Optimize AI Results Using Prompt Controls

You can optimize the responses AI gives you by giving it relevant instructions, data, and tools. Learn about the many prompt controls included with Procore AI. This tutorial is for the following audiences:

• Procore customer wanting to optimize AI results after getting started in Procore.

• Datagrid customers wanting to optimize AI results after getting started in the Datagrid product.

• Prerequisites:

  • To optimize as a Procore customer with Procore AI: Your account must have an active subscription to one of the Procore Digital Coworker packages, and a Company Admin enabled your project and assigned you a role. See Enable Procore AI.

  • To optimize as a Datagrid without Procore AI: Your account has purchased and enabled the Datagrid product.

• Credit consumption: Credits are consumed by every AI question, agent task, and background process such as storing and syncing your data. Learn more about AI credits.

• Validation: AI can make mistakes. Always check all AI work, results, and answers against your own records to ensure accuracy.

1. Choose Your AI Mode

The AI mode determines the processing power and the toolset available for your request. You can find the mode selector at the top of the prompt composer. Note: 'Ask' mode cannot support certain connections via the 'Actions' menu nor the use of agents to create or modify items in Procore. Select a higher mode for those features.

2. Select Your Agent

By default, Datagrid uses Auto-Select Agent, which identifies the best specialized tool for your prompt. However, you can manually override this to use a specific agent for more precision with your task.

Click the mode name (e.g., Ask, Execute) to toggle offAuto-Select Agent and select a different agent from the dropdown list. You can select one or more specialized agents (e.g., Bill's Deep Search Agent or Budget Overrun Analyzer). Or, learn how to create an agent.

Note: Make sure the mode you selected in the previous step matches the outcome you need from the agent. 'Ask' is best for quick and simple assistance. 'Execute' is built for deep analysis and multi-step tasks.

Once selected, the agent’s icon will appear in the header next to the mode name so you always know who you’re chatting with.

3. Equip Your Toolkit - Actions and Tools

In Extended or Execute mode, you can toggle specific 'Actions' and 'Tools' on or off to give the AI more reach. Some examples include the following:

• Procore Enterprise: Enabled by default for Procore AI customers. This allows the AI to fetch and submit content directly to your Procore tools (RFIs, Submittals, etc.).

• Web Search: Enables the AI to look up information on the internet, such as external regulations or the current weather.

• Calculate: Helps with budget reconciliations or complex quantity take-offs.

• MCP (Model Context Protocol): An advanced data connector that allows the AI to securely access additional data sources, such as an internal database. This feature is only available to Enterprise accounts when using the Datagrid standalone product. See the 'MCP Servers' section of Configure an Agent for set up instructions.

• Connect Apps and data: Click thePaperclip icon to attach local files or connect to external data sources:

  • Connect Apps: Opens the 'Manage Apps' menu to link Datagrid to your Box, SharePoint, Google Drive, or Slack accounts.

  • Attach Local Files: Upload a PDF or spreadsheet directly into the chat for immediate analysis.

4. Set the AI's Knowledge Base

Define the AI’s context by connecting it to deep project records and learned agent history

@Use Knowledge

Guide the AI to read and retrieve facts from your project data. Select one or more data sources from @Use Knowledge to focus the AI on specific datasets or page environments.

• When to use it: Use this when you need to research specific details or facts buried in large files—like drawings, specifications, or historical project records. This is your best tool for "deep dives" into the thousands of pages that make up your project’s knowledge base.
  

  What's the difference between 'Knowledge' and 'Actions'?

  • Knowledge: This route is often best for retrieval of buried or complex information because it allows the AI to search through data that is pre-organized in a searchable index. It looks for facts in thousands of pages of information. This indexed data refreshes on a schedule, such as every 15 minutes or hour, depending on your Datagrid plan.

  • Actions: Use this route for the most current data (such as something added to a system 5 minutes ago) or when you want the AI to modify or create items for you. Selecting an Action or MCP connection (or telling the AI to use those modes) bypasses the index and gives the AI a "live" look and access to your project records, such as an RFI or submittal. When you do this to retrieve information, you'll need to be more exacting with your request because it looks for an exact match of items and fields.

Agent Memories

Click thebrain icon to view the personal knowledge the AI has stored about you.

• This includes stored context like your name, role, time zone, and project preferences to make responses more relevant over time. If the AI is consistently getting your title or location wrong, you can edit or clear these memories here.

5. Choose the Vehicle

Precision starts with the right structure. Manage your large language model (LLM) and prompt selections to optimize results.

Choose Your LLM

Click the model icon via thesparkle or brand logo at the top right of the prompt composer to switch the LLM.
Note: Model options can vary by account, depending on company configurations such as cloud platform.

Prompt Library

Click Prompts to save your most effective and complex prompts here to reuse them later. Toggle prompts between "Only for me" and "Share with team" to share your most effective prompts with your team and help standardize workflows.

6. Verify and Audit the Result

Never take an AI's word as final. Use the verification tools at the bottom of every response to audit the work:

• Reasoning: Click this to see the "hidden" step-by-step logic the AI used. It shows you the planning, the tools it selected (like the MCP or Calculate tool), and the documents it consulted to reach its conclusion.

  • Tip: If an agent or response seems "project blind," you can click Go to Configuration from themore menu next to it's name at the beginning of one of its responses and check whether it has access to the necessary datasets. Learn more about configurating an agent, or for a more temporary solution, add the necessary data to your current prompt via the 'Actions' or 'This Knowledge' menus.

• Citations: Click the source links to jump directly to the specific document or Procore item or PDF the AI used as a reference.

7. Try this prompt

To see these controls in action, try a prompt with your desired controls in place. For example, try the following:

A. Prompt Control Setup:

1. Select Execute mode

2. Enabled these in your Actions menu:

   • Procore Enterprise
     Note: If a relevant Procore dataset is already added to your page environment or teamspace, you do not need to select it here. Or if you have Procore AI, this will already be enabled.

   • Web Search

   • Calculate
     Note: If you do not manually select 'Web Search' or 'Calculate', the AI will likely still employ those tools if it can deduce from the prompt that they are relevant to the task.

3. Change the LLM to a Claude model.

4. Select the Deep Search Agent

B. Submit a prompt:

For example, type: Audit our latest submittal against LEED requirements. Identify any conflicts and estimate potential cost impacts.

C. Optional extension:

Ask the AI to draft a response or take follow-up actions. For example, type: Draft a rejection letter for the subcontractor.