{"id":223464,"date":"2024-09-06T13:46:59","date_gmt":"2024-09-06T05:46:59","guid":{"rendered":"https:\/\/www.grab.com\/sg\/?post_type=editorial&p=223464"},"modified":"2026-01-08T11:36:44","modified_gmt":"2026-01-08T03:36:44","slug":"tech-behind-in-house-copywriting-tool-ai-llms-engineering","status":"publish","type":"editorial","link":"https:\/\/www.grab.com\/sg\/inside-grab\/stories\/tech-behind-in-house-copywriting-tool-ai-llms-engineering\/","title":{"rendered":"The AI technology that gives Grab\u2019s copywriting tool Mystique its voice"},"content":{"rendered":"\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t

When Grab\u2019s creative teams craft messages for our users and partners they can rely on Mystique, our powerful, in-house copywriting tool.<\/p>

Mystique spits out highly personalised content for different channels\u2014including in-app messages, push notifications, and emails\u2014in different languages, for different scenarios, all in the Grab voice.<\/p>

(Also read: Meet Mystique, the AI-powered copywriting tool that helps Grab communicate effectively<\/a>)<\/strong><\/p>

Mystique\u2019s ability to generate personalised content at scale is driven by a sophisticated blend of AI technologies. These include Machine Learning models (ML) which help us predict the user segment an individual belongs to; Large Language Models (LLMs)\u2014these are good at generating creative copy; and Retrieval-Augmented Generation (RAG)\u2014a technique for enhancing the accuracy and reliability of Mystique\u2019s output.<\/p>

Here\u2019s a closer look at the technical foundation that makes Mystique a gamechanger for content creation at Grab.<\/p>

Understanding user preferences<\/h5>

The first step required us to understand and segment the different types of user groups on our platform, along with their behaviours, preferences, and engagement patterns.<\/p>

Grab\u2019s Marketing Analytics team meticulously developed distinct personas and lifecycles based on our proprietary data, gathered from the activities of millions of Grab users and partners across the region.<\/p>

The Marketing Analytics team developed a comprehensive feature bank, which allows us to apply machine learning models to classify and predict which user group a specific individual is most likely to fall into.<\/p>

With this methodology, we are able to make data-driven assumptions about the needs and desires of each individual, based on the behaviours of the group, or persona, they most closely resemble\u2014possibly a student looking for affordable meal options, or a parent seeking kid-friendly deals, or an office worker needing a quick lunch solution.<\/p>

We also understand which part of the lifecycle the individual is in. Are they new to Grab, or a long-term loyal user? Have they interacted with a certain new feature before, or never?<\/p>

Mystique uses the rich insights embedded in our user persona and lifecycle data to create messages that are highly relevant to each segment.<\/p>

How Mystique leverages user lifecycles and personas<\/h5>

Let\u2019s demonstrate how that works with a simple example by looking at Grab\u2019s communications for Saver Delivery<\/a>. (This is an option for users who don\u2019t mind waiting longer for their food delivery, if it costs less.)<\/p>

Mystique tailors messages based on the consumer’s lifecycle stage, or previous engagement with Saver Delivery, creating distinct messages for those who have used Saver Delivery before, those who haven’t, and those who have stopped.\u00a0<\/p>

Nudging users towards an action, like trying out Saver Delivery if they\u2019ve never done it before, becomes easier if our messages are on point and meet them exactly where they are in their engagement with the feature.<\/p>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t
Mystique tailors messages based on different user lifecycles<\/figcaption>\n\t\t\t\t\t\t\t\t\t\t<\/figure>\n\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t

Mystique also leverages the user personas predicted from our machine learning models to tailor content to the context that is the most fitting for them.<\/span><\/p>

So, for instance, messages would be different depending on whether they\u2019re likely a student, making an order as part of a family household with kids, or an office worker.<\/span><\/p>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t
Mystique tailors messages based on different user personas.<\/figcaption>\n\t\t\t\t\t\t\t\t\t\t<\/figure>\n\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t

Going even further, Mystique also uses attributes from the user\u2019s transaction history to make it hyper-personalised by referencing previous interactions and past orders.<\/span><\/p>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t
Mystique tailors messages based on different user attributes.<\/figcaption>\n\t\t\t\t\t\t\t\t\t\t<\/figure>\n\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t

These examples illustrate how Mystique speaks directly to each user group, making the content more engaging and relatable.<\/p>\n

Our experimentation results showed a 25-50 per cent increase in engagement rate across our communication channels when using Mystique, compared to content created without the help of the AI copywriting tool.<\/p>\n

Architectural flow and RAG implementation<\/h5>\n

Now, let\u2019s delve even deeper into what goes on under Mystique\u2019s hood. <\/p>\n

The big thing to take away here is that Mystique isn\u2019t simply churning out content based on prompts and inputs, but that it\u2019s an integrated system with a built-in automated validation and feedback loop to improve its effectiveness.<\/p>\n

The flow can be explained in six steps:<\/p>\n

Step 1: The creative teams at Grab define parameters on Mystique, like target audience, content type, and channel, guiding the LLM to generate relevant content. This process is similar to the creative brief form our teams were already working with prior to Mystique.<\/p>\n

Step 2: Mystique uses Retrieval-Augmented Generation (RAG) to access Grab\u2019s databases, retrieving data that ensures content aligns with the user\u2019s persona and lifecycle, and references of content styles that we have previously validated to work well in terms of engagement.<\/p>\n

Step 3: The LLM then generates content using the retrieved data, enriched with contextual information from the inputs, current data and historical metrics.<\/p>\n

Step 4: Next comes the validation and feedback loop in which generated content is validated against brand guidelines, with feedback from automated systems and human reviewers, which helps us refine input prompts and the RAG process.<\/p>\n

Step 5: The generated and validated content is now ready to be deployed. Mystique can generate content for at least:<\/p>\n

    \n
  • 11 channels, including push notifications, emails, and messaging apps<\/li>\n
  • 12 user personas, including students and families<\/li>\n
  • 9 user lifecycles, including those new to Grab and long-term users<\/li>\n
  • 15 languages, including English, Bahasa Melayu, Bahasa Indonesia<\/li>\n<\/ul>\n

    This represents up to <\/b><\/p>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

    \n\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t<\/span>\n\t\t\t\t0<\/span>\n\t\t\t\t<\/span>\n\t\t\t<\/div>\n\t\t\t\t\t\t\t
    variations<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t

    Step 6: We track content performance using metrics like clickthrough rates and open rates. These insights are then fed back into the database where we can retrieve them to enhance future content.<\/p>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t

    This entire flow, as depicted in the diagram, demonstrates how Mystique works with advanced AI techniques to streamline content creation.<\/p>

    Scalability and Future-Proofing<\/h5>

    Mystique\u2019s architecture is built to scale, ensuring that it can handle the growing demands of Grab\u2019s content needs. The system is designed with modular components, allowing for easy integration of new features, such as the planned expansion into visual content generation.<\/p>

    Mystique\u2019s infrastructure is cloud-based and can scale horizontally, meaning we can easily increase the load as more Grab teams and regions adopt the tool.<\/p>

    This flexibility also makes it easier to roll out updates and improvements, ensuring that Mystique remains at the cutting edge of AI-driven content creation for everyone that uses it.<\/p>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t

    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t<\/div>\n\t\t","protected":false},"parent":180237,"menu_order":0,"template":"grab21-default","acf":[],"_links":{"self":[{"href":"https:\/\/www.grab.com\/sg\/wp-json\/wp\/v2\/editorial\/223464"}],"collection":[{"href":"https:\/\/www.grab.com\/sg\/wp-json\/wp\/v2\/editorial"}],"about":[{"href":"https:\/\/www.grab.com\/sg\/wp-json\/wp\/v2\/types\/editorial"}],"version-history":[{"count":33,"href":"https:\/\/www.grab.com\/sg\/wp-json\/wp\/v2\/editorial\/223464\/revisions"}],"predecessor-version":[{"id":256582,"href":"https:\/\/www.grab.com\/sg\/wp-json\/wp\/v2\/editorial\/223464\/revisions\/256582"}],"up":[{"embeddable":true,"href":"https:\/\/www.grab.com\/sg\/wp-json\/wp\/v2\/editorial\/180237"}],"wp:attachment":[{"href":"https:\/\/www.grab.com\/sg\/wp-json\/wp\/v2\/media?parent=223464"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}