Add Copilot Studio WorkData pattern for structured AI-assisted form and entity collection

[jirihoracek-creator]

Problem or Use Case

Copilot Studio agents often need to help users complete structured, multi-turn forms or assessments where the AI orchestrator should inspect current state, extract missing information from natural language, ask follow-up questions, and persist collected evidence.

There are two common variants:

1. Fixed questionnaire matrix: predefined rows where each row is a question/checklist item and answer/status/evidence columns are gradually filled until complete.
2. Dynamic entity collection: homogeneous table where user or AI can add as many records as needed, such as AI features, risks, stakeholders, or requirements. A seed/example row may expose schema to orchestrator, then be removed on first real insert.

Without reusable guidance, these topics are easy to scatter across topic list, hard to name consistently, and error-prone to author manually.

Proposed Solution

Add reusable WorkData authoring pattern to Copilot Studio Author plugin.

Pattern should generate or guide creation of three topics per structured data section:

• WorkData
  Initialize: OnActivity Message topic that initializes or seeds AI-visible global variable when blank.
• WorkData
  Record: OnRecognizedIntent setter/upsert topic with AutomaticTaskInput inputs and shouldPromptUser=false. For fixed questionnaire data, update existing row by key. For dynamic collections, add new row or patch existing row by key.
• WorkData
  Get: OnRecognizedIntent getter/status topic returning full current table/record for review and completeness checks.

Also include naming guidance, global variable guidance, agent instruction snippets, optional Dataverse JSON save/load guidance, and examples for fixed questionnaire matrices and dynamic entity collections.

Skill attached

SKILL.md

openai.yaml

workdata-pattern.md

Alternatives Considered

No response

Area

Skills

Additional Context

Pattern has been used in AI Act assessment agent that collects information / evidence for AI Act assessment. If the user does not provide any details, it leads them through the questionnaire and collects information. In case the details are provided by the user (e.g. a tool description), it extracts all the required information, records it and then only asks for the missing details.

Examples:

• SystemDetails: fixed questionnaire matrix with question_code, question_name, question_description, user_answer. Record topic updates answer by question_code.
• FeatureDetails: dynamic collection of AI feature records. Initializer seeds EXAMPLE row to show schema. Record topic removes example row on first real insert, generates feature_number, and patches only non-empty fields on update.

[jirihoracek-creator]

OK, I have added one more scenario - Card-driven multi-branch questionnaire mode

For fixed questionnaires, support an optional card-driven interaction mode where the AI does not directly record inferred answers. Instead, the AI processes only the current question, optionally infers an answer and evidence comment from conversation context, and passes those values as prefill inputs to a question card topic.

The card topic shows the current question to the user, pre-fills inferred values when available, lets the user confirm or correct them, and only then calls the WorkData Record topic. This keeps the user experience consistent, avoids asking some questions in chat and others in cards, and reduces orchestration complexity because the AI always calls the card topic for each question.

Skill files:

SKILL.md

workdata-pattern.md

openai.yaml

A sample questionnaire definition:

                =[
                  {
                    code: "ProductName",
                    text: "Product Name",
                    description: "Enter the product name to keep the assessment evidence traceable. This does not affect the CRA scoping outcome.",
                    type: "free_text",
                    user_answer: "",
                    user_comment: "",
                    answer_source: "",
                    confirmation_status: "",
                    uncertainty: "",
                    flow_yes_next: "ProductNumber",
                    flow_yes_exit_description: "",
                    flow_no_next: "ProductNumber",
                    flow_no_exit_description: ""
                  },
                  {
                    code: "ProductNumber",
                    text: "Product Number",
                    description: "Enter the product number, product ID, SKU, or internal reference if available. This does not affect the CRA scoping outcome.",
                    type: "free_text",
                    user_answer: "",
                    user_comment: "",
                    answer_source: "",
                    confirmation_status: "",
                    uncertainty: "",
                    flow_yes_next: "01",
                    flow_yes_exit_description: "",
                    flow_no_next: "01",
                    flow_no_exit_description: ""
                  },
                  {
                    code: "01",
                    text: "Is this product made available on the EU market?",
                    description: "The product is sold, licensed, subscribed to, or otherwise supplied for use in at least one EU/EEA country.",
                    type: "yes_no",
                    user_answer: "",
                    user_comment: "",
                    answer_source: "",
                    confirmation_status: "",
                    uncertainty: "",
                    flow_yes_next: "02",
                    flow_yes_exit_description: "",
                    flow_no_next: "Exit",
                    flow_no_exit_description: "The Cyber Resilience Act only applies to products made available on the EU market."
                  },
                  {
                    code: "02",
                    text: "Does the product contain software or programmable logic?",
                    description: "Includes embedded software, firmware, applications, scripts, services, or any programmable digital component.",
                    type: "yes_no",
                    user_answer: "",
                    user_comment: "",
                    answer_source: "",
                    confirmation_status: "",
                    uncertainty: "",
                    flow_yes_next: "03",
                    flow_yes_exit_description: "",
                    flow_no_next: "Exit",
                    flow_no_exit_description: "Products without software or programmable digital elements are generally outside the scope of the Cyber Resilience Act."
                  },
                  {
                    code: "03",
                    text: "Is the product intended to connect to another device, network, or cloud service in normal use?",
                    description: "Includes internet, LAN, Bluetooth, Wi-Fi, USB, APIs, companion apps, gateways, or cloud backends. Factory-only or debug-only connections do not count.",
                    type: "yes_no",
                    user_answer: "",
                    user_comment: "",
                    answer_source: "",
                    confirmation_status: "",
                    uncertainty: "",
                    flow_yes_next: "04",
                    flow_yes_exit_description: "",
                    flow_no_next: "05",
                    flow_no_exit_description: ""
                  },
                  {
                    code: "04",
                    text: "Does the product rely on remote data processing or remote services provided under the manufacturer's responsibility?",
                    description: "Examples include cloud services, update servers, configuration portals, remote control, APIs, or analytics services required for core functionality.",
                    type: "yes_no",
                    user_answer: "",
                    user_comment: "",
                    answer_source: "",
                    confirmation_status: "",
                    uncertainty: "",
                    flow_yes_next: "06",
                    flow_yes_exit_description: "",
                    flow_no_next: "06",
                    flow_no_exit_description: ""
                  },
                  {
                    code: "05",
                    text: "Does the product contain software but operate fully offline in normal use?",
                    description: "The product contains software but has no intended connectivity or dependency on remote services in normal operation.",
                    type: "yes_no",
                    user_answer: "",
                    user_comment: "",
                    answer_source: "",
                    confirmation_status: "",
                    uncertainty: "",
                    flow_yes_next: "Exit",
                    flow_yes_exit_description: "Products with software that operate fully offline and without remote services are generally outside the scope of the Cyber Resilience Act.",
                    flow_no_next: "06",
                    flow_no_exit_description: ""
                  },
                  {
                    code: "06",
                    text: "Is the product supplied in the course of a commercial activity?",
                    description: "Includes paid products, subscriptions, licensed software, freemium offerings, or products bundled with paid services.",
                    type: "yes_no",
                    user_answer: "",
                    user_comment: "",
                    answer_source: "",
                    confirmation_status: "",
                    uncertainty: "",
                    flow_yes_next: "07",
                    flow_yes_exit_description: "",
                    flow_no_next: "Exit",
                    flow_no_exit_description: "The Cyber Resilience Act does not apply to products not supplied in the course of a commercial activity."
                  },
                  {
                    code: "07",
                    text: "Is the product only a code sample, demo, or non-standalone example?",
                    description: "Applies to tutorial code, documentation examples, proof-of-concept snippets, or test code not offered as a standalone product.",
                    type: "yes_no",
                    user_answer: "",
                    user_comment: "",
                    answer_source: "",
                    confirmation_status: "",
                    uncertainty: "",
                    flow_yes_next: "Exit",
                    flow_yes_exit_description: "Standalone obligations under the Cyber Resilience Act do not apply to demo code or non-product examples.",
                    flow_no_next: "08",
                    flow_no_exit_description: ""
                  },
                  {
                    code: "08",
                    text: "Is the product security-critical or related to network, identity, or safety functions?",
                    description: "Examples include operating systems, VPNs, firewalls, IAM systems, cryptography, routers, smart locks, cameras, alarms, or similar components.",
                    type: "yes_no",
                    user_answer: "",
                    user_comment: "",
                    answer_source: "",
                    confirmation_status: "",
                    uncertainty: "",
                    flow_yes_next: "Exit",
                    flow_yes_exit_description: "The product is in scope of the Cyber Resilience Act and may qualify as an important or critical product requiring enhanced compliance efforts.",
                    flow_no_next: "Exit",
                    flow_no_exit_description: "The product is in scope of the Cyber Resilience Act as a standard product with digital elements."
                  }
                ]

[GiorgioUghini]

Can you share some screenshot or a solution that contains an agent with this pattern applied? I'm not sure I'm following 100%

[jirihoracek-creator]

OK, let's say the agent needs to collect information for a form like this (this is scenario 1 in the skill)

There are two ways how to get the answers:

• The user does not provide any background information (like product description). Then the agen will ask the questions one by one and save the answers to a JSON that will be then used to generate final report.
• The user provides the product description - then the agent infers the answers from the description and only asks the user the questions that could not be inferred.

There are three topics defined

They have the same prefix so they sit together in the topics list> WorkData prefix has been selected so that it sits at the bottom of the list and does not mess it up too much.

Initialize topic

In this case I created the JSON right in the topic but it can be also e.g. fetched from a text file or made up using certain rules.

[
  {
    question_code: "use",
    question_name: "External or internal use?",
    question_description: "External offering / xxxxx internal use / research only",
    user_answer: ""
  },
  {
    question_code: "solution_name",
    question_name: "Name of the solution",
    question_description: "Name of the IT Service / External offering",
    user_answer: ""
  },
  {
    question_code: "product_number",
    question_name: "Product number",
    question_description: "Mandatory only for external offerings",
    user_answer: ""
  },
  {
    question_code: "description",
    question_name: "Solution description",
    question_description: "Give a short description of the solution",
    user_answer: ""
  },
  {
    question_code: "ai_included",
    question_name: "AI? Yes or no",
    question_description: "IS AI part of the System's functionality?",
    user_answer: ""
  },
  {
    question_code: "ai_purpose",
    question_name: "Purpose of the AI functionality",
    question_description: "Describe shortly the purpose. What shall AI do in the solution? In case there are multiple features utilizing AI, describe each of them separately.",
    user_answer: ""
  },
  {
    question_code: "ai_vendor",
    question_name: "Vendor of the AI",
    question_description: "Who provides the AI models",
    user_answer: ""
  },
  {
    question_code: "ai_vendor_product",
    question_name: "Vendors Product name",
    question_description: "Vendor's Product",
    user_answer: ""
  },
  {
    question_code: "ai_vendor_version",
    question_name: "Product Version Number",
    question_description: "Vendor's Product Version",
    user_answer: ""
  }
]

Get topic
It simply returns the global variable content. It is called by the orchestrator after multiple calls of the Record topic to make sure the JSON is now complete ot what is still missing before proceeding with next steps (e.g. generating final report).

Record topic
This is used to record an answer to a single question to the JSON and possibly (like in this case) persist the JSON to a dataverse.

ForAll(
  Global.One_BasicInfo,
  If(
    question_code = Topic.InQuestionCode,
    Patch(
      ThisRecord,
      { user_answer: Topic.InUserAnswer }
    ),
    ThisRecord
  )
)

In the runtime, it then looks like this (based on the instructions of course)

[jirihoracek-creator]

Scenario no. two is then collecting a table of homogenous records. E.g. you need to collect/infer information about individual features in a product so you can perform a risk analysis on them. There may be N features and again, the AI can extract them from the description and save one by one, or the user can add them manually. Or a mixed mode.

Scenario 3 is then a structured questionnaire with branching logic. Just like MS Forms but with the possibility that the AI actually infers the answers from the provided description and change the direction of the questions according to the pre-defined questionnaire logic. Incl. that it can end the questionnaire if it hits an exit answer. In my agent I have modified this one so that it does not record the questions immediately but it only shows the question and inferred answer in an adaptive card and wait's for the user's confirmation of the answer to make 100% sure that the answer is what the user wants to submit (legal stuff).

[jirihoracek-creator]

Giorgio Ughini (@GiorgioUghini), comments added above