Back to Blog

    Agentic AI

    Copilot Studio: when generative AI helps and when it doesn't

    Answer open questions generatively, query data via a connector, make changes through validated actions: our decision rule for Copilot Studio agents.

    June 21, 20243 min read
    Marcel Haas

    Marcel Haas

    Solution Architect, CEO

    marcel.haas@cnext.ch
    CNEXT AI Agent
    In this article

    Our experience from building agents in Copilot Studio: generative AI pays off where language has to be understood or written. A simple data query runs more efficiently through a connector such as the Dataverse connector, and a change to data through an explicit, validated action. Copilot Studio supports both in the same agent; the work lies in choosing the right route for each action.

    The guiding question

    The most important question when designing an agent is not ‘How much AI is possible?’ but ‘What must happen reliably in this task?’ An open question benefits from a generative answer. A query, a booking or a data change needs clearly defined inputs and controls.

    Which mechanism for which task

    TaskExampleMechanism
    Knowledge question‘What is the travel expense rule?’Generative answer from approved sources, with citation
    Understanding or writing textDrafting a reply to a customer, summarising a documentGenerative AI
    Querying dataRetrieving a list of records from DataverseDataverse connector with a fixed query
    Changing dataChanging a customer address, approving a paymentExplicit action with validation, permission check and confirmation

    Why a data query does not need AI

    Copilot Studio: retrieving a list of data through the Dataverse connector

    Retrieving a list of data: no AI needed.

    For simple queries, the Dataverse connector is the more efficient choice:

    1. 1Accuracy: generative AI produces probabilistic results. The connector returns exactly the data requested, without that variability.
    2. 2Cost: running AI models can be expensive. For routine access, the connector is the cheaper option.
    3. 3Speed: connectors are designed for fast data queries; generative AI takes longer.
    4. 4Energy: generative models are compute-intensive. For a simple query, that effort is unnecessary.

    When data is changed

    When someone changes a customer address or approves a payment, the agent captures identity and inputs, validates the values, checks authorisation and asks for confirmation where needed. Only then may an action write to Dataverse or another system. Free-form text does not replace these rules.

    A simple design test

    1. 1Write down the expected outcome for three typical and two difficult cases.
    2. 2Mark every point where the state of a system changes.
    3. 3For knowledge questions, use limited sources and citations.
    4. 4For changes, use explicit actions with validation, permission checks and an error path.
    5. 5Test the cases in the agent and record where people have to step in.

    Limits of generative answers

    With generative answers, test whether outdated sources are excluded and whether the agent admits when it does not know something. Microsoft describes the capabilities and limits of generative answers, and its Copilot Studio guidance on topics explains how topics, tools and knowledge sources are orchestrated. Whether a step runs generatively or on a guided path depends on the risk of the individual task, not on the label ‘agent’.

    Agentic AIMicrosoft Copilot
    Teilen:

    AI tools may have supported the creation of this article. CNEXT is responsible for the published content. Product features and prices can change; check the linked primary sources for current details.

    Marcel Haas

    Marcel Haas

    Solution Architect, CEO

    6x Microsoft Applied Skills

    Discuss your agent design with CNEXT

    We work out with you which steps of your agent run generatively and which run through connectors and validated actions.