why it matters
This is a concrete example of AI agents being connected to the machinery of a marketing platform, not just asked to draft copy in a chat window. They can work with campaign history, analytics, audience segments and customer journeys, and some can take actions or tailor decisions for individual customers.
That makes the permission boundaries and review options part of the product story. MoEngage says its design reuses existing access controls and lets teams choose between autonomous and human-reviewed execution. The case study describes the architecture, but does not provide an independent assessment of its security or the claimed customer uplift.
Our read
The useful idea here is not “AI writes marketing copy”; that part has become table stakes. It is the attempt to give agents access to existing tools without letting the model decide whose data it can reach. If that separation works as described, it is a more consequential design choice than adding another chat box. The proof will be in what the agents do in production, and how clearly teams can inspect and reverse their actions.
What to watch
- Whether MoEngage publishes independent or more detailed results for customer outcomes such as Glance’s reported click-through uplift.
- How teams set permissions and choose between autonomous and human-reviewed actions.
- Whether the MCP connector and agent tools remain aligned as MoEngage updates its underlying APIs.
Discussion spark: Should marketing agents be allowed to make customer-by-customer decisions autonomously when teams can inspect their activity, or should human approval remain the default?
Sources and evidence
- How MoEngage built agentic customer engagement on Amazon Bedrock AgentCore (16 November 2026, 18:16 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)