Logic Conductor best practices

Logic Conductor builds each tool from your design: the name, the description and the arguments. An AI chooses a tool by reading those, and fills in inputs by reading the argument descriptions. Good tools come from designs written with that reader in mind.

Names

• Use a verb and an object. Approve Credit Limit, Calculate Shipping Cost, Check Quote Eligibility. The AI matches the name against what the user asked.
• Keep names unique and stable. Renaming an exposed artifact renames the tool, and agents or saved prompts that call it by name stop working.
• Avoid internal abbreviations. CalcDiscV2 tells an AI nothing; Calculate Volume Discount does.

Descriptions

The description is the tool’s instructions. Write it for someone who has never seen your project.

Weak	Strong
“Discount logic.”	“Returns the maximum discount percentage a sales rep may give on a quote, based on the customer’s tier and the quote total. Use before offering or approving a discount.”
“Credit check flow.”	“Checks whether an account’s requested credit limit can be approved, and if so updates the account’s credit limit. Changes data: call only when the user confirms.”

• Say what it does, when to use it, and whether it changes data.
• Mention important limits, such as “only for active accounts” or “amounts in the account’s currency”.

Inputs

• Describe every input argument. “Customer tier: Gold, Silver or Bronze, from the account’s Customer Tier column” is far better than a blank description.
• Use specific types. A lookup instead of a text ID, a choice instead of free-text status, a number instead of a number stored as text. Typed inputs keep the AI’s calls valid.
• Keep inputs few. If a tool needs ten inputs, the AI will guess some of them. Look values up inside the flow instead, from the record ID.

Outputs

• Return what the user needs to act on, with clear names: Approved, Approved Limit, Reason.
• Return a reason or message with yes/no answers, so the AI can explain the result.
• Use lists when there are several results, for example a list of eligible products. Remember that list arguments rule out Expose as Dataverse Action for the same block or flow.

Control what AI can do

• Expose only what you intend. Leave helper blocks and sub-flows unexposed. Expose the business-level entry point instead.
• Separate reads from writes. Prefer a block or a read-only flow for questions, and a separate flow for changes. Say clearly in the description which flows change data.
• Use business processes for governed work. Exposing a business process as start_… lets an AI start it, but stages, tasks, approvals and the process precondition still apply.
• Scope the project list. Expose only the projects whose logic AI should run. Fewer, sharper tools are easier for an AI to choose between.
• Plan restarts. New and changed tools appear after Logic Conductor restarts. Include it in your release steps.

Test before you expose

Run the block or flow yourself with realistic inputs (see Run and test), or generate and run tests with Logic Assurance. Then ask an AI client a few natural questions and check that it picks the right tool and fills in the inputs correctly. If it doesn’t, improve the description, not the question.

Related

• How Logic Conductor works
• Expose logic to AI and APIs
• Inputs and outputs