Multi-Agent Workflow Orchestration Architect

Company or project name: B2B GrowthOps AI

Business process to orchestrate: Inbound lead qualification and sales handoff

Business goal: Qualify inbound leads quickly, prioritize high-fit companies, update CRM with structured notes, and route qualified leads to the correct sales owner without creating duplicate CRM records or unapproved outreach.

Agents involved: Intake Agent, Research Agent, Lead Scoring Agent, Validation Agent, CRM Update Agent, Sales Communication Agent, Reporting Agent, Escalation Agent

Current process: A form submission creates a lead. A human sales assistant checks the company website, reviews the lead message, updates CRM notes, assigns lead priority, and alerts the sales team. This is slow and inconsistent.

Current tools: Website form, CRM, Slack, email, company website research, internal lead scoring criteria

Data sources: Lead form fields, CRM records, company website, approved enrichment notes, sales qualification criteria

Actions agents may take: Classify leads, enrich company context from approved sources, draft CRM notes, draft sales handoff summary, draft outreach email, create internal summary, create reporting summary

Actions requiring human approval: CRM write actions, sales outreach emails, high-value lead routing exceptions, disqualification of enterprise leads, custom pricing promises

Customer-facing outputs: Draft sales outreach email only after approval

Internal-only outputs: Lead score, research notes, CRM update draft, sales handoff summary, reporting summary

Risk level: Medium

Sensitive data involved: Lead contact details, company information, CRM notes, sales qualification data

Memory requirements: Case memory for each lead. Team memory for approved scoring criteria, sales tone, and disqualification rules. No long-term storage of private lead messages unless authorized.

Tool access requirements: Read-only CRM access for most agents. CRM write only after approval. Drafting access for Sales Communication Agent. Slack notification draft after validation.

Validation requirements: Required lead fields, duplicate CRM record check, consent check, scoring evidence, confidence score, disqualification flags, approval before CRM update or outreach

Escalation rules: Escalate enterprise leads, unclear consent, conflicting CRM records, low-confidence scoring, missing required fields, pricing requests, legal or procurement questions

Human approval roles: Sales Operations Manager for CRM updates and routing exceptions. Account Executive for outreach email approval.

Audit requirements: Log workflow run ID, lead ID, agent outputs, score, evidence, validation result, approval decision, CRM update status, and final routing outcome.

Error-handling expectations: If research fails, continue with form data and mark enrichment incomplete. If CRM duplicate is detected, stop CRM update and escalate. If approval is delayed, hold outreach draft.

Fallback behavior: Route incomplete or conflicted leads to human review.

Need output: Workflow architecture, agent role map, input/output contracts, approval gates, memory policy, audit log design, implementation roadmap, and JSON blueprint

Special constraints: Do not send customer-facing messages automatically. Do not update CRM without approval. Do not store sensitive lead details in long-term memory.

=== MULTI-AGENT WORKFLOW ARCHITECTURE ===

Workflow name: Inbound lead qualification and sales handoff

Company/project: B2B GrowthOps AI

Business goal: Qualify inbound leads quickly, prioritize high-fit companies, update CRM with structured notes, and route qualified leads to the correct sales owner without duplicate CRM records or unapproved outreach.

Risk level: Medium

Recommended orchestration pattern: Sequential pipeline with validation gates and human-in-the-loop approval before CRM writes or customer-facing outreach.

Why this pattern: The workflow has a clear order: intake, enrichment, scoring, validation, approval, CRM update, sales communication draft, and reporting. Because CRM updates and outreach can affect customer relationships and sales operations, write actions and outbound messages require human approval.

Agents:

1. Intake Agent Purpose: Collect and normalize lead submission data.

Responsibilities:

• capture form fields
• check required fields
• assign workflow run ID
• detect missing lead data
• identify consent field if available

Not responsible for:

• scoring the lead
• updating CRM
• sending outreach

Inputs: Lead form submission

Outputs: Lead intake packet

Tools allowed: Read-only form data and drafting output

Memory allowed: Case memory only

Escalation triggers: Missing email, missing company name, unclear consent

Human owner: Sales Operations Manager

2. Research Agent Purpose: Enrich lead context from approved sources.

Responsibilities:

• review company website if allowed
• summarize business model
• identify industry
• identify potential use case
• collect evidence for scoring

Not responsible for:

• deciding qualification
• updating CRM
• contacting the lead

Inputs: Lead intake packet

Outputs: Company research packet

Tools allowed: Approved research sources only

Memory allowed: Case memory only

Escalation triggers: Research unavailable, conflicting company identity, restricted source required

Human owner: Sales Operations Manager

3. Lead Scoring Agent Purpose: Score lead fit using approved qualification criteria.

Responsibilities:

• apply scoring rubric
• identify fit level
• identify urgency, company size signal, budget signal, and use case
• flag disqualification risks
• provide confidence score

Not responsible for:

• final disqualification of enterprise leads
• pricing promises
• CRM writes
• outreach

Inputs: Lead intake packet and company research packet

Outputs: Lead scoring packet

Tools allowed: Read approved scoring criteria

Memory allowed: Team memory for approved scoring criteria, case memory for current lead

Escalation triggers: Low confidence, enterprise lead disqualification, pricing request, legal/procurement question

Human owner: Sales Operations Manager

4. Validation Agent Purpose: Check whether the lead packet is complete, consistent, supported by evidence, and safe to route.

Responsibilities:

• check required fields
• check duplicate CRM risk
• check consent status
• verify scoring evidence
• check confidence
• decide whether approval or escalation is required

Not responsible for:

• changing the score without evidence
• contacting the lead
• writing CRM records

Inputs: Lead intake packet, research packet, scoring packet

Outputs: Validation result

Tools allowed: Read-only CRM check if available

Memory allowed: Case memory only

Escalation triggers: Duplicate CRM record, missing consent, conflicting data, low-confidence score

Human owner: Sales Operations Manager

5. CRM Update Agent Purpose: Prepare CRM update draft after validation.

Responsibilities:

• draft CRM note
• draft lead priority field
• draft routing recommendation
• prepare update request for approval

Not responsible for:

• writing CRM without approval
• creating duplicate records
• deciding final owner without approval

Inputs: Validated lead packet

Outputs: CRM update draft

Tools allowed: Drafting tools; CRM write only after explicit approval if authorized

Memory allowed: Case memory only

Escalation triggers: Duplicate record, unclear owner, missing required CRM field

Human owner: Sales Operations Manager

6. Sales Communication Agent Purpose: Draft outreach message for approved leads.

Responsibilities:

• draft personalized outreach email
• use approved sales tone
• avoid unsupported claims
• include relevant lead context
• prepare approval request

Not responsible for:

• sending email without approval
• making pricing promises
• making contractual commitments

Inputs: Validated lead packet and sales routing decision

Outputs: Outreach email draft

Tools allowed: Drafting tools only

Memory allowed: Team memory for sales tone, case memory for current lead

Escalation triggers: Pricing request, legal claim, procurement request, compliance-sensitive message

Human owner: Assigned Account Executive

7. Reporting Agent Purpose: Summarize workflow outcomes.

Responsibilities:

• count leads processed
• summarize qualification distribution
• track escalation reasons
• track approval outcomes
• create weekly report

Not responsible for:

• changing lead decisions
• contacting leads

Inputs: Final workflow logs

Outputs: Reporting summary

Tools allowed: Read-only workflow logs

Memory allowed: Aggregated team memory, no raw sensitive lead details unless authorized

Escalation triggers: Audit gap, unusual rejection rate, missing workflow logs

Human owner: Sales Operations Manager

=== RESPONSIBILITY MATRIX ===

Step: Lead intake

Responsible: Intake Agent

Accountable: Sales Operations Manager

Consulted: None

Informed: Validation Agent

Approval required: No

Step: Company research

Responsible: Research Agent

Accountable: Sales Operations Manager

Consulted: Lead Scoring Agent

Informed: Validation Agent

Approval required: No

Step: Lead scoring

Responsible: Lead Scoring Agent

Accountable: Sales Operations Manager

Consulted: Research Agent

Informed: Validation Agent

Approval required: No for draft score; yes for enterprise disqualification

Step: Validation

Responsible: Validation Agent

Accountable: Sales Operations Manager

Consulted: Lead Scoring Agent, CRM Update Agent

Informed: Escalation Agent if needed

Approval required: Only if risk triggers appear

Step: CRM update draft

Responsible: CRM Update Agent

Accountable: Sales Operations Manager

Consulted: Validation Agent

Informed: Reporting Agent

Approval required: Yes before CRM write

Step: Sales outreach draft

Responsible: Sales Communication Agent

Accountable: Assigned Account Executive

Consulted: Lead Scoring Agent

Informed: Sales Operations Manager

Approval required: Yes before sending

Step: Reporting

Responsible: Reporting Agent

Accountable: Sales Operations Manager

Consulted: Workflow audit logs

Informed: Sales leadership

Approval required: No for internal summary

=== INPUT/OUTPUT CONTRACT PACK ===

Handoff: Intake Agent -> Research Agent

Purpose: Provide normalized lead data for enrichment.

Input contract: Required fields:

• workflow run ID
• lead ID
• name
• email
• company name
• submitted message
• source
• timestamp
• consent status if available

Optional fields:

• role/title
• phone
• company website
• industry
• budget signal
• urgency signal

Missing-data behavior: If company name or email is missing, route to validation and request human review.

Output contract: Research Agent must produce:

• company summary
• industry
• website evidence if available
• use case hypothesis
• research confidence
• missing research fields
• source notes

Validation rules:

• do not invent company facts
• mark enrichment incomplete if evidence is weak
• do not use restricted sources unless approved

Failure behavior: Continue with intake-only lead packet and flag research incomplete.

Audit fields:

• workflow run ID
• lead ID
• research sources
• research confidence
• missing fields

Handoff: Research Agent -> Lead Scoring Agent

Purpose: Provide evidence for lead qualification.

Input contract: Required fields:

• lead intake packet
• company research packet
• approved scoring criteria

Output contract: Lead Scoring Agent must produce:

• lead score
• fit category
• urgency signal
• budget signal
• disqualification flags
• evidence notes
• confidence score
• recommended next action

Validation rules:

• score must reference evidence
• low-confidence scores must be flagged
• enterprise disqualification requires approval

Failure behavior: If scoring evidence is insufficient, mark as needs human review.

Audit fields:

• score
• scoring criteria used
• evidence summary
• confidence score
• escalation flag

Handoff: Validation Agent -> CRM Update Agent

Purpose: Permit CRM update drafting only after validation.

Input contract: Required fields:

• validated lead packet
• duplicate check result
• consent status
• approved score
• routing recommendation
• approval requirement

Output contract: CRM Update Agent must produce:

• CRM note draft
• field update draft
• routing recommendation
• approval request
• risk notes

Validation rules:

• no CRM write without approval
• no duplicate CRM record creation
• no unsupported claims in CRM notes

Failure behavior: If duplicate record exists, stop and escalate.

Audit fields:

• duplicate check result
• CRM draft fields
• approval status
• final CRM write status

=== VALIDATION GATES ===

Gate 1: Required field completeness

Checks:

• name
• email
• company name
• submitted message
• source
• timestamp

Fail behavior: Request missing information or route to human review.

Gate 2: Consent check

Checks:

• lead consent status if available
• outreach permission if required

Fail behavior: Do not draft outreach as ready to send. Escalate to human.

Gate 3: Duplicate CRM check

Checks:

• existing contact
• existing company
• existing open opportunity
• duplicate lead source

Fail behavior: Stop CRM update and escalate.

Gate 4: Scoring evidence check

Checks:

• score has evidence
• assumptions are marked
• confidence is included
• disqualification flags are justified

Fail behavior: Return to Lead Scoring Agent for revision or escalate.

Gate 5: CRM write approval

Checks:

• Sales Operations Manager approval
• approved CRM fields
• duplicate status clear

Fail behavior: Hold CRM update draft.

Gate 6: Outreach approval

Checks:

• Account Executive approval
• no pricing promise
• no unsupported claim
• correct lead context

Fail behavior: Hold email draft.

=== HUMAN APPROVAL STEPS ===

Approval 1: CRM update approval

Approver: Sales Operations Manager

Required before: Any CRM write action

Approval record:

• lead ID
• proposed fields
• lead score
• evidence
• duplicate check
• approver decision
• timestamp
• notes

Approval 2: Sales outreach approval

Approver: Assigned Account Executive

Required before: Any customer-facing email is sent

Approval record:

• email draft
• lead context
• personalization notes
• risk flags
• approver decision
• timestamp
• notes

Approval 3: Enterprise lead disqualification approval

Approver: Sales Operations Manager or sales leadership

Required before: Enterprise or strategic lead is marked disqualified

Approval record:

• disqualification reason
• evidence
• alternative routing considered
• final decision
• timestamp

=== MEMORY POLICY ===

Allowed memory: Case memory:

• current lead packet
• research summary
• scoring result
• validation result
• approval status
• routing outcome

Team memory:

• approved scoring criteria
• sales tone guidelines
• disqualification rules
• routing rules
• CRM field mapping

Long-term memory: Only aggregated, non-sensitive process learnings if authorized.

Must not store:

• raw private lead messages in long-term memory unless authorized
• sensitive personal data beyond the case requirement
• unauthorized CRM details
• unapproved pricing exceptions
• private customer information unrelated to the workflow

Retention: Case memory should persist only as long as required for the workflow and audit process.

Access rules: Research Agent can access intake packet. Lead Scoring Agent can access intake and research packets. CRM Update Agent can access validated lead packet only. Reporting Agent should use aggregated workflow data unless authorized.

=== TOOL ACCESS POLICY ===

Read-only tools:

• form submission data
• approved company research sources
• approved scoring criteria
• read-only CRM duplicate check
• workflow logs

Drafting tools:

• CRM note draft
• sales email draft
• routing summary
• reporting summary

Write actions requiring approval:

• CRM record update
• lead owner assignment
• Slack notification to sales team if it includes sensitive details
• sales outreach email

Forbidden autonomous actions:

• send customer-facing email without approval
• update CRM without approval
• delete CRM records
• create pricing promises
• disqualify enterprise leads without approval
• store sensitive lead details in long-term memory

=== ERROR HANDLING AND FALLBACK PLAN ===

Failure mode: Missing required lead data

Fallback: Return missing-data request to Intake Agent or route to human review.

Failure mode: Research unavailable

Fallback: Continue with form data only, mark enrichment incomplete, lower confidence, and route to human if score is uncertain.

Failure mode: Duplicate CRM record detected

Fallback: Stop CRM update and escalate to Sales Operations Manager.

Failure mode: Low-confidence score

Fallback: Send to Validation Agent and escalate if confidence remains low.

Failure mode: Approval delayed

Fallback: Hold CRM update or outreach draft. Do not proceed automatically.

Failure mode: Conflicting agent outputs

Fallback: Validation Agent compares evidence and routes to human if conflict affects routing, scoring, or customer-facing communication.

Failure mode: Tool unavailable

Fallback: Pause workflow, log tool failure, notify owner, and retry within defined limits.

Stop conditions:

• missing consent
• duplicate CRM conflict
• approval rejected
• unauthorized action requested
• required data unavailable after retry
• policy-sensitive issue without owner review

=== AUDIT LOG DESIGN ===

Required audit fields:

• workflow run ID
• lead ID
• timestamp
• agent name
• step name
• input summary
• output summary
• confidence score
• evidence notes
• validation result
• approval requirement
• approver decision
• CRM update status
• outreach status
• escalation flag
• error details
• final routing outcome

Review owner: Sales Operations Manager

Review frequency: Weekly during pilot, then monthly after stable rollout.

=== CONFLICT PREVENTION RULES ===

Single source of truth: CRM remains the system of record for lead status.

CRM write control: Only CRM Update Agent may draft CRM updates, and writes require approval.

Scoring ownership: Lead Scoring Agent produces the score, but Validation Agent checks evidence and confidence.

Outreach ownership: Sales Communication Agent drafts outreach, but Account Executive approves sending.

Duplicate prevention: Use lead ID, email, company domain, and CRM duplicate check before any record creation or update.

Conflict rule: If scoring, research, or CRM data conflict, Validation Agent must stop the workflow and escalate.

=== IMPLEMENTATION ROADMAP ===

Phase 0 — Process discovery: Document current lead qualification steps, required CRM fields, routing rules, approval owners, and disqualification policy.

Phase 1 — Contract design: Create lead intake packet, research packet, scoring packet, validation result, CRM update draft, and outreach draft schemas.

Phase 2 — Prototype: Run workflow manually using sample leads. No CRM writes. No customer-facing emails.

Phase 3 — Validation gates: Add completeness, consent, duplicate, scoring evidence, CRM approval, and outreach approval gates.

Phase 4 — Limited rollout: Use read-only CRM access and draft-only outputs for real leads. Human performs all writes and sends.

Phase 5 — Approved write workflow: Allow CRM Update Agent to write only after explicit approval if the platform supports approved write actions.

Phase 6 — Monitoring and optimization: Track completion rate, escalation rate, approval rate, duplicate detection, time saved, and sales acceptance rate.

Exit criteria:

• required fields are consistently captured
• duplicate detection works
• CRM updates are accurate
• outreach drafts are approved at acceptable rate
• audit logs are complete
• no unauthorized customer-facing messages are sent

=== JSON WORKFLOW BLUEPRINT ===

{ "workflow_name": "Inbound lead qualification and sales handoff", "business_goal": "Qualify inbound leads quickly, prioritize high-fit companies, update CRM with structured notes, and route qualified leads to the correct sales owner without duplicate CRM records or unapproved outreach.", "risk_level": "medium", "orchestration_pattern": "sequential_pipeline with human_in_the_loop approval", "agents": [ { "name": "Intake Agent", "role": "intake", "purpose": "Collect and normalize lead submission data.", "responsibilities": [ "capture form fields", "check required fields", "assign workflow run ID", "detect missing lead data", "identify consent field if available" ], "not_responsible_for": [ "lead scoring", "CRM updates", "customer outreach" ], "inputs_required": [ "form submission", "lead source", "timestamp" ], "outputs_required": [ "lead intake packet" ], "tools_allowed": [ "read_only", "drafting" ], "memory_scope": "case", "escalation_triggers": [ "missing email", "missing company name", "unclear consent" ], "human_owner": "Sales Operations Manager" }, { "name": "Research Agent", "role": "research", "purpose": "Enrich lead context from approved sources.", "responsibilities": [ "summarize company context", "identify industry", "identify likely use case", "collect evidence for scoring" ], "not_responsible_for": [ "qualification decision", "CRM update", "outreach" ], "inputs_required": [ "lead intake packet" ], "outputs_required": [ "company research packet" ], "tools_allowed": [ "read_only" ], "memory_scope": "case", "escalation_triggers": [ "research unavailable", "conflicting company identity", "restricted source required" ], "human_owner": "Sales Operations Manager" }, { "name": "Lead Scoring Agent", "role": "analysis", "purpose": "Score lead fit using approved qualification criteria.", "responsibilities": [ "apply scoring rubric", "identify fit category", "identify urgency signal", "identify budget signal", "flag disqualification risks", "provide confidence score" ], "not_responsible_for": [ "final enterprise disqualification", "pricing promises", "CRM writes", "outreach" ], "inputs_required": [ "lead intake packet", "company research packet", "approved scoring criteria" ], "outputs_required": [ "lead scoring packet" ], "tools_allowed": [ "read_only", "drafting" ], "memory_scope": "team and case", "escalation_triggers": [ "low confidence", "enterprise lead disqualification", "pricing request", "legal or procurement question" ], "human_owner": "Sales Operations Manager" }, { "name": "Validation Agent", "role": "validation", "purpose": "Check completeness, consent, duplicate risk, scoring evidence, and approval requirements.", "responsibilities": [ "check required fields", "check duplicate CRM risk", "check consent status", "verify scoring evidence", "flag approval requirements" ], "not_responsible_for": [ "sending outreach", "writing CRM records" ], "inputs_required": [ "lead intake packet", "research packet", "scoring packet" ], "outputs_required": [ "validation result" ], "tools_allowed": [ "read_only" ], "memory_scope": "case", "escalation_triggers": [ "duplicate CRM record", "missing consent", "conflicting data", "low-confidence score" ], "human_owner": "Sales Operations Manager" }, { "name": "CRM Update Agent", "role": "execution", "purpose": "Prepare CRM update draft after validation.", "responsibilities": [ "draft CRM note", "draft field updates", "prepare routing recommendation", "request approval before write" ], "not_responsible_for": [ "writing CRM without approval", "creating duplicate records", "sending outreach" ], "inputs_required": [ "validated lead packet" ], "outputs_required": [ "CRM update draft", "approval request" ], "tools_allowed": [ "drafting", "write_with_approval" ], "memory_scope": "case", "escalation_triggers": [ "duplicate record", "unclear owner", "missing required CRM field" ], "human_owner": "Sales Operations Manager" }, { "name": "Sales Communication Agent", "role": "communication", "purpose": "Draft outreach message for approved leads.", "responsibilities": [ "draft personalized outreach email", "use approved sales tone", "avoid unsupported claims", "prepare approval request" ], "not_responsible_for": [ "sending email without approval", "pricing promises", "contractual commitments" ], "inputs_required": [ "validated lead packet", "sales routing decision" ], "outputs_required": [ "outreach email draft" ], "tools_allowed": [ "drafting" ], "memory_scope": "team and case", "escalation_triggers": [ "pricing request", "legal claim", "procurement request", "compliance-sensitive message" ], "human_owner": "Assigned Account Executive" }, { "name": "Reporting Agent", "role": "reporting", "purpose": "Summarize workflow outcomes.", "responsibilities": [ "count leads processed", "summarize qualification distribution", "track escalation reasons", "track approval outcomes" ], "not_responsible_for": [ "changing lead decisions", "contacting leads" ], "inputs_required": [ "workflow audit logs" ], "outputs_required": [ "reporting summary" ], "tools_allowed": [ "read_only", "drafting" ], "memory_scope": "team", "escalation_triggers": [ "audit gap", "unusual rejection rate", "missing workflow logs" ], "human_owner": "Sales Operations Manager" } ], "workflow_steps": [ { "step_number": 1, "step_name": "Lead intake", "responsible_agent": "Intake Agent", "input_contract": { "required_fields": [ "name", "email", "company_name", "submitted_message", "source", "timestamp" ], "optional_fields": [ "role", "phone", "company_website", "budget_signal", "urgency_signal" ], "accepted_formats": [ "form submission", "JSON lead object", "CSV lead row" ], "missing_data_behavior": "Request missing fields or route to human review." }, "output_contract": { "required_fields": [ "workflow_run_id", "lead_id", "normalized_lead_packet", "missing_fields", "consent_status" ], "confidence_required": "medium", "evidence_required": false, "schema_required": true }, "validation_gate": { "gate_type": "completeness", "pass_behavior": "Proceed to Research Agent.", "fail_behavior": "Request missing data or escalate." }, "approval_required": false, "fallback_behavior": "Route incomplete lead to human review.", "audit_fields": [ "workflow_run_id", "lead_id", "missing_fields", "timestamp" ] }, { "step_number": 2, "step_name": "Company research", "responsible_agent": "Research Agent", "input_contract": { "required_fields": [ "lead_intake_packet" ], "optional_fields": [ "company_website", "industry_hint" ], "accepted_formats": [ "lead packet" ], "missing_data_behavior": "Continue with available data and mark enrichment incomplete." }, "output_contract": { "required_fields": [ "company_summary", "industry", "use_case_hypothesis", "source_notes", "research_confidence" ], "confidence_required": "medium", "evidence_required": true, "schema_required": true }, "validation_gate": { "gate_type": "evidence", "pass_behavior": "Proceed to Lead Scoring Agent.", "fail_behavior": "Mark enrichment incomplete and continue or escalate." }, "approval_required": false, "fallback_behavior": "Use intake-only packet with lower confidence.", "audit_fields": [ "research_sources", "research_confidence", "missing_research_fields" ] }, { "step_number": 3, "step_name": "Lead scoring", "responsible_agent": "Lead Scoring Agent", "input_contract": { "required_fields": [ "lead_intake_packet", "company_research_packet", "approved_scoring_criteria" ], "optional_fields": [ "purchase_intent_signal", "budget_signal", "urgency_signal" ], "accepted_formats": [ "structured lead packet" ], "missing_data_behavior": "Score with lower confidence or escalate." }, "output_contract": { "required_fields": [ "lead_score", "fit_category", "evidence_notes", "confidence_score", "recommended_next_action", "escalation_flag" ], "confidence_required": "medium", "evidence_required": true, "schema_required": true }, "validation_gate": { "gate_type": "quality", "pass_behavior": "Proceed to Validation Agent.", "fail_behavior": "Return for revision or escalate." }, "approval_required": false, "fallback_behavior": "Route low-confidence lead to human review.", "audit_fields": [ "lead_score", "fit_category", "confidence_score", "evidence_notes" ] }, { "step_number": 4, "step_name": "Validation", "responsible_agent": "Validation Agent", "input_contract": { "required_fields": [ "lead_intake_packet", "research_packet", "scoring_packet" ], "optional_fields": [ "CRM_duplicate_check_result" ], "accepted_formats": [ "validated schema objects" ], "missing_data_behavior": "Stop or escalate if critical fields are missing." }, "output_contract": { "required_fields": [ "validation_status", "duplicate_status", "consent_status", "approval_requirements", "escalation_reason" ], "confidence_required": "high", "evidence_required": true, "schema_required": true }, "validation_gate": { "gate_type": "risk", "pass_behavior": "Proceed to CRM update draft and outreach draft.", "fail_behavior": "Escalate to Sales Operations Manager." }, "approval_required": false, "fallback_behavior": "Stop workflow and escalate if duplicate, consent, or policy conflict exists.", "audit_fields": [ "validation_status", "duplicate_status", "consent_status", "approval_required", "escalation_reason" ] }, { "step_number": 5, "step_name": "CRM update draft", "responsible_agent": "CRM Update Agent", "input_contract": { "required_fields": [ "validated_lead_packet", "CRM_field_mapping", "approval_requirement" ], "optional_fields": [ "routing_notes" ], "accepted_formats": [ "validated lead packet" ], "missing_data_behavior": "Hold draft and request missing mapping." }, "output_contract": { "required_fields": [ "CRM_note_draft", "field_update_draft", "routing_recommendation", "approval_request" ], "confidence_required": "high", "evidence_required": true, "schema_required": true }, "validation_gate": { "gate_type": "human_approval", "pass_behavior": "Apply approved CRM write only if authorized.", "fail_behavior": "Hold or revise CRM update draft." }, "approval_required": true, "fallback_behavior": "Do not write to CRM until approved.", "audit_fields": [ "CRM_update_draft", "approver", "approval_decision", "CRM_write_status" ] }, { "step_number": 6, "step_name": "Sales outreach draft", "responsible_agent": "Sales Communication Agent", "input_contract": { "required_fields": [ "validated_lead_packet", "approved_sales_tone", "assigned_owner" ], "optional_fields": [ "product_interest", "pain_point", "urgency_signal" ], "accepted_formats": [ "validated lead packet" ], "missing_data_behavior": "Create generic draft or request clarification." }, "output_contract": { "required_fields": [ "outreach_email_draft", "personalization_notes", "risk_flags", "approval_request" ], "confidence_required": "medium", "evidence_required": true, "schema_required": true }, "validation_gate": { "gate_type": "human_approval", "pass_behavior": "Approved human may send outreach.", "fail_behavior": "Hold or revise draft." }, "approval_required": true, "fallback_behavior": "Do not send customer-facing message automatically.", "audit_fields": [ "outreach_draft", "approver", "approval_decision", "send_status" ] } ], "memory_policy": { "allowed_to_store": [ "case-level lead packet", "research summary", "lead score", "validation result", "approval status", "aggregated reporting metrics" ], "must_not_store": [ "raw private lead messages in long-term memory without authorization", "sensitive lead data beyond workflow need", "unauthorized CRM details", "unapproved pricing exceptions" ], "retention": "Keep case memory only as long as required for workflow completion and audit requirements.", "access_rules": [ "Agents access only the packets required for their role.", "Reporting Agent uses aggregated data by default.", "Team memory is limited to approved scoring criteria, sales tone, and routing rules." ] }, "tool_policy": { "read_only_tools": [ "lead form data", "approved company research sources", "read-only CRM duplicate check", "workflow logs" ], "drafting_tools": [ "CRM note draft", "sales outreach draft", "routing summary", "reporting summary" ], "write_actions_requiring_approval": [ "CRM update", "lead owner assignment", "sales outreach email", "sensitive Slack notification" ], "forbidden_actions": [ "send customer-facing email without approval", "update CRM without approval", "delete CRM records", "create pricing promises", "disqualify enterprise leads without approval" ] }, "audit_log": { "required_fields": [ "workflow_run_id", "lead_id", "timestamp", "agent_name", "step_name", "input_summary", "output_summary", "confidence_score", "evidence_notes", "validation_result", "approval_status", "CRM_update_status", "outreach_status", "escalation_flag", "final_routing_outcome" ], "review_owner": "Sales Operations Manager", "review_frequency": "weekly during pilot, monthly after stable rollout" }, "success_metrics": [ "lead qualification time", "CRM update accuracy", "duplicate record prevention", "approval rate", "sales acceptance rate", "escalation rate", "audit completeness rate", "unauthorized action count", "workflow completion rate" ] }