AI Agents, Standard

MoEngage’s Merlin AI connects customer engagement agents to live platform data

Symbolic diagram: MoEngage’s Merlin AI connects customer engagement agents to live platform data

Introduction to MoEngage and its AI-driven customer engagement platform.

MoEngage has built Merlin AI, a collection of agents for AI-driven customer engagement that use live platform data to help marketing and product teams create campaigns, investigate performance, and make decisions for individual customers. Its approach connects conversational interaction to existing platform functions: a request can lead to analysis, content generation, or a change to a customer journey.

The reported results are substantial. Glance, a MoEngage customer, reached campaign go-live 50% faster and recorded a 38% uplift in push notification click-through rates. Across its platform, MoEngage reports a 60–70% reduction in manual effort for decisions about channel, timing, and frequency. These figures come from an AWS technical account co-written with MoEngage contributors, rather than an independent evaluation. They show the outcomes the companies report, with important limits on what can be inferred about causation and repeatability.

The central development is architectural. MoEngage connected agents to its existing APIs instead of rebuilding the platform around them. Amazon Bedrock supplies model inference, Amazon Bedrock AgentCore runtime hosts the Model Context Protocol (MCP) server, and MCP gives agents a standardized way to access platform capabilities. Existing authentication and permissions remain part of that path.

According to the account, more than 1,350 consumer brands use MoEngage across channels including push, email, in-app messages, SMS, WhatsApp, and Rich Communication Services. That multichannel scope helps explain the problem Merlin AI addresses: planning and operating a campaign can involve analytics, audience selection, content, and journey management. Bringing those functions into a conversation could reduce the effort spent moving between tools. Whether it improves customer outcomes depends on a separate question: how well the resulting campaigns and decisions perform.

Overview of Merlin AI and its functionalities.

Symbolic diagram illustrating: Overview of Merlin AI and its functionalities.

Merlin AI has four agent modes, covering both work performed for marketing teams and decisions made for individual customers. This distinction matters because drafting a message, diagnosing a campaign, and continuously selecting an offer involve different responsibilities and different levels of delegated control.

Agent modeDocumented roleUser control or context
Embedded agentsHelp create copy, images, segments, templates, and customer journeysOperate within relevant product workflows
Conversational agentsAnswer questions and act across campaigns, analytics, segments, and flowsUse plain-language requests to select tools
Custom agentsRun continuously using MoEngage data and toolsFollow team-defined rules, with autonomous or human-reviewed execution
Decisioning agentsSelect content, timing, frequency, and channel for individual customersWork toward goals within team-defined guardrails

Within Merlin AI Studio, Copywriter generates personalized push and transactional messages, while Designer produces marketing images from text descriptions. Segment Assist translates audience descriptions into attributes and filters. Jinja Assist generates Jinja and Liquid markup for dynamic content. In-app Template Builder produces HTML templates for in-app messaging, including surveys and lead-capture formats.

Flow Assist handles a broader task: translating a high-level request into a customer journey. The documented process interprets the request, proposes paths, constructs the flow, refines it using historical data, and prepares it for deployment. This extends the scope beyond generating a single piece of content to assembling a sequence of customer interactions.

Conversational agents connect these functions with analysis. The Campaign Insights Agent delegates work to specialized sub-agents that examine multiple metrics, investigate underperformance, and compare periods or segments. A subsequent step formats their findings. The account describes this division of labor but does not provide an evaluation of diagnostic accuracy.

Custom agents add continuous operation and an activity log covering data accessed, decisions, channels, and content sent. Teams can require review before execution or permit autonomy. Decisioning agents operate at the customer level: Offer Decisioning selects offers on web and app surfaces, while Campaign Decisioning handles individualized choices across email, push, SMS, and WhatsApp.

How the agents reach platform data

The technical foundation separates the agent doing the reasoning from the server exposing tools. Internal Merlin AI clients and external MCP-compatible clients connect to MoEngage’s MCP server. The server runs on AgentCore runtime and calls existing MoEngage APIs; it does not run its own model. Agents use foundation models in Amazon Bedrock to plan tool calls, generate content, and summarize results.

MoEngage used FastMCP to generate tools from its existing OpenAPI 3.0-compatible specifications. Those tools cover campaigns, analytics, segmentation, content, and flows. Each operation has a schema, description, and parameter validation. Reusing these specifications provides a route from established API functions to agent-accessible operations, while keeping the underlying platform APIs responsible for permissions and data residency rules.

Identity is handled outside the model’s tool arguments. AgentCore runtime validates the user’s JSON Web Token, and workspace and user context pass through request headers. The account says the model cannot see or change those context values. Individual sessions run in isolated environments. Amazon Bedrock Guardrails apply content safety and personally identifiable information filters to responses, and a backend layer validates generated content structure before the interface displays it.

These are documented safeguards, not proof that every possible failure is excluded. Their architectural significance is that access boundaries rely on authentication, request context, and session isolation rather than on a model choosing the correct tenant.

Impact of AI agents on marketing efficiency.

Symbolic diagram illustrating: Impact of AI agents on marketing efficiency.

Merlin AI targets two related sources of work: assembling campaigns across platform functions and repeatedly deciding how to contact customers. The first involves moving between analytics, segments, content, and flows. The second involves channel, timing, frequency, and offer choices. MoEngage’s reported efficiency gains should be read against those specific tasks.

The source illustrates the workflow with a weekend flash sale. A marketer asks for a push notification aimed at a high-value segment, using the tone of a successful December campaign. The agent retrieves campaign performance, checks the segment, accesses brand content, and returns copy variants. This is an example supplied by the companies, rather than a documented Glance campaign.

A follow-up request can ask which push notifications fell below a 2% click-through rate and what they share. Another asks the agent to inspect an onboarding flow and pause an SMS branch if conversion is below 5%. In that example, the agent pauses the branch after user confirmation. The percentages are conditions in an illustrative workflow, not performance findings or recommended targets.

The efficiency argument is that a conversation can coordinate several existing operations. The user still supplies the objective and reviews consequential actions when review mode applies, but the agent handles the intermediate tool calls. That gives the reported reduction in manual effort a concrete mechanism: fewer separate steps for gathering context and carrying out a decision.

MoEngage reports a 60–70% reduction in manual effort specifically for channel, timing, and frequency decisions, alongside a 20–30% improvement in conversion rates from goal-based decisioning. The supplied account does not disclose the sample size, measurement period, baseline conversion rates, or a controlled comparison. Those omissions prevent treating either range as an expected result for every customer or every marketing task.

There is a separate infrastructure claim. Managed, pay-per-token inference through Amazon Bedrock reportedly reduced AI infrastructure costs by more than 50% compared with self-hosted models. That comparison concerns AI infrastructure, not the full cost of running a marketing organization or implementing Merlin AI. The account does not supply a detailed cost breakdown or matched workload measurements.

Operational control remains relevant even when intermediate work becomes faster. Custom agents expose activity logs, and teams can choose human-reviewed execution. AgentCore runtime supplies session metrics, logs, and traces through Amazon CloudWatch. These features provide documented ways to inspect activity; the evidence does not establish how often human intervention is needed or how consistently generated campaigns meet a brand’s requirements.

Case study: Glance’s success with MoEngage.

Glance provides the most concrete customer example in the supplied evidence. The AWS and MoEngage account reports three outcomes: campaign go-live was 50% faster, push notification click-through rates rose by 38%, and teams using Merlin AI Copywriter produced about three times as many content variants per campaign.

Those results concern different parts of the campaign process. Faster go-live describes execution speed. More variants describe content production. A higher click-through rate describes audience response. Taken together, they suggest that the reported benefits extended from preparation into engagement, although the account does not establish how much each capability contributed to the final result.

Reported Glance outcomeWhat it measuresImportant missing context
Campaign go-live 50% fasterCampaign execution speedAbsolute time saved and the baseline process
38% uplift in push notification click-through ratesEngagement with push notificationsStarting rate, campaign mix, and measurement period
About three times more content variants per campaignCopy production volumeReview workload and the performance of individual variants

The 38% uplift should not be restated as a 38-percentage-point increase. The source provides no underlying click-through rates from which to calculate an absolute change. Similarly, faster campaign delivery does not reveal whether the time saved amounted to hours or days. Preserving the reported wording avoids implying precision the evidence does not contain.

More content variants also require careful interpretation. The figure supports a claim about production capacity, but the account does not explain how variants were selected, tested, or reviewed. It would therefore go beyond the evidence to attribute Glance’s entire engagement improvement to generating more copy.

The source reports that Merlin AI agents on this architecture had been in production for more than six months. That establishes reported production use, rather than only a demonstration. It does not establish that Glance’s metrics were measured over that full period. The supplied material provides no detailed study design, control group, or independent audit of the outcomes.

For teams assessing a similar deployment, Glance’s case suggests three distinct questions to measure: whether campaigns reach launch sooner, whether content production becomes more efficient, and whether customers respond more often. That is an editorial implication of the reported results, not an additional claim about Glance. Keeping those measures separate would help distinguish workflow gains from improvements in campaign effectiveness.

Conclusion and future implications for AI in customer engagement.

MoEngage’s implementation shows a practical route for adding agents to an established customer engagement platform: expose existing API operations as MCP tools, retain authentication and tenant context outside model control, and use a managed runtime for hosting and session isolation. The source describes a platform made accessible to agents through its existing capabilities, rather than a replacement for those capabilities.

Its broader implication is conditional. Software providers with mature API specifications may be able to reuse a similar pattern, as the AWS account argues. That does not establish equivalent results elsewhere. The quality of the available operations, the access controls around them, and the ability to inspect actions remain central to what an agent can safely and usefully accomplish. AgentCore feature availability also varies by AWS Region, a deployment constraint explicitly noted in the source.

For marketing teams, the evidence supports a promising combination of conversational workflow assistance and individualized decisioning. It also leaves key questions unanswered: how reproducible the reported gains are across customers, what baselines underpin them, and how review requirements affect the net time saved. MoEngage’s reported results make those questions worth measuring. The next useful evidence would connect the documented architecture and controls to transparent, comparable customer outcomes.