A custom AI agent on Zapier MCP can help a real estate team interpret new lead activity, propose a bounded score observation, and send an approved update to its CRM. The safer pattern is not to let a model score and write anything it wants. Use a verified event source, retrieve the contact by stable CRM ID, give the model a narrow interpretation task, validate its structured result, then write it to a verified destination or route it for review.
This guide uses Zapier’s customer story about Rush Home as a starting point, then separates reported behavior from current vendor capabilities and proposed implementation choices. It is an architecture and operations guide, not a reproduction of Rush Home’s private build. Zapier MCP can expose configured app actions to a compatible AI client, but it does not define a scoring policy or provide validation, approval, authorization, or duplicate prevention by itself.
The practical model is source event, permitted context, bounded interpretation, structured observation, validation, then CRM action or human review. That distinction matters when a brokerage wants useful automation without turning an opaque model output into an untraceable CRM change.
What a custom AI agent on Zapier MCP can do for lead scoring
Zapier describes MCP as a server that exposes configured app actions as tools to a compatible client. Its current documentation says Zapier MCP uses Streamable HTTP and does not support SSE. This is separate from the Zapier MCP Client product, which has different transport support. Check compatibility for the specific client before connecting it. See Zapier’s MCP overview and the MCP quickstart.
In a lead-scoring design, MCP is the action interface. It can help a compatible client reach configured CRM actions, while the model interprets permitted unstructured context. Deterministic rules should decide whether a record is eligible for processing, whether an event is new, which fields may be written, and whether the result requires review.
A model may interpret a note that suggests someone is considering a purchase, summarize evidence, or suggest a next step. It should not decide whether the contact is the intended record, whether the person has opted out, whether a score field is writable, or whether an unvalidated update is authorized.
What the Rush Home story establishes, and what it leaves open
Zapier’s May 12, 2026 customer story describes Marcus Rush’s AI agent, Russ, as a system for scoring leads, preparing ranked daily briefs with follow-up suggestions, and piloting email-driven calendar management. The story reports that the workflow covered more than 11,000 contacts and used signals including lead source, buyer or seller indicators, geography, price points, response timing, and call duration.
These are customer-reported details, not independently audited implementation specifications. The public story does not disclose the scoring weights, prompts, field mappings, API requests, event identifiers, logs, review rules, or performance measurements. It describes calendar management as a pilot, not as a verified production capability.
The current Rush Home team page identifies Marcus Rush as founder and team leader and lists eight people as agents. Because team rosters change, that current page does not establish the exact team composition on the story’s publication date.
Use the customer story for what Rush Home reportedly did and vendor documentation for current product capabilities. Do not infer which trigger, custom field, prompt, or write mechanism Russ used. The Follow Up Boss MCP page lists actions that may be exposed through MCP; it is not a Russ template or a documented scoring build.
Choose the right job for MCP, the model, and CRM rules
Keep three responsibilities distinct:
- Rules: determine whether an event is new, the contact has a stable ID, the record is active and assigned, the person has not opted out, and the fields pass type and range checks.
- AI: interpret permitted free-text notes for likely buyer or seller intent, summarize evidence, extract a geographic preference, or propose a next step. Require structured output rather than an unrestricted CRM instruction.
- MCP and CRM actions: expose only the tools required and use a verified destination after checking current integration support, permissions, and field semantics.
Fixed tool lists can limit which tools an MCP server exposes, which supports a least-privilege design. They do not replace field-level authorization, approval gates, validation, audit history, or idempotency. Keep CRM and personal data access-controlled, minimize notes sent to the model, protect credentials, and name the person responsible for reviewing exceptions.
In real estate, the brokerage’s compliance owner should review the scoring policy. Do not use protected characteristics, or proxies for them, to prioritize or deprioritize people. Handle geographic signals only under a documented and lawful business policy.
Use deterministic rules to decide whether a record may be processed or changed. Use AI to interpret ambiguous context and propose a bounded result.
Design the lead-scoring pipeline around interaction events
Zapier currently documents Follow Up Boss triggers such as New Contact and Updated Contact, searches such as Find a Contact, and actions including Create Note, Log Call, Create Task, Add Tags, Change Stage, and an API Request action marked Beta. A documented contact trigger is not proof that every call, email, or note is available as an instant event. Confirm the exact trigger, account permissions, action, and target field before building around it. The Follow Up Boss setup guide is the relevant current reference.
A proposed scoring flow is:
- Receive a verified Follow Up Boss New Contact or Updated Contact event, or another verified interaction event.
- Retrieve the contact by its stable Follow Up Boss ID, not by a name alone.
- Gather only relevant, permitted context such as the new event, selected CRM fields, and a limited notes excerpt.
- Ask the model for a bounded interpretation and a structured observation.
- Validate identity, eligibility, field types, ranges, policy version, confidence, and event freshness.
- Write to a verified CRM destination or create a review task.
Keep three records distinct. An InteractionEvent is one source-system interaction. A LeadScoreObservation is one evaluation of one contact against one event and one policy version. The CRM’s current score is a projection of whichever approved observation the brokerage chooses to display. This preserves history when a later event or policy version changes the result.
The following is a hypothetical contract, not a vendor template or Rush Home export. Its identifiers and values are illustrative.
{
"contact_id": "FUB-illustrative-id",
"evaluated_event_id": "event-illustrative-id",
"score_value": 68,
"score_policy_version": "lead-score-v1",
"model_id": "configured-model-identifier",
"confidence": 0.82,
"rationale": [
"Recent response suggests active purchase interest"
],
"review_status": "pending",
"created_at": "2026-11-20T14:35:00Z"
}
Before a CRM write, confirm that the contact ID resolves to the intended record, the event is present and new, numeric fields parse and fall within policy limits, required fields are allowed, and the contact remains eligible. Store the source event ID, policy version, model identifier, rationale, timestamp, review state, and write status with the observation. Set a confidence threshold for automatic action, and route low-confidence or conflicting results to the assigned reviewer.
Use event-level idempotency. Recommended uniqueness rules are:
- InteractionEvent: unique on
source_systemplussource_record_id. - LeadScoreObservation: unique on
contact_idplusevaluated_event_idplusscore_policy_version.
A lookup followed by an insert can race when two workers process the same event concurrently. Enforce uniqueness with a database constraint, transactional upsert, or equivalent atomic operation. If the source provides no event ID, define a stable composite key from source, contact, event type, and source timestamp, then test whether it can distinguish genuinely separate events. Do not use contact ID plus date alone when several interactions can occur on the same day.
Missing location, buyer or seller information, or price data should be recorded as unknown rather than treated as a negative signal. Contradictory evidence should be quarantined or sent for review.
Turn scores into a daily brief agents can act on
Zapier reports that Russ sent agents daily ranked lead briefs with suggested follow-up or introduction tactics. The story does not provide a measured productivity result or the exact schedule and delivery configuration. For a brokerage’s own brief, define the grain as one contact per brief run, not one row per interaction or one row per score citation.
A contact, an interaction, a score observation, and a daily brief item are different records. Use an event key for interactions, an event-and-policy key for score observations, and a brief-run-and-contact key for brief items. A contact ID plus date can collapse separate same-day events.
A useful brief item includes the brief run ID, reporting window, assigned agent, contact ID, score-observation ID, rank, evidence-based reason, suggested action, CRM link, source-data cutoff, and generation time. Exclude contacts that are closed, opted out, archived, or reassigned under brokerage rules. Sort by score descending, then last interaction time descending, then contact ID ascending so ties resolve consistently.
Each brief should show why a contact was ranked, not merely a numerical score. Store agent corrections separately from the original model output so operations can audit recommendations and determine whether the brief reflects current CRM data.
A failed retrieval or ranking run should create an explicit failure notice or show the last successful brief with its timestamp. It should not silently send an empty or stale list.
| Workflow | Trigger or source | AI job and validation | Action and fallback |
|---|---|---|---|
| Score update | Verified CRM event, then find contact by stable ID | Interpret permitted notes. Validate identity, eligibility, ranges, confidence, policy version, and event-observation uniqueness. | Write an approved note or score representation to a verified destination, or create a review task. Integration owner handles authorization and rate-limit errors. |
| Daily brief | Scheduled run, then retrieve eligible contacts | Explain ranking and suggest a next step. Validate one item per contact per run, assignment, freshness, exclusions, and deterministic tie ordering. | Send to the assigned agent with a CRM link. Notify the operations owner on stale or failed retrieval or ranking. |
| Calendar pilot | Email thread, then extract request and check availability | Extract details and flag ambiguity. Validate time zone, duration, permissions, availability, attendee policy, and duplicate key. | Create a draft or approved event. Require human approval for uncertain times, external attendees, or changes to existing events. |
Keep calendar automation at a separate approval boundary
The Rush Home story reports an email-based calendar-management pilot, but does not identify the exact Gmail or calendar actions, permissions, approval policy, or duplicate handling. Treat scheduling as a separate workflow with its own owner and action log, not as an automatic extension of lead scoring.
For a hypothetical pilot, start with an email thread ID, participants, requested time, time zone, duration, calendar ID, and candidate availability. Have AI extract a proposed request and flag ambiguity. Do not let it infer consent to invite people or modify an existing meeting from unclear language.
Confirm the time zone and availability immediately before creating an event. Require a human to approve ambiguous times, external attendees, and changes to existing events. Use the calendar provider’s event ID for updates or cancellation. Deduplicate with a key based on the email thread and normalized request, not merely the subject and date.
Pilot, measure, and expand only after the write path is trustworthy
Begin in shadow mode: generate observations and briefs without changing production score fields. Compare recommendations with agent decisions and corrected records. Track duplicate-event rate, rejected or invalid writes, review rate, score changes later reversed, brief freshness, and agent adoption. If the brokerage wants to test conversion impact, define a baseline and comparison method first. The customer story supplies no independently measured outcome figures.
Check operational cost before increasing volume. Zapier’s current usage documentation says each successful MCP tool call consumes two tasks, including successful tests, while failed calls do not consume tasks. Recheck this time-sensitive rule before launch. A high-volume workflow that performs multiple searches and writes per interaction can consume tasks quickly.
The Follow Up Boss Zapier integration guide documents paid-account and permission requirements, a limit of up to five custom fields for the integration, and possible HTTP 429 rate-limit responses. Plan retries for transient failures such as rate limits, but do not blindly retry invalid data or authorization failures.
Follow Up Boss documents API-key authentication and OAuth for integration partners. Its API guidance says API keys provide access comparable to the user’s login credentials, so protect them and do not share unrestricted keys broadly. OAuth requires a registered system and client application. Verify the current integration path and exact write operation before choosing between standard Zapier actions and a direct API design.
- Confirm the event source, stable identifiers, and supported CRM destination.
- Assign owners for reviews, write errors, stale briefs, and failed runs.
- Enforce unique event and observation keys with an atomic write strategy where concurrency matters.
- Approve the scoring policy, allowed signals, exclusions, and confidence threshold.
- Budget expected MCP calls and test task-limit, permission, and rate-limit behavior.
- Set shadow-pilot acceptance measures before enabling production writes.
Expand only when ownership, audit history, exception handling, and measurable acceptance criteria are in place. Teams assessing current integration actions can explore Zapier automation consulting. Teams reviewing field ownership and score history may benefit from CRM systems consulting, while AI agent design and implementation can help define the agent’s bounded role and operational controls.
