Real estate agents can use AI to organize incoming inquiries, spot a stated request, and send each message to the right human queue. They should not let a model decide who is worthy of service, predict who will close, or make housing, lending, legal, or fair-housing judgments.

That distinction matters because “qualify the lead” can describe two very different jobs. One is administrative triage: identify whether a person asked about showings, selling, financing, property management, or something else. The other is a consequential judgment about the person. AI can help with the first when the categories are observable and every route receives appropriate attention. The second should stay with qualified humans operating under brokerage policy and applicable rules.

OpenAI’s new Decisions API makes the line timely. Released in public beta on October 6, the developer endpoint evaluates text, images, or both and returns a probability, a choice from fixed options, or a score against a defined rubric. OpenAI says the only model currently available is gpt-6-luna. This is an API for developers, not a new button in the standard ChatGPT interface, and beta behavior can change.

Use AI to route the work, not rank the person

A sound first use is deciding where a message goes. A website inquiry that says “I need to move in June and want to understand the showing process” might be routed to a buyer-information queue. A message asking for a valuation conversation might go to a seller queue. A maintenance request should go somewhere entirely different.

The categories should describe the requested work, not the perceived quality of the human. Avoid labels such as “good lead,” “low value,” “serious buyer,” or “unlikely to close.” Those labels can hide assumptions, produce uneven follow-up, and encourage a team to treat a prediction like a fact.

Make the fallback route obvious. OpenAI’s documentation specifically recommends an “other” option when fixed choices do not cover every input. In real estate, an unknown or low-confidence inquiry should reach a general human-review queue—not disappear, receive a colder experience, or wait indefinitely.

Probabilities are not proof

The API can return probabilities or confidence values, but a high number does not prove that the classification is correct. It is the model’s estimate for the question and evidence it received. A short message, typo, mixed request, unfamiliar language, or missing context can move the result.

OpenAI advises setting thresholds with labeled examples from the actual application and considering the cost of false positives and false negatives. For a real estate team, that means testing fictional or properly approved sample inquiries that a human has already categorized. Measure how often urgent requests are missed, how often messages reach the wrong queue, and whether the fallback catches uncertainty.

Do not begin by connecting the live CRM. First prove that the categories are distinct, the sample set reflects the messages the team receives, and every result remains visible to a person. A routing tool that saves thirty seconds but causes one important request to vanish is not an improvement.

Keep sensitive judgments outside the model

Lead routing can touch fair housing, privacy, lending, and brokerage obligations quickly. Do not ask AI to infer protected characteristics, financial capacity, family status, neighborhood fit, trustworthiness, or eligibility from a name, writing style, photo, address, or other proxy. Do not use it to decide who receives listings, responsiveness, opportunities, or encouragement.

Use the smallest data set the task requires. A routing test may need the inquiry text and requested service; it probably does not need a full contact record, financial documents, identification, or transaction history. Review the tool’s current data controls, account configuration, retention terms, regional availability, and brokerage policy before any real information is processed.

A safer AI lead-routing checklist

  • Define categories by the work requested, not a person’s predicted value.
  • Remove protected traits, sensitive details, and unnecessary identifiers.
  • Include an “other” or human-review route for uncertain messages.
  • Give every legitimate inquiry a timely service path.
  • Test with fictional or properly approved, human-labeled examples.
  • Track wrong routes and missed urgent requests, not just speed.
  • Keep the original message visible beside the AI result.
  • Require a person to review consequential decisions and exceptions.
  • Confirm privacy, brokerage, fair-housing, and vendor requirements before live use.
  • Pause the workflow if categories, model behavior, or business rules change.

The practical decision today

Do not automate “lead qualification” as one vague task. Separate it into observable routing and human judgment. Test one draft-only sorter with a few neutral categories, a generous fallback, and a clear review owner. If it routes work reliably without ranking people or changing service, it may earn a limited role.

If you want help turning that idea into a small, reviewable business experiment, join the free AI Agents for Agents Skool community. The detailed prompts, worksheets, and implementation lessons live there.

Primary source: OpenAI, Decisions API guide, public beta documented October 6, 2026; accessed October 7, 2026.

This article is educational and does not provide individualized legal, privacy, lending, security, fair-housing, tax, or compliance advice. Follow your brokerage’s policies and obtain qualified guidance for your market and workflow.

Route the request. Respect the person.

Build useful AI handoffs with a human still accountable.

Join the free community for practical lessons on testing AI workflows, setting review boundaries, and keeping client relationships human.

Join AI Agents for Agents free on Skool