NAR 2025 in the United States: around 68% had tried AI

In the 2025 NAR survey in the United States, 32% said they had not actively tried AI in their business. Using a tool is not the same as changing an operation. Panama has no equivalent public measurement, so the useful starting point is measuring your own operation.

The uncomfortable figure: high adoption, little reported impact

The annual technology survey from the US National Association of Realtors (NAR, 2025 Technology Survey) received 1,241 responses from 49,233 invitations, a rate of 2.5%. Conversely, around 68% had tried AI in their business; 17% reported a significantly positive impact and 7% used chatbots to attract or communicate with clients. These are responses from NAR members, not the Panamanian market.

In that same survey, AI-generated content was the most commonly reported application, at 46% of respondents. That may improve a writing task, but does not by itself demonstrate an improvement in response, follow-up or closings.

Two commercial surveys show adoption among US consumers, using different samples and questions. Realtor.com reported that 82% of 1,000 adults interested in buying or selling, or who had done so in the previous two years, used AI for real estate information (Realtor.com, 2025). Veterans United reported that 45% of 859 prospective buyers had used AI during the buying process (Veterans United, Q2 2026). Neither sample allows adoption in Panama to be inferred.

Writing with ChatGPT is not transformation

Be clear about what does and does not count as a change.

Opening ChatGPT, pasting an apartment's details and getting a polished paragraph improves personal productivity. That is fine. But it is a tool you open and close: it is not part of your operation, cannot see your leads, does not know your inventory and does not access your calendar. On a day when you are in a meeting, that ‘AI’ does absolutely nothing for you, because you are its only user and you are busy.

Transformation requires connecting AI to a process and maintaining oversight. It can prepare answers, organize context or support automation, but relationships, negotiations and closing still need a person responsible.

Four maturity levels: from a prompt to an operation

Level 1: individual use. Someone requests copy, summarizes a call or suggests email subject lines. The potential benefit is saving some of the time spent on a task. The risk lies in pasting personal data, publishing incorrect information or treating a draft as a decision. The rule here is simple: minimal data, human review and a record of the original source.

Level 2: shared process. The team agrees on templates, tone, required fields and a review before publication. AI no longer depends on each agent's improvisation, but it still does not know the actual state of the inventory. The main safeguard is an authoritative record: price, availability, floor area and terms must come from the system, not the model's memory.

Level 3: connected information. The tool receives structured property and client data with defined permissions. It can draft a description from approved fields or summarize context for the person responsible. This level requires traceability: it must be possible to identify what information went in, what output was produced and who approved it.

Level 4: supervised automation. An action happens without someone pressing every button: classifying an inquiry, scheduling a reminder or preparing a message, for example. The greater the effect on a client, the greater the control required. Price, availability, contractual terms, financial advice and commitments should never depend on an unreviewed inference.

Moving up a level is not always the right answer. If inventory is outdated or no one owns follow-up, a more complex integration only moves the disorder faster. Define the process first; then decide which part warrants assistance.

What can be measured: an operating process

An operation can measure response delay, contact, qualification and the next action without turning an external association into a sales promise.

A historical InsideSales study observed more than 15,000 web leads and 100,000 call attempts at six US companies. It found associations between shorter delays, contact and qualification (Lead Response Management Study, 2007). It did not measure closings, Panamanian brokers or WhatsApp conversations. We therefore do not apply its multipliers to local commercial expectations.

A mystery-shopping study of 100 inquiries to more than 25 US firms found that 47% received no response, with an average of 8 hours and 17 minutes among those answered (Mike DelPrete, US, 2024). It is not a Panama statistic. Use it as a signal to measure your own operation, not to copy the percentage.

The practical difference is in the system: record, assign, respond and leave a next action. AI can assist with some of these tasks; the team remains responsible for what is sent and what is promised.

To measure a trial, choose a few metrics the team can audit: percentage of inquiries with an owner; median time to first response; percentage with a next action and date; incomplete records; human corrections before sending; and conversations that required escalation. “We use AI” is not a metric. “Of 40 drafts, 11 needed price or availability corrections” does provide an actionable signal.

The limits a professional operation must put in writing

Personal data. Do not paste documents, full conversations or financial information into a tool without understanding its policy, retention, access and contract. Use the minimum necessary and separate identifying data when a task can be completed without it.

Property facts. A model can fill a gap with a plausible sentence. Price, area, maintenance, parking, delivery date and availability must come from an approved source. If a field is empty, the correct response is to request confirmation, not invent it.

Discrimination and profiling. A recommendation must not exclude people based on protected characteristics or turn indirect signals into a housing decision. Define which fields automation may not use, and reserve sensitive decisions for a person with identifiable responsibility.

Commercial promises. AI must not guarantee appreciation, returns, bank approval, delivery dates or negotiation outcomes. It can present scenarios with visible assumptions; it must not turn a scenario into a guarantee.

Identity of the person speaking. If a conversation uses automation, the client must be able to understand who they are speaking with and how to reach a person. The escalation path matters as much as initial speed.

The category already exists abroad and is at an early stage in Latin America

On February 2, 2026, Lofty announced a product in the United States that the company itself called “Agentic AI OS” real estate. It is evidence of how a provider presents its product, not of broad adoption or validated results for the category.

The existence of that category in the United States does not prove that the same product or level of adoption exists in Latin America. Nor does it justify presenting a roadmap as finished software. A useful comparison examines each task and the actual status of each tool.

For a Panamanian broker, the existence of this category helps compare tasks, limits and responsibilities. It does not prove that the Latin American market is empty or that a local offering already has those capabilities.

What this means for you in Panama

Local context makes follow-up valuable. A residential sale recorded in ACOBIR's MLS took an average of 248 days to close, with 9% off the list price (Xposure system via Servmor, second half of 2025, n=235 closings). It is an MLS proxy, not a proxy for the entire market, but it shows why a visible next action matters over months.

WhatsApp can be part of that process when the operation needs it. FARO offers it as a managed integration using Meta’s official API, without presenting the inbox or autonomous agents as native FARO 0.5 capabilities.

The useful question is not whether AI writes or executes on its own. It is which tasks it can assist with while preserving human responsibility. The managed CRM and bot layer is explained separately from the current product.

A 30-day test that does not require faith

Week 1: baseline. Without changing the process, measure inquiries received, the person responsible, first response, next action and inventory errors. The baseline prevents attributing an existing problem or improvement to the tool.

Week 2: a single task. Choose a low-risk, high-volume task, such as turning approved fields into an initial description draft. Define the source, template, reviewer and the errors that require stopping the test.

Week 3: compare quality and time. Record how long the task took before, how long it takes with assistance and how many corrections it requires. A fast draft that requires checking every sentence may not improve the process. Include review time in the calculation.

Week 4: decide. Keep, modify or remove the flow. Only then consider connecting another task. Document the decision and its evidence: observed volume, time, errors, risks and owner. This turns adoption from a demonstration into an operational practice.

In FARO 0.5, the operational source connects information, inventory, the website, clients and content. That does not mean every process is autonomous. CRM and WhatsApp integrations are implemented as managed services when the business subscribes to them. That distinction between an available product and a configured service is part of our approach, not a limitation to hide.