Overview
Identifying the owner of a property is straightforward — county records list their name. But turning that name into an actionable prospect profile is where the real work begins. Who is this person professionally? What is their estimated net worth beyond the property? How do you reach them? These questions traditionally require hours of manual research across LinkedIn, company databases, public records, and various people-search sites. Plotbook's AI owner research agent automates this entire workflow. Given an owner name and property address, the agent conducts an autonomous investigation across eight data sources, cross-references findings for accuracy, and delivers a comprehensive profile — typically in 10 to 30 seconds. The agent is built as a ToolLoopAgent powered by Google Gemini 3 Pro with high-effort reasoning capabilities, meaning it actively thinks through its research strategy, reasons about which of its 8 specialized tools to invoke next, evaluates intermediate results, and adapts its approach based on what it finds. This is not a rigid pipeline or a simple database lookup — it is an autonomous research agent that intelligently orchestrates its tools the way a skilled analyst would investigate, making real-time decisions about what to search for next based on what it has already discovered.
Agent Architecture
The AI agent is built as a ToolLoopAgent powered by Google Gemini 3 Pro (via OpenRouter) and the Vercel AI SDK v6's agentic framework. Unlike a rigid pipeline that always follows the same steps in the same order, this is a true autonomous agent — it thinks before it acts, reasons about what information it needs, and dynamically chooses which of its 8 specialized tools to invoke next. The agent operates with reasoning enabled at high effort, meaning the model actively thinks through its research strategy before each action. It can execute up to 30 reasoning steps per search and run up to 2 tools in parallel when independent queries would benefit from concurrent execution. The agent receives a structured prompt containing the owner's name, property address, and any additional context from the property record. Rather than following a fixed sequence, it formulates a research plan and adapts that plan based on intermediate results. If an early web search reveals that the owner is a CEO of a specific company, the agent immediately pivots to target that company in its professional database queries — rather than running generic searches that would waste time. The entire process runs server-side with real-time progress streamed to the browser via server-sent events (SSE). As the agent reasons, calls tools, and evaluates results, you see exactly what is happening: which databases are being searched, what results are coming back, and how the agent is adapting its strategy. This transparency is critical for wealth managers who need to trust the quality of the intelligence they are acting on. The architecture supports automatic retries if a tool call fails and graceful degradation — if one data source is unavailable, the agent intelligently routes around it using alternative sources.
Web Discovery & Identity
The agent typically begins its research with broad web discovery to establish who the property owner is beyond their county record — though as an autonomous agent, it may choose a different starting point if context suggests a more efficient approach. It uses neural search powered by Exa to find the person across the open web. Three specialized web search tools are available for this phase. The exa_people_search tool casts a wide net, looking for any web mentions of the person — news articles, company bios, professional directories, social media profiles, and personal websites. The exa_linkedin_search tool targets LinkedIn profiles matching the owner's name and geographic area, which often provides the most reliable professional context. The exa_public_records tool queries nine major people-search databases simultaneously — including WhitePages, FastPeopleSearch, TruePeopleSearch, Spokeo, BeenVerified, Intelius, ThatsThem, ZabaSearch, and PeopleFinder — to find recorded addresses, phone numbers, and associated persons. The agent reasons about which of these tools to invoke and may run multiple searches in parallel when it determines that concurrent execution would be faster. If the owner name is common (e.g., John Smith), the agent uses the property address, city, and state to disambiguate and find the correct individual. Based on what it discovers, the agent dynamically decides what to investigate next — it might immediately pivot to professional databases if it finds a strong LinkedIn match, or run deeper web research if initial results are sparse.
Professional Database Search
When the agent determines it has enough identity context to query professional databases effectively, it invokes the search_rocketreach and search_apollo tools. These platforms maintain verified records of hundreds of millions of professionals, including current and past employers, job titles, verified email addresses, phone numbers, and company details. The agent may query both databases in parallel — leveraging its smart parallelization capability to run 2 tools simultaneously — or it may choose to query one first and use those results to refine its query to the other. The search_rocketreach tool finds professional profiles with verified contact information, work history, and education. RocketReach is particularly strong for individual professionals, executives, and business owners. The search_apollo tool provides an alternative and complementary data source with emphasis on company context — organizational structure, company size, industry classification, and technology stack. By querying both databases, the agent maximizes coverage. Some professionals appear in RocketReach but not Apollo, and vice versa. The overlap between the two sources also provides a valuable cross-validation signal — when both databases agree on an email address or employer, confidence in that data point increases significantly. If neither database returns a match, the agent does not give up. It adapts its strategy — perhaps running deeper web research with exa_deep_research, trying alternative name variations, or leveraging whatever contact details it has already gathered from web sources and public records.
Contact Enrichment
Once the agent has identified the likely person and found matching professional records, it reasons about whether to perform full enrichment to retrieve the most detailed information available. The enrich_rocketreach tool returns the complete professional profile: all verified email addresses (personal and professional), direct phone numbers, full employment history with dates and titles, education background, and social media profile links. The enrich_apollo tool provides complementary details: company revenue estimates, employee count, organization LinkedIn URL, industry classification, and the person's specific role within the company hierarchy. The enrichment step is distinct from the search step because it retrieves substantially more data. The search tools find the person; the enrichment tools retrieve the full depth of information available about them. This two-step approach minimizes API costs by only performing expensive enrichment queries on confirmed matches rather than speculative searches. The agent intelligently decides which enrichment tools to invoke — if a RocketReach search returned a strong match but Apollo did not, the agent may only enrich through RocketReach rather than wasting a call on a weak Apollo match. The agent also extracts family connections during the research process, identifying spouses, children, parents, and siblings from public records when available. Family relationships are normalized into consistent categories and include source attribution and confidence scores so you know how reliable each connection is.
Deep Research & Gap Filling
Throughout the research process — and especially as the agent nears completion — it evaluates what information is still missing and performs targeted deep research to fill those gaps. The exa_deep_research tool uses Exa's neural search to find company websites, news articles, SEC filings, conference speaker bios, board memberships, and other sources that provide context beyond what contact databases contain. This is particularly valuable for high-net-worth individuals who may be retired, self-employed, or working at private companies that do not appear in standard professional databases. Because the agent is autonomous rather than following a fixed sequence, deep research can happen at any point in the investigation. If early web discovery reveals that the owner is a retired founder, the agent might invoke deep research immediately to find acquisition news or company valuations — rather than waiting until the end. The agent's reasoning capabilities mean it actively thinks about what would be most useful to search for next, adapting in real time. Once the agent determines it has gathered sufficient information, it synthesizes all findings into a final structured profile. It resolves conflicts between sources (e.g., if RocketReach and Apollo disagree on a current employer), weights more recent and authoritative sources higher, and produces a confidence score reflecting the overall reliability of the assembled profile. The entire autonomous research process typically completes in 10 to 30 seconds depending on how much information is available and how many sources the agent needs to consult.
The 8 Specialized Tools
The AI agent has eight specialized tools at its disposal, each optimized for a specific type of data retrieval — and the agent autonomously decides which to invoke, when, and in what combination. The exa_people_search tool performs broad web discovery using neural search to find mentions of a person across news, directories, and social media. The exa_linkedin_search tool targets LinkedIn specifically to find professional profiles, job titles, and company affiliations. The exa_deep_research tool conducts focused web searches for company context, board memberships, speaking engagements, and news coverage. The exa_public_records tool queries nine people-search databases simultaneously for address history, phone numbers, and associated persons. The search_rocketreach tool finds professional profiles in RocketReach's database of hundreds of millions of verified contacts. The search_apollo tool queries Apollo's database for person and company records with employment details. The enrich_rocketreach tool retrieves complete contact details including all emails, phones, and full work history for a confirmed match. The enrich_apollo tool retrieves detailed company intelligence, organization data, and the person's role within their company. The agent's reasoning engine — powered by Gemini 3 Pro with high-effort thinking — decides which tools to call and in what order based on its evolving understanding of the search. It can run up to 2 tools in parallel when their queries are independent, skip tools if earlier results already provide sufficient information, or invoke additional tools if initial results are ambiguous. This intelligent orchestration means no two searches follow exactly the same path — the agent adapts its strategy to each unique individual.
Confidence Scoring
Every profile produced by the AI agent includes a confidence score from 0 to 100, reflecting how reliable the assembled information is. The scoring model weights multiple factors. Each data source that confirms a piece of information contributes to the confidence score — a single source might provide 50 to 75 base points depending on its reliability tier. When multiple sources agree on the same data point, the score increases by 5 points per additional confirming source, up to a maximum of 100. Source reliability tiers are pre-configured based on historical accuracy. WhitePages is weighted at 75 points (high reliability), FastPeopleSearch at 70, BeenVerified at 65, and so on down to general web sources at lower tiers. Professional databases like RocketReach and Apollo are weighted heavily because their contact verification processes produce more reliable data than aggregator sites. Family connections include their own per-relationship confidence scores and source audit trails, so you can see not just that a person has a spouse but which source identified the relationship and how confident the system is in that connection. This multi-source validation approach means that a high-confidence profile (85+) has been verified across several independent data sources, while a lower-confidence profile (50-70) may rely on fewer or less authoritative sources. Both are useful — the confidence score helps you calibrate how much additional verification to do before acting on the information.
Real-Time Streaming
Unlike traditional batch-processing research tools that make you wait for a final result, Plotbook streams the AI agent's progress to your browser in real time. As the agent begins its search, you see which tool it is calling, what query it is running, and what results come back — all updating live on screen. This streaming architecture is built on server-sent events and the Vercel AI SDK's streaming capabilities. The agent's reasoning steps, tool calls, and intermediate results are pushed to the browser as they occur. You do not wait 30 seconds staring at a loading spinner — you watch the research unfold step by step. The practical benefit is transparency and trust. When you see the agent searching LinkedIn, querying RocketReach, cross-referencing public records, and synthesizing the results, you understand exactly where each piece of information came from. This auditability is essential for wealth managers who need to justify their prospecting data to compliance teams. The streaming display also shows tool call status indicators so you can see which sources returned results and which came up empty, giving you an intuitive sense of the profile's completeness before you even read the final output.
What You Get
The final output of an AI owner research query is a comprehensive structured profile containing the owner's full name, professional title, and current organization. Contact details include verified email addresses (both personal and professional where available), direct phone numbers, and LinkedIn profile URL. Professional background includes complete employment history with company names, titles, and date ranges, plus education history with institutions, degrees, and graduation years. Company intelligence includes the organization's website, LinkedIn page, industry classification, and estimated size. The wealth estimation section combines the property value with professional and organizational signals to produce a net worth range estimate. Family connections list identified relationships with names, relationship types, and confidence scores. Each profile also includes a research summary — a narrative paragraph written by the AI that synthesizes the key findings and highlights the most relevant information for a wealth management context. All of this information is saved to your Plotbook account and can be exported or integrated into your prospecting workflow. The profile becomes part of your saved prospects library, where you can add notes, track outreach status, and build your pipeline over time.