InformationTitleOrchestration - AI Inference Node Use CasesURL NameOrchestration-AI-Inference-Node-Use-CasesStep-by-StepContents: OverviewInbound Triage Use CasesData Quality and Enrichment Use CasesRouting and Rep Enablement Use CasesBest PracticesSummary Overview The AI Inference Node lets you run a generative AI prompt during routing and save the response into LeanData variables for use in downstream nodes. This article walks through 10 common use cases for the node. Each one covers the problem it solves, how to set up the node, and how to route on the result. For setup steps, prerequisites, and configuration details, see Orchestration - AI Inference Node. Please Note: The AI Inference Node uses your own AI provider API key. You are responsible for AI provider usage costs and token consumption. Several tips below help you limit unnecessary calls. Inbound Triage Use Cases 1. Screen Out Spam, Vendor Pitches, and Junk Submissions The challenge: Contact Us and demo request forms collect more than buyers. Vendors pitching your company, agencies, job seekers, test submissions, and gibberish all land in the same queue. When these reach reps, they waste time and erode trust in routing. How to set up the node: Place the node near the top of your Lead graph, before matching and assignment.Pass in the form comments, along with name, title, company, and email.Ask the model to classify the submission into a short, fixed list (for example, Genuine, Vendor Solicitation, Job Seeker, Spam, or Uncertain) and return a one-sentence reason. How to route on the result: Send Spam, Vendor Solicitation, and Job Seeker records to a review path or disqualify them with an alert. Send Genuine and Uncertain records into normal routing. Tips: Let trusted sources (for example, event lists, list imports, or webinar registrants) bypass the node so you only spend tokens where spam is likely.Tell the model to be conservative. When it is not sure, it should return Uncertain so real buyers are not filtered out.You can also add a second node that checks the name, title, company, and email for placeholder or fabricated values and returns a true/false junk flag. 2. Keep Support, Billing, and Cancellation Requests Out of Sales Routing The challenge: Existing customers often use the sales form when they need support, billing help, or want to cancel. Keyword rules misfire because words like "cancel" or "support" also appear in real sales inquiries. How to set up the node: Pass in the form comments from the Lead or Contact.Ask the model to return exactly one category, for example: TECHNICAL_SUPPORT, BILLING_SUPPORT, CANCELLATION_REQUEST, NEW_SALES_INQUIRY, or UNCERTAIN. How to route on the result: Support requests: email the support inbox and recycle or disqualify the record.Billing requests: email the billing team.Cancellation requests: notify the billing team and create a task for the account owner so the churn risk is visible.New sales and Uncertain requests: continue into normal sales routing. Tips: Decide deliberately where the Time Out and Error edges go. If they fall through to sales routing, any failed call sends a support request to a rep. 3. Verify the Person and Company Are Real The challenge: Creating Accounts and Contacts for fake submissions or companies that do not exist clutters your CRM and skews reporting. How to set up the node: Use two nodes in sequence. Person check (web search off): Score how likely it is that the submission came from a real person, looking for disposable email domains, placeholder names, or spam in the notes. Return a 0-100 score and a short reason code.Company check (web search on): Check the email domain, website, and the company's public presence. Return a 0-100 score for whether the company is real and operating. How to route on the result: Create the Account only when both scores clear your threshold. Send everything else to a person with the model's reason, so they can settle it quickly. Tips: Cap the number of web searches in the prompt (for example, "run at most 2 searches"). Uncapped searches increase response time, cost, and the chance of a timeout.If the company check cannot confirm, you can add a third, deeper research node that runs only for those records. Data Quality and Enrichment Use Cases 4. Identify the Company from an Email Domain The challenge: Shorter forms convert better, but a form that only asks for an email address leaves the record without a company name, which matching and routing depend on. How to set up the node: Pass in the email address.Ask the model to return the most likely company name for the domain, or Generic Domain for free email providers. List the providers you want treated as generic (for example, gmail, yahoo, outlook, and icloud). How to route on the result: Write the company name to the record and continue into matching. Send Generic Domain records to a separate path, such as an outreach sequence plus a rep alert. 5. Find a Missing Website Before Matching or Account Creation The challenge: Records with a blank or invalid website often fail enrichment and matching and end up stuck, waiting for someone to fix them by hand. How to set up the node: Run the node only when the website fails your validation check.Turn on Web Search. Pass in the name, email, and company.Ask for the company's website as a full https URL, or an empty value if none is found. How to route on the result: Write the website back to the record, then continue into matching or Account creation. Tips: Add a decision node that checks for an empty result before writing it back, so you never write placeholder text into the website field. 6. Fill Missing Firmographics as an Enrichment Fallback The challenge: Enrichment providers leave gaps, and switching providers can leave Accounts without the HQ address, revenue, industry, or employee size that ICP scoring and assignment depend on. How to set up the node: Use one node per field, such as HQ address, annual revenue, industry and employee band, and website.Put a decision node in front of each one so it only runs when that field is empty.Include your exact picklist values in the prompt so the output can be written straight to the field.Turn on Web Search for values that need current public data, such as revenue. How to route on the result: Write the values back to the Account, then continue into ICP tiering and owner assignment. Tips: The empty-field checks keep token usage down because enriched Accounts skip the node entirely.Use a production API key with enough quota. Free or trial keys can hit provider rate limits during bulk updates. 7. Standardize Industry Across Data Sources The challenge: Each data provider uses its own industry taxonomy, but segmentation and territory rules need one clean industry value on every Account. How to set up the node: Pass in the industry values from each source (for example, industry fields and NAICS codes) plus the company description.Ask the model to map them to exactly one value from your internal industry list. Tell it which source to prefer, and to use web search only when the sources conflict.Return the industry plus a one-line explanation of which values drove the decision. How to route on the result: Write the industry to the Account. For a sub-industry, branch to one node per industry that picks from that industry's sub-industry list. Tips: Write each node's output to its own dedicated field. If two nodes write to the same field, they can overwrite each other. 8. Review Likely Duplicate Accounts The challenge: Rule-based duplicate matching can flag false positives such as parent and subsidiary companies, divisions, franchises, or companies with similar names. Reviewing every match by hand does not scale. How to set up the node: Place the node after your duplicate match.Pass in the name, website, domain, phone, and address for both the routed Account and the matched Account.Ask the model to return Duplicate, Not Duplicate, or Needs Review, a 0-100 confidence score, and a reason tag (for example, Exact Duplicate, Parent/Subsidiary, Division, or Franchise). How to route on the result: Flag high-confidence duplicates for merge, send mid-confidence and Needs Review results to a manual review queue, and let Not Duplicate records continue. Tips: Start with a high confidence threshold for automated merges and lower it as you build trust in the results.Tell the model which signals do not prove identity in your market, such as shared directories or listing sites that many companies use.Instruct the model to return Needs Review when its confidence is low, since a merge is hard to undo. Routing and Rep Enablement Use Cases 9. Route to Product Specialists from Inconsistent Campaign Names The challenge: You want to send a lead to the specialist for the product it showed interest in, but the product is buried in campaign names that are not consistently formatted. A decision tree that parses campaign names breaks every time naming changes. How to set up the node: Pass in the campaign name and any related campaign fields.Ask the model to return one product line from a fixed list, or Other. How to route on the result: Add one branch per product line to the matching specialist or round robin pool, and send Other to your default path. Tips: A single AI Inference Node can replace a long chain of decision nodes that parse campaign names, and it keeps working when your naming conventions change. 10. Write a Rep Briefing into the Routed Task The challenge: When leads go straight to sales, reps lose the qualification context an SDR would normally provide, and they have to hunt across Salesforce to learn why someone is reaching out. How to set up the node: Pass in the lead source, recent campaign, webinar, or event engagement, Contact Us text, title, and company.Ask for three outputs: a short intent phrase; a 2-3 sentence summary of who the person is, what they engaged with, and their likely product interest; and a recommended next step and channel.Consider a separate node per lead source (for example, webinar, event, or Contact Us) so each prompt fits the context. How to route on the result: After assignment, create a task for the owner with the briefing in the description. Tips: A second node can rewrite the draft to match your team's style, such as converting campaign codes into plain names.Asking reps to rate each briefing when they close the task gives you data to tune the prompt over time. Best Practices Follow these practices to get consistent results from the AI Inference Node and avoid common setup issues. Define a fixed set of return values and match them exactly. If the model returns "Spam" and your decision node checks for "spam," records skip the branch. Spell out the exact values in the output Instruction box.Plan your Time Out and Error edges. Decide on purpose where records go when the AI provider is slow or unavailable, rather than letting them fall through to a default path.Pass in every field your prompt refers to. If your prompt tells the model to use the email domain, add the email as a variable. The model cannot see fields that are not passed in.Gate the node with decision nodes. Only send records that need AI to the node. This reduces cost and the chance of hitting rate limits.Cap web searches. When Web Search is on, set a maximum number of searches in your prompt to control response time and cost.Use dedicated output fields. Do not reuse fields that hold other data, and do not have two nodes write to the same field.Keep your model current. AI providers retire older models on their own schedules. Check the supported models in the LeanData AI Feature Overview and update nodes before a model is retired.Test with Get Results. Routing Preview does not call AI providers, so use the node's Get Results test area with sample values before deploying. Summary The AI Inference Node is most valuable where routing depends on text, judgment, or public information that rules alone cannot handle: screening inbound submissions, filling and cleaning data, and giving reps context. Start with one use case, test it with sample records, and add a fallback path for timeouts and errors before you deploy. For questions or additional assistance, please contact LeanData Support.