MoEngage's Merlin AI Is the Right Architecture — and the Numbers Still Need a Second Opinion
Every marketing vendor is now selling you an "agent platform." MoEngage's Merlin AI is worth a closer look for one reason: it didn't rebuild the platform around the agent. It wired the agent into the platform that already existed — and that architectural choice is the whole story.
The take
Merlin AI connects agents to live MoEngage platform data through an MCP server running on Amazon Bedrock AgentCore, with Bedrock supplying the model inference. The agents don't get a new toy database; they call the same APIs the product already uses, with the same authentication and permissions. FastMCP generated the tools straight from MoEngage's existing OpenAPI specs. That's the boring-sounding decision that actually matters: the agent inherits the platform's guardrails instead of inventing its own.
Four agents, four different kinds of trust
- Embedded agents draft copy, images, segments and journeys inside the workflows marketers already use. Low stakes, high volume.
- Conversational agents answer questions and act across campaigns, analytics and flows — and delegate diagnostics to specialised sub-agents when a campaign underperforms.
- Custom agents run continuously on team-defined rules, with human review optional. This is where the trust question gets real.
- Decisioning agents pick content, timing, frequency and channel for individual customers. The highest-leverage mode, and the one where a bad call is invisible until the metrics arrive.
Notice what's not delegated: the model never sees or changes workspace identity. AgentCore validates the JWT, context travels in headers, sessions are isolated, and Bedrock Guardrails filter responses. Access control sits outside the model — which is exactly where it belongs.
The numbers are impressive. They're also vendor numbers.
Glance reportedly went live 50% faster, saw a 38% uplift in push click-through rates, and produced about 3x the content variants per campaign. Platform-wide, MoEngage claims 60–70% less manual effort on channel, timing and frequency decisions, and 20–30% better conversion from goal-based decisioning. Infrastructure costs allegedly dropped over 50% versus self-hosted models.
Here's the catch, and the original article is honest about it: all of it comes from an AWS account co-written with MoEngage contributors. No sample size. No measurement period. No baseline conversion rates. No controlled comparison. A 38% uplift is not 38 percentage points, and "50% faster go-live" tells you nothing about whether the saving was hours or days. These are the outcomes the companies report — not findings an independent evaluation would stand behind.
That doesn't make them false. It makes them unpriced. Treat them as a vendor's best case until someone else reproduces them.
Why this one still matters
Strip the marketing and the pattern here is genuinely good: reuse the APIs you already trust, expose them through a standard protocol, keep identity and permissions outside the model, and let humans review the consequential moves. The flash-sale demo — ask for a push notification in the tone of a winning December campaign, get variants back with the segment and performance data attached — is the workflow every marketing team actually wants. The question is whether the decisioning agents keep earning that trust at scale, one customer's inbox at a time.
Go deeper
Read the full analysis — including the architecture breakdown, the Glance case study with its missing context, and what the safeguards do and don't prove — on Agentic Affairs:
MoEngage's Merlin AI connects customer engagement agents to live platform data
Originally published on Agentic Affairs (https://agentic-affairs.com)
Top comments (0)