Quick start: Complete your first conversational analysis

Updated at:
Copy as MD

Complete an end-to-end conversational data analysis in ADA: connect a dataset, ask business questions in natural language, get conclusions with step-by-step execution details you can expand and verify, and then save the analysis as a daily automated report. Two tasks are all you need to pinpoint an issue. Following the main flow takes about 10 minutes.

Business scenario

Operations staff frequently need to answer a question like this: What are the order volumes and week-over-week changes by region over the last 7 days, which region has the largest fluctuation, and what might be the cause?

The traditional approach is: operations submits a data request, the data team schedules it, writes queries, and exports data. By the time the conclusions reach operations, the decision window has passed. If you want to add another dimension, you have to submit yet another request.

The following walkthrough uses this single question to drive the entire flow. The main path takes the shortest route — uploading an order details spreadsheet with no dependency on any pre-existing data source — to verify the end-to-end flow as quickly as possible. After completing the walkthrough, you can get order volume details, week-over-week comparisons, and root-cause clues in a single session, with the calculation logic and intermediate results expandable for verification at every step.

If your data is already in a data warehouse or database, or if metric definitions are scattered across documents, simply replace the data connection method in the main path. All subsequent steps remain the same. See "Other data starting points" below.

Solution architecture

The flow involves two entry points. The ADA portal console handles activation, subscription, and usage management. The AdaWorks workbench handles data connection, conversation execution, and result viewing. These are parallel entry points.

The data connection method is switched within the workbench and does not affect subsequent conversation steps. Therefore, changing the data starting point does not alter any operations after Step 3.

Procedure

The main path has five steps: confirm service availability, bring data into the conversation, run the analysis and verify the process, ask follow-up questions to pinpoint the issue, and save the analysis as a daily report. Each step provides a completion checkpoint so you can determine whether to proceed to the next step.

Step 1: Confirm service availability

This step confirms that your account can use ADA and determines the region for subsequent operations. Before you start, you need an Alibaba Cloud account that has completed real-name verification.

  1. Log on to the Alibaba Cloud console and go to the AI-Native Big Data Service (ADA) product page. In the region selector, select the target region. The region must match the engine instance you plan to connect later, because the instance list only loads instances in the current region. Available regions are determined by the options in the region selector, and the capabilities offered may vary across regions.

  2. Open the Overview page of the portal, expand the Subscribe to Adawork to Use Collaborative Agents section, and follow the card instructions starting from the Subscribe to Adawork card to complete the subscription. If the current account is not within the eligible scope, the page displays "Ada is in preview" and provides an Apply for Preview entry point.

  3. In the left-side navigation of the portal, check the subscription status under Activated Services.

Completion checkpoint: The "Version Subscription" card under "Activated Services" displays the number of seats, the quota per cycle, and the current quota cycle. At this point, the service is ready — all that is missing is data.

Step 2: Bring data into the conversation

This step gives the agent analyzable data. Open the AdaWorks workbench, click New Task in the sidebar, click Add Content on the left side of the input box, and select Local File to upload your order details spreadsheet.

The spreadsheet must contain columns such as order date, region, city, channel, and amount. The date column must use a standard date format (for example, 2026-09-06) and cover at least two complete periods; otherwise, week-over-week changes cannot be calculated. Use your own business data if available. If you do not have data on hand, use the sample data, which covers 14 days, four regions, and five channels, with columns and formats prepared according to the requirements above.

The sample data is available in the attachment . Download and upload it directly Sample Data - Regional Order Details.xlsx.

You can upload up to 6 files at a time, with a combined size limit of 48 MB. Files apply only to the current session, which is suitable for verifying a complete analysis first.

Completion checkpoint: The file entry appears above the input box and can be removed from the preview area before sending. At this point, the data is attached to the current session, and you can start the analysis.

Step 3: Run the analysis and verify the process

This step identifies where the issue lies while also verifying the execution path and calculation process, rather than just returning a number.

  1. Describe what you want to analyze in natural language. You do not need to decide which dimensions to break down in advance.

    Compare the order volume of the last 7 days with the previous week, and list the week-over-week changes by region
  2. After you send the message, Ada returns an Execution Plan. Review the data scope and breakdown dimensions in the plan, and proceed with execution only after confirming they are correct. If the plan does not match your expectations, add constraints directly in the conversation — you do not need to start a new session.

  3. During execution, a collapsible section appears above the response. While execution is in progress, the header reads Processing; after completion, it changes to Processed. Expand it to view the actual code executed and intermediate results. Tool calls that involve data access display a permission confirmation card with four options: allow this time, deny this time, always allow, and always deny. For a first-time execution with uncertain scope of impact, select allow this time. The four names above describe the behavior of each option; the actual button labels on the confirmation card may differ.

  4. After the response is complete, verify the calculation logic. Match the key values in the conclusions to the corresponding calculation steps in the execution process, and confirm that the statistical definition and time range are consistent with the task requirements. The uploaded file itself is visible in the Files tab on the right.

When cloud resource access is required for the first time, the page displays a Service-Linked Role Authorization Required prompt. This service-linked role allows ADA to access resources in other Alibaba Cloud services on behalf of the current identity. After you agree and create the role, execution continues automatically.

Completion checkpoint: The session generates a response that contains data tables and analysis conclusions, and the tool call cards can be expanded. If this step succeeds, the pipeline from describing a task to getting conclusions is established.

For reference when using the sample data: the total volume drops about 5%, which alone does not reveal the issue. However, breaking it down by region shows that East China drops by more than 10% — the only region with a significant decline — while the other three regions remain roughly flat. An aggregate number masking structural issues is a common scenario, which is why the task includes a group-by dimension. If your results diverge from this direction, first check the time range in the task, then expand the execution process to see whether the date column was treated as text. If the numbers do not match, first confirm the statistical definition (for example, whether canceled orders are included) and then re-describe the task.

Step 4: Ask follow-up questions to pinpoint the issue

The previous step identified that East China is declining. This step drills down to an actionable granularity. City and product category are the two angles that come to mind as soon as you see a regional decline. Writing them in a single task is faster than splitting them into two rounds. Continue in the same session — Ada retains the current data context, so you do not need to restate the background. Simply write "East China" directly.

Break down the reasons for the decline in East China order volume, listing week-over-week changes by city and product category respectively

For reference when using the sample data: East China's decline is concentrated in Hangzhou, while Shanghai and Suzhou show no decline. In Hangzhou, almost the entire drop comes from home appliances, with the other three product categories remaining flat. The conclusions may also mention that the mini-program channel has the largest decline, but that is because home appliance orders in Hangzhou are primarily placed through the mini-program channel. The channel is a symptom; the category decline is the actual issue.

Completion checkpoint: The response attributes the decline to specific cities and product categories rather than stopping at the regional level. Two tasks are sufficient to pinpoint the issue. The entire chain of analysis is preserved in the session — you can reopen it from the history list and continue analyzing.

After pinpointing the specific product category, if you need to determine whether the issue is a demand decline or a supply problem, ask one more follow-up question. This question can only be formulated after you have the results from the previous step:

For home appliances in Hangzhou, is the decline due to fewer orders placed or more order cancellations?

When using the sample data as a reference, you will see that fewer orders placed and more order cancellations occur simultaneously. The cancellation rate is significantly higher than that of other product categories. The combination of both points to a stock shortage in this category rather than a decline in demand.

Tables and report cards in a conversation can be previewed online and downloaded. To retain results long-term or share them across members, convert the output into artifacts. For more information, see Artifact management.

Step 5: Convert into a daily report

In this step, you replace "asking manually each time" with "running automatically every day." Validated analyses can be executed on a recurring schedule, so the output no longer depends on manual initiation. Scheduled tasks run in independent conversations, and each trigger's results are saved to the execution history for review. The entry point is Scheduled Tasks in the left-side menu of the workbench. This feature is in preview. If your account is not included in the preview, you will be prompted to apply for access and cannot create tasks. In that case, skip this step — the first four steps already constitute a complete analysis workflow. For details about how to create scheduled tasks, execution frequency, and execution history, see Scheduled tasks.

After the report runs automatically, the next step is to deliver the conclusions directly to your team instead of waiting for someone to check the platform. Configure a group bot Webhook URL for the task, and each execution's conclusions will be sent directly to DingTalk, Lark, or WeCom — the platform is determined by the URL. If you need to specify particular recipients or groups, select an IM bot instead. Your team can view the day's conclusions in their regular communication channels, and anomalies will not be missed simply because no one opened the platform. For configuration details, see Scheduled tasks.

Completion checkpoint: The scheduled task appears on the "My Scheduled Tasks" tab with a status of Running and a displayed next execution time. If the task does not trigger on time, check the "Execution History" tab for the most recent execution status and failure information.

Other data starting points

The following two starting points only replace Step 2. Step 3 and all subsequent steps remain exactly the same.

If your data already resides in a supported engine instance such as MaxCompute or Hologres, register a connector to query the data directly. A connector is the access configuration between ADA and an engine instance. Once registered, it can be reused long-term. Go to Connectors under "Extensions" in the sidebar, click Register Connector, select Big Data Engine as the service type, choose the engine type, confirm the region, select an instance from the instance list, and fill in the data location and access credentials. Write a clear description of what the data is and what scenarios it is used for, such as "E-commerce operations team order details for regional order analysis." The agent uses the description to determine when to invoke the connector. An overly vague description may cause the agent to select the wrong connector when multiple connectors are available. After saving, return to the new task and attach the connector in "Add Content." If the instance list is empty, verify the region first — only instances in the current region are loaded. For credential policies and a complete description of all fields, see Create a data connection.

When metric definitions are scattered across documents, create a knowledge base first to unify the definitions. This approach is suited for scenarios where the same metric, such as "active users," has different definitions across departments. A knowledge base centralizes these definition documents so the agent answers based on the definitions within it. Go to Knowledge Base in the sidebar, click Create Knowledge Base, upload metric definitions, region mapping rules, and other specification documents, and wait for the wiki build to complete before the content can be referenced. Then attach the knowledge base in "Add Content." Responses with an attached knowledge base will include Wiki References. You can expand the reference list and reference graph to see which definitions were used to derive the conclusions. This type of reference does not appear when only spreadsheets are uploaded. A knowledge base provides metric definitions rather than business data, so it must be used together with uploaded spreadsheets or connectors. In the task description, require the agent to compute metrics according to the definitions in the knowledge base and to state which definitions were used. A conversation can be associated with one or more knowledge bases, and each can be removed individually. For more information, see Knowledge base management.

Costs and risks

Conversations and task executions consume credits. Credits are allocated monthly based on your subscription plan and can also be purchased on a pay-as-you-go basis. The consumption during this quick start is minimal. For billing rules, see Billing overview.

After you authorize the service-linked role, ADA can access associated cloud resources according to the configured policies. We recommend that you minimize the scope of authorization to what is necessary for your business, and use row-level and column-level rules in data access control policies to restrict the data visible to each member, preventing data leaks in analysis scenarios. For more information, see Security and compliance.

Summary

You have now completed all four stages: data onboarding, conversational analysis, iterative drill-down, and automated scheduling. This validates the end-to-end workflow from a business question to a conclusion with a verifiable process. Uploading spreadsheets is suitable for validating the workflow, but the accuracy of metric definitions depends on the completeness of your knowledge assets. Data retrieval inferred solely from table schemas can still produce deviations when metric definitions are complex, and conclusions will not include definition references.

Choose one of the following directions to continue based on your needs:

  • Record the definitions of frequently used metrics in a knowledge base so that conclusions include definition references. For more information, see Knowledge base management.

  • Replace the data used in this analysis with a direct connection to an engine instance. For more information, see Create a data connection.

  • Subscribe to official agents for the engine you primarily work with. For more information, see AI Capability Center.

  • Control costs and the scope of permissions. For more information, see Billing overview and Security and compliance.