8ndpoint Repurchase: AI-Powered Precision Marketing for E-commerce | MoBagel

In 2021, MoBagel applied its core technology, Decanter AI, to the marketing vertical field, launching a new product line called 8ndpoint. It covers all stages of the marketing funnel, creating personalized consumer experiences through data-driven methods such as ad monitoring, member management, and market planning, ultimately leading to high conversion rates. Data shows that without good member management, e-commerce businesses can easily lose over 70% of their members. Repurchase is an AI product designed for the member management stage, aiming to allow customers to easily apply AI capabilities for effective member management.

I joined the team as a product management intern in the second half of 2021, with the goal of helping 8ndpoint validate the value of AI in e-commerce member management/precision marketing, and driving the MVP of 8ndpoint's AI precision member marketing product (Repurchase) from 0 to 1. The 8ndpoint team had fewer than 10 members in total. During this process, the main challenges I faced included sharing development resources (front-end and back-end), and coinciding with a new round of fundraising. As Repurchase was the only 8ndpoint product that had not yet been launched, it needed to iterate quickly to validate in the market and secure funding through growth potential.

Product Planning

The core value proposition of Repurchase is that AI can produce more precise marketing distribution lists than traditional RFM models. After proposing this idea to our well-known brand e-commerce partners, we launched several rounds of experiments, comparing the performance after adjusting various parameters for RFM. Due to the confidentiality of the collaboration, I can't share many experimental details, but ultimately, this partner increased customer repurchase rates by nearly 10% and achieved NTD 1.3 million in revenue, successfully validating our hypothesis and product value.

After validating the product value, the goal was to create an MVP within one month. Based on the data analysis process, I defined several basic product modules:

• Data upload
• Basic BI reports showing repurchase situations
• Marketing goals and list settings
• AI analysis
• Export lists

During this period, I would typically use Wireframes and Flowcharts to discuss product specifications with the product manager and data scientists. For example, I needed to discuss with data scientists the data required for specific business scenarios, data formats, and model performance evaluation methods. As the model performance evaluation methods were usually too technical, I needed to translate them into ways that general users could understand and present them in the interface.

After clarifying the basic product specifications, I began the product design process, continuously adjusting the design and specifications through internal feedback and ongoing interviews with partners. Below, I will introduce the design and thought process of the final launched product.

Product Design

Completing an AI prediction involves three main stages: data input, data analysis, and presentation of analysis results. Each stage has its own core objectives to achieve. Below, I will analyze the product design approach by stage.

Customer Persona

Before presenting our design, let me briefly introduce the product's customer persona, as all designs are centered around this customer persona.

8ndpoint's initial target group is mainly SMBs (small and medium-sized businesses) for two main reasons: (1) Our deep collaboration with iProspect allowed their primary client base - SMB customers - to switch to 8ndpoint's ad monitoring product and deliver successful case studies, making it relatively easy to upsell these customers to new services; (2) These customers have digital transformation needs but typically haven't yet formed comprehensive data analysis practices and may not have dedicated data analysts. Additionally, due to their less complex organizational structures compared to large enterprises and shorter decision-making processes, introducing user-friendly AI tools is relatively attractive to SMBs.

For the precision marketing product for members, to ensure rapid validation, we chose "online proprietary e-commerce" as the initial target industry to facilitate data source acquisition and align with our partners' data types. This customer group typically has a membership base of around 10,000, with marketing activities planning and related analyses usually conducted by in-house marketing specialists.

Data Upload

Initially, I didn't design a feature for automated import from databases or other data sources. Instead, I chose to let users directly upload data in our specified format, while simultaneously verifying if the data format meets the requirements during the process. The advantage of this design for users is that it eliminates difficulties in data retrieval and integration. For us, the benefit is ensuring data quality and controlling the development costs at this stage.

However, having users upload data themselves essentially increases the behavioural cost, potentially leading to a lower product activation rate. To avoid this issue, I introduced a gamification concept where each time a user uploads a piece of data, they unlock an initial analysis performed by Repurchase using that data, presented through a dashboard. This design aims to encourage users to continue uploading data to receive the latest analyses.

After uploading data, users can update it as needed, but I intentionally did not include a data deletion button in the interface. From the user's perspective, the accuracy and value of the analysis improve with larger amounts of data, so there's no need for deletion. Even if outdated files are accidentally uploaded, we can identify and remove duplicate data directly. From the product's perspective, this allows customer data to accumulate within the product, helping to increase customer retention.

Although we need a total of three separate pieces of data to perform all types of predictions, we allow users to start using our list recommendation feature by uploading just the "member transaction data". For prediction categories that require more types of data, the interface will remind users to return to the homepage to upload additional information.

Data Analysis

In member management, common business scenarios include "remarketing," "specific product promotion," and "member segmentation." We have abstracted these scenarios, allowing users to obtain recommended lists through simple clicks without having to consider which parameters or dimensions to set.

For example, if a user wants to stimulate existing members to repurchase by sending remarketing messages, they only need to select "Stimulate Member Repurchase" and set the time period for sending marketing messages to get our recommended distribution list. For users who want to promote specific products, they just need to choose the product category they want to promote.

During the data analysis process, we aim to ensure the quality of the analysis results, so we require users to provide the most recent data possible. I achieved this by deliberately limiting the selectable date range. When the date of the last uploaded data is D, users can select up to D+90. The inability to select a distribution date will prompt users to update their member data.

In addition to needing the latest data, Repurchase also needs to ensure that there is sufficient data for our model to produce high-quality analysis results. If Repurchase determines that there is insufficient data, it will remind users that more data is needed to conduct the analysis.

Once these two data quality verification mechanisms are passed, users can obtain Repurchase's recommended list in a very short time. By integrating MoBagel's proprietary Decanter AI technology, most lists can be predicted within one minute, even with large amounts of data.

Presentation of Prediction Results

When the prediction is complete, Repurchase not only provides users with a recommended list for download but also allows them to set the unit cost of the marketing message for that campaign and the number of people they plan to send messages to. As these numbers change, the marketing effectiveness changes accordingly. This section can actually produce a lot of analyses, but I have simplified the information as much as possible, hoping that users can determine how many people to include in the list download based on the marketing effectiveness information.

Market Entry Strategy

After the product launch, I immediately began planning the Go-to-market Strategy with the product marketing manager, which mainly included content strategy, customer acquisition channels, and delivery materials.

We realigned Repurchase's positioning within 8ndpoint and how Repurchase differentiates itself in the MarTech market. Specifically, even though there are many similar competitors in the current market, they lack modularity and ease of use, factors that increase the cost of implementing data tools for SMBs and reduce their willingness to adopt. On the technical level, MoBagel's Decanter AI is the fastest and most accurate AutoML technology in the market, and when combined with the AI consulting team, it can achieve industry-specific application implementation. Therefore, considering cost, user experience, and technology, we are confident in meeting the needs of our target customers. Finally, we also clarified Repurchase's customer base and value proposition, presenting this content in the Sales Kit and product website.

Due to our deep collaboration with iProspect, this will be our main channel for acquiring seed customers. Additionally, we continue to collaborate with educational institutions on courses and projects. Currently, MoBagel has signed letters of intent with dozens of Taiwanese universities to introduce the 8ndpoint product line into collaborative courses, training management school and MBA/EMBA students to use 8ndpoint products, and further promoting them in the students' work environments, achieving Product-Led Growth.

Lastly, by researching other MarTech SaaS pricing and calculating the returns the product can bring, I defined Repurchase's subscription prices and charging model. This charging model is tied to the usage of AI analysis resources because if customers truly perceive the product's value, their resource usage should continue to increase. The specific prices are not currently public, but if you're interested, please contact a representative through the official website.

Data Tracking

Our main data analysis tool is Amplitude, so based on Amplitude's documentation, I compiled the necessary Events and Event Properties for the frontend and backend engineers to implement tracking. The entire tracking framework consists of three main parts: (1) North Star Metric (2) Monetization Metrics (3) User Experience Metrics, which I'll briefly introduce below.

North Star Metric

The North Star Metric for the entire 8ndpoint product line should be consistent, which is "how often users utilize predictive analysis." This corresponds to our value proposition, which is to make their work more efficient through AI predictions. Therefore, if users perceive the product's value, the North Star Metric should show positive performance.

Monetisation Metrics

Monetization metrics are linked to our Go-to-market Strategy's business model. We hope to answer questions through data, including (1) How much data have users uploaded? (2) On average, how many marketing campaigns do users execute per month? (3) What is the marketing budget for each user's campaign? (4) How many customers do users expect to reach with each marketing campaign? If we can answer these questions, it will help us examine the rationality of different tier plan divisions and whether the overall pricing model is reasonable.

User Experience Metrics

Since we don't have a very complex product structure at present, the current focus is on hoping users can fully experience the product. Therefore, we will measure (1) whether users can complete data uploads and (2) whether users can complete the prediction process.

Customer Experience Research and Product Iteration

After the product launch, we recruited product usability testers through partner companies, asking customers to try the product and provide feedback. Through three sessions of approximately 1-2 hour tests and interviews, we were able to better understand the business processes of marketers in managing repurchases, and also summarized many product requirements worth prioritizing for iteration.

Key Research Findings

1. Revenue is King

Among the customers we recruited, we found that customers don't care about cost control, but rather whether they can meet their revenue KPIs. Additionally, users don't specifically calculate the cost of each distribution channel, but choose the most effective channels based on past experience. If they don't meet targets, they then choose new channels. The operation panel for list settings and the presentation of predicted performance information may undergo significant design changes due to these two findings. For example, in the final presentation of prediction information, the focus should be on using revenue information to help users decide on the number or frequency of deployments rather than through ROI.

2. No Need to Deliberately Seek User Trust in AI

When initially designing the product, we found that investors wanted to see a lot of information demonstrating AI capabilities, such as how much more revenue Repurchase could generate compared to traditional RFM models. Although we removed much of the information we felt customers might not necessarily care about after demoing to investors, we still retained the ROI curve of the RFM model, indirectly demonstrating through the curve height that Repurchase's recommended lists would perform better.

However, in the interviews, we found that this information, intended to reinforce Repurchase's superiority over traditional models, actually caused confusion and distraction for users. We further learned that users felt this information didn't help with actual tool operation. Although users were indeed curious about how AI generates these results, they also understood that the machine learning process is essentially a black box, so the absence of this information doesn't affect their willingness to use Repurchase.

In the current version, we have already removed the ROI curve of the RFM model. In the future, we won't design features or information presentations that compare with traditional methods in our customer-facing products.

3. Users Possess Certain Data Processing Capabilities

In the data upload process, I was originally concerned that if marketers' data processing abilities were poor, it might increase the threshold for uploading data. Through these studies, we found that the data preprocessing we currently ask users to do is not at all difficult for them, and they can process it very quickly.

This finding allows us to focus more on how the analysis results in the dashboard can inspire marketers to design member engagement activities when designing future features, rather than on what information would attract users to upload data.

Other Future Product Directions

In addition to the above product requirements, considering Repurchase's role in 8ndpoint and 8ndpoint's future product strategy, besides optimizing Repurchase's product experience itself, it also includes data import at the front end and data utilization at the back end.

First, we will try to integrate and reuse data across products, using data from ad monitoring and inventory optimization to improve prediction accuracy.

Second, the current recommended lists can be further applied to other MarTech applications, allowing post-deployment performance data to feed back to the existing model as a way to calibrate the model.

Finally, Repurchase's capabilities could potentially be opened up via API, made into other apps, and listed on marketplaces similar to Shopify. MoBagel has always aimed to build an open AI platform, making AI accessible to everyone, which is one of our biggest differences from other AutoML platforms.

We look forward to 8ndpoint becoming such a product, and hope that Repurchase can play a key role in the future.

中文版：AI 會員精準行銷 - 8ndpoint Repurchase 0-1 落地

2021 年 MoBagel 將其核心技術 Decanter AI 應用在行銷垂直領域推出了全新產品線 8ndpoint，並覆蓋行銷漏斗全階段，透過廣告監測、會員經營，與市場規劃等主產品，以數據驅動的方式創造個人化消費體驗，最終促成高轉化率。數據顯示，如果缺乏良好的會員經營，電商很容易就產生 70% 以上的會員流失。Repurchase 就是針對會員經營階段而打造的 AI 產品，目標是讓客戶能夠非常簡單的就能應用 AI 能力來做到良好的會員經營。

我在 2021 下半年加入團隊擔任產品經理實習生，目標即是協助 8ndpoint 驗證 AI 在電商會員經營/精準行銷上的價值，推動 8ndpoint AI 精準會員行銷產品（Repurchase）的 MVP 從 0 到 1 落地。8ndpoint 團隊全員不到 10 人，在這個過程中我主要面臨的挑戰包含開發人力的共用（前後端），且適逢新一輪募資階段，Repurchase 作為 8ndpoint 唯一沒有落地的產品，需要快速迭代來落地到市場上驗證，並透過增長潛力爭取募資。

產品規劃

Repurchase 的核心價值主張就是 AI 能產出比傳統 RFM 模型更精準的行銷投放名單。在把這個想法提案給我們的知名品牌電商合作夥伴後，就開展了數輪的實驗，針對 RFM 調整各種參數後的表現去進行成效比較。基於合作的保密性我沒辦法分享太多實驗細節，但最終此合作夥伴提升了近 10% 的顧客回購率以及 NTD 130 萬的營業額，成功驗證我們的假設與產品價值。

在驗證了產品價值後，目標就是用一個月的時間打造出 MVP。我依據資料分析流程定義出了幾個基本的產品模塊：

• 上傳資料
• 基本 BI 報表展示回購情況
• 行銷目標與名單設定
• AI 分析
• 匯出名單

期間我通常會使用 Wireframe 與 Flowchart 與產品主管和資料科學家討論產品規格。舉例而言，我需要與資料科學家討論特定業務場景所需之資料、資料格式與模型表現評估方式，而模型表現的評估方式通常太過技術，我還需要再轉譯成一般使用者能夠理解的方式再呈現於介面中。

在明確了基本的產品規格後，我就開始進行產品設計並透過內部回饋與持續訪談合作夥伴的方式，在產品設計過程中持續調整設計與規格，以下我將介紹最終上線產品的設計與思路。

產品設計

完成一個 AI 預測會經歷三大環節：資料輸入，資料分析，呈現分析結果。每個階段都有各自圍繞的核心目標需要達成，以下分階段分析產品設計思路。

客戶畫像

在呈現我們的設計前，先簡介一下產品的客戶畫像，因為設計皆是圍繞這個客戶畫像進行設計。

8ndpoint 初期的目標族群主要是 SMB（中小企業），主要原因有二：(1) 我們與 iProspect 的深度合作中讓其主要客群 — 中小企業客戶轉用 8ndpoint 廣告監測產品，並交付成功案例，因此 upsell 這群客戶使用新服務相對容易；(2) 這群客戶擁有數位轉型需求，但通常尚未形塑完整的數據分析作法，也不一定擁有專職的數據分析人員；再加上由於組織結構相對大企業不複雜，決策流程較短，導入方便易用的 AI 工具對 SMB 而言是相對有吸引力的手段。

對於會員精準行銷產品而言，為講求快速驗證，我們選擇「線上自有電商」作為初期鎖定的產業，以便取得資料源，也符合合作夥伴的資料型態。此類客群所擁有的會員數大致在萬人規模，規劃行銷活動與進行相關分析的人員，多為企業內部之行銷專員。

上傳資料

初期我並沒有設計從資料庫或其他數據源自動化匯入的功能，而是選擇讓使用者直接上傳我們指定格式的資料，同時在過程中驗證資料格式是否符合要求。這樣的設計對使用者而言的好處是免去取數與串接上的困難，而對我們來說好處是可以確保資料的品質以及控制這個階段的開發成本。

然而相對的，讓使用者自主上傳資料實質上是增加了行為成本，可能因此導致較低的產品啟用（Activation）率。為避免此問題，我引入了遊戲化的概念，每次使用者上傳一筆資料，就能解鎖 Repurchase 使用該資料幫他們進行的初步分析，並透過儀表板來呈現，期望這樣的設計能鼓勵使用者持續上傳資料，以持續獲得最新的分析。

使用者在上傳資料後，可以依需求更新資料，但我刻意不在介面上設置資料刪除按鈕。從使用者的角度而言，分析的精準度與價值會隨著資料量越大而提升，因此沒有刪除的必要，即使是不小心上傳成過往的檔案，我們也都能辨識並直接去除重複資料。從產品的角度而言，這能讓客戶的資料沉澱在產品內，幫助增加客戶的留存。

雖然我們總共需要三筆分開的資料才能進行全部類型的預測，但我們讓使用者只要上傳「會員交易資料」，即可開始使用我們的名單推薦功能。部分預測類別需要更多類型的資料，介面上也會提醒使用者回到首頁上傳。

資料分析

在會員經營中，常見的幾個業務場景就是「再行銷」、「特定商品促銷」，以及「會員分級」。我們抽象化了這些場景，讓使用者不用去想需要設定哪些參數或維度，透過簡單的點擊即可完成獲得推薦名單所需的設定。

舉例來說，假設使用者想要刺激現有會員回購，投放再行銷訊息，使用者只要選取「刺激會員回購」並設定要投放行銷訊息的時間區間，即可取得我們推薦的投放名單。而對於要針對特定商品促銷的使用者，則只要再選擇想促銷的商品類別即可。

在資料分析的過程中，我們希望確保分析結果的品質，因此會要求使用者盡可能提供最新的資料。我透過刻意限制了可選的日期範圍來達成這個目的，當上傳的最後一筆資料日期是 D，則最多可以選到 D+90 的日期。只要沒辦法選擇投放日期的問題，就會促使使用者更新會員資料。

除了需要最新的資料，Repurchase 也需要確保資料量足夠我們的模型產出夠高品質的分析結果。假設 Repurhcase 判定資料量不足，就會提醒使用者需要更多的資料才能進行分析。

只要通過以上兩個資料品質的驗證機制，使用者就可以在很短的時間內取得 Repurchase 推薦的名單。透過整合 MoBagel 自研的 Decanter AI 技術，即使資料量龐大，多數名單皆能在一分鐘之內預測完畢。

預測結果呈現

當預測完畢，Repurchase 除了會提供推薦名單給使用者下載，也讓使用者自行設定該次行銷訊息的單位成本，以及預計要投放訊息給多少人，隨著這些數字的改變，行銷效益也會隨之改變。這個區塊實際上可以做出非常多的分析，但我讓資訊盡可能簡化，期望使用者可以透過行銷效益的資訊來判斷該下載多少人數的名單。

市場進入策略

在產品上線後，我與產品行銷經理立刻展開了市場進入策略（Go-to-market Strategy）的規劃，主要包含了內容策略、客戶獲取渠道，以及交付材料。

我們重新對齊了 Repurchase 在 8ndpoint 的定位，以及 Repurchase 如何在 MarTech 市場中實現差異化。具體而言，即使當前市場已經有許多同類競品，但缺乏了模組化能力且易用性不佳，這些因素增加了 SMB 導入數據工具的成本並降低了導入意願。在技術層面，MoBagel 的 Decanter AI 為市場上最快最準的 AutoML 技術，在結合 AI 顧問團隊後，能夠實現行業化應用落地。因此考量成本、使用者體驗、技術三個層面，我們能夠有自信滿足目標客群的需求。最後，我們也再次明確了 Repurchase 的客群與價值主張，並將這些內容呈現在了 Sales Kit 以及產品網站之中。

由於我們與 iProspect 的深度合作，這將會是我們獲取種子客戶的主要管道。另外我們也持續與教育單位進行課程與專案合作，目前 MoBagel 已與數十間台灣大學簽訂合作意向書，將嘗試導入 8ndpoint 產品線於合作課程，訓練管理學院與 MBA/EMBA 學生使用 8ndpoint 產品的能力，進一步推廣到學生工作的場域中，實現 Product-Led Growth 增長。

最後，透過研究其他 MarTech SaaS 計費以及計算產品能帶來的回報，我定義了 Repurchase 的訂閱價格與收費模式。這個收費模式綁定了 AI 分析的資源使用量，因為如果客戶確實感受到了產品價值，那麼資源使用量就應該會持續提升。具體價格目前暫時沒有公開，但如果有興趣，歡迎至官網聯繫專員。

數據埋點

我們使用的數據分析工具主要為 Amplitude，因此依據 Amplitude 的文件，我將所需的 Event 和 Event Property 彙整給前後端工程師並實現埋點。整個埋點框架主要包含三個部分：(1) 北極星指標 (2) 變現指標 (3) 使用者體驗指標，以下簡單做介紹。

北極星指標

整個 8ndpoint 產品線的北極星指標應該都是一致的，也就是「使用者多常使用預測分析」。這對應到了我們的價值主張，也就是希望透過 AI 預測的結果讓他們的工作更有效率。因此如果使用者感受到了產品價值，北極星指標就應該會有正向的表現。

變現指標

變現指標連結到了我們的 Go-to-market Strategy 中的商業模式。我們期望透過數據回答的問題包含 (1) 使用者上傳了多少資料？ (2) 使用者平均一個月執行了多少行銷活動？ (3) 使用者每次活動的行銷預算是多少？ (4) 使用者期望每次行銷活動能夠觸及到多少的顧客？如果能夠回答這些問題，那麼將有助於我們檢視不同等級方案的切分合理性，以及整體的收費模式是否合理。

使用者體驗指標

由於目前沒有太過複雜的產品架構，現階段的重點是期望使用者能夠完整體驗產品，因此在這邊我們會去量測 (1) 使用者是否能夠完成資料上傳？(2) 使用者是否能完成預測流程？

客戶體驗研究與產品迭代

在產品上線後，我們透過合作企業招募產品易用性測試者，在請客戶試用產品並給予我們回饋。透過三場約 1~2 小時的測試與訪談，我們得以更加了解行銷人員經營回購時的業務流程，也總結出許多值得排入迭代的產品需求。

研究主要發現

1. 營業額為王

在我們招募的客戶中，我們發現客戶並不在意成本的控制，反而是能不能達標營業額的 KPI。另外，使用者也不會特地去計算每個投放渠道的成本，而是根據過往經驗選擇最有效的渠道，假設沒有達標，再選擇新的渠道。名單設定的操作面板與預測成效的資訊呈現可能會因為這這兩點發現而大幅度的改變設計。例如在最終預測資訊的呈現上，應著重在運用營業額資訊幫助使用者決定投放人數或次數而非透過 ROI。

2. 不用刻意爭取使用者對 AI 的信任度

在一開始設計產品時，我們發現投資人希望能夠看到許多展現 AI 能力的資訊，例如 Repurchase 相比傳統的 RFM 模型可以多創造出多少營收。雖然我們在 Demo 給投資人看之後就拿掉了許多我們認為顧客並不一定在意的資訊，我們還是保留了 RFM 模型的 ROI 曲線，透過曲線高度間接展現 Repurchase 推薦的名單將有更好的成效。

然而在訪談中我們發現，這些為了強化 Repurchase 贏過傳統模型的資訊對於使用者而言反而會造成困惑與干擾。我們進一步了解到，使用者認為這些資訊對實際的工具操作並沒有幫助，雖然使用者確實會好奇 AI 是怎麼產生這些結果的，但他們也能理解機器學習的運作過程其實是黑盒子，所以即使沒有這些資訊也不影響他們使用 Repurchase 的意願。

在目前的版本中，我們已經先移除了 RFM 模型的 ROI 曲線，未來我們在 customer-facing 的產品中就不會再去設計與傳統方法進行比較的功能或資訊呈現。

3. 使用者具備一定的數據處理能力

在上傳資料的流程中，我原本擔心行銷人員在處理數據的能力如果不佳，可能會增加上傳資料的門檻，透過這幾次研究發現，我們現階段請使用者進行的資料預處理對他們而言完全沒有難度，非常快速就可以處理完。

這個發現讓我們未來在設計功能時，可以更加著重在儀表板的分析結果怎麼啟發行銷人員去設計會員經營活動，而不是放哪些資訊會吸引使用者上傳資料。

其他未來產品方向

除了上述的產品需求以外，綜觀整個 Repurchase 於 8ndpoint 的角色和 8ndpoint 未來的產品策略，除了 Repurchase 本身的產品體驗需優化以外，也包含前段的資料匯入，以及後段的資料運用。首先，我們還會嘗試去做跨產品的資料整合與覆用，透過廣告監測與庫存優化的資料來提升預測的準確度。其次，目前的推薦名單可以做其他進一步的 MarTech 應用，讓投放後的成效數據能夠回饋給現有模型，作為校準模型的一個方式。最後，Repurchase 的能力可能可以透過 API 開放，製作成其他 App 上架到類似 Shopify 的 Marketplace。MoBagel 一直以來都希望能夠打造一個開放的 AI 平台，讓 AI 與所有人都沒有距離，這也是我們與其他 AutoML 平台最大的差異之一。我們期待 8ndpoint 能夠成為這樣的產品，也希望 Repurchase 能在未來扮演關鍵角色。