Your marketing problem. A workable system.
Turn a marketing operations problem into a workflow, field map, failure checklist and implementation brief. Explore a worked example or request an AI-assisted plan.
Marketing Workflow Architect
By Akshay Shaligram. Build a brief, map your workflow, review required data fields and define acceptance checks. JavaScript is required for the interactive builder. AI generation is subject to availability and server-enforced limits. The worked example is available without API usage.
Worked example: From demo request to accountable pipeline
Synthetic teaching scenario, not a personalised AI result.
A proposed lead-routing workflow for a synthetic B2B SaaS team. Keep identity, consent and routing deterministic. Use AI only to suggest a short problem summary for a person to review.
Assumptions
- The form is allowed to collect and store enquiry data; marketing permission is a separate choice.
- HubSpot is the contact system of record. Available API scopes, account features and routing rules still need checking.
- Campaign context is observational attribution, not evidence of incremental revenue.
Workflow
Capture the enquiry — Rule
Website form: Validate required fields and record the purpose of collection. Create a unique submission ID before sending.
Input: Name, reply email, problem and consent state.
Output: Validated enquiry with submission ID and timestamp.
If it fails: Keep the visitor’s text if delivery is unconfirmed. Do not show success before receipt.
Resolve identity — Rule
CRM integration: Look up an exact normalised email match and upsert the contact. Store submission IDs separately so repeated enquiries are not lost.
Input: Validated enquiry and submission ID.
Output: One contact reference and a distinct enquiry record.
If it fails: Send ambiguous matches to review. Retry transient failures with the same idempotency key.
Suggest a problem summary — AI-assisted
Restricted AI step: Send only a redacted problem description. Request a short summary and unknowns; do not infer purchase intent or sensitive attributes.
Input: Redacted free text with names and contact details removed.
Output: An explicitly AI-generated draft summary.
If it fails: If unavailable, continue with the original enquiry. This step must not block routing.
Assign accountable ownership — Rule
HubSpot: Use an agreed territory or service rule. Preserve an existing owner unless reassignment is explicitly approved.
Input: CRM reference, service selection and current owner.
Output: Named owner and follow-up task.
If it fails: Route missing owners to a shared review queue with an alert.
Review and respond — Human review
Sales or consultant: Review the original request alongside the draft. Decide whether and how to reply; enquiry consent is not a newsletter opt-in.
Input: Original enquiry, owner and optional AI draft.
Output: Reviewed follow-up and recorded next step.
If it fails: Escalate overdue tasks. Never treat an email click as a sent message.
Reconcile the handoff — Rule
CRM reporting and GA4: Compare form receipt IDs with CRM enquiry IDs. Join campaign context only where consent and identity policy permit.
Input: Submission receipt, contact reference and observed campaign fields.
Output: A report of matched, missing, duplicate and unresolved enquiries.
If it fails: Keep unmatched rows visible. Never silently count a missing attribution value as direct traffic.
Field map
submission_id
Form integration → Enquiry record. Stable key for receipt reconciliation and duplicate delivery prevention.
reply_email
Visitor input → CRM contact. Reply and exact identity matching; exclude from AI prompts.
problem_original
Visitor input → Enquiry record. Keep the unmodified request separate from any AI interpretation.
processing_consent
Form permission → Enquiry record. Record purpose, timestamp and wording version; keep marketing consent separate.
owner_id
Approved routing rules → CRM contact and task. Make responsibility explicit without silently overwriting an owner.
campaign_context
Consented session context → Enquiry record. Record observed source and its measurement window, including unknown values.
Implementation checklist
01 / Map and agree
Document form fields, permission language, CRM schema, owners and API permissions. Walk through five synthetic scenarios.
Ready when: A named owner approves the field map and routing rules; missing permissions are listed.
02 / Build the reliable path
Implement validation, stable submission IDs, contact lookup, ownership and retry handling before adding AI.
Ready when: Duplicate, timeout, invalid-input and missing-owner tests pass without losing an enquiry.
03 / Add the optional AI step
Use redacted examples. Compare summaries with the originals, test unsupported claims and provide a no-AI fallback.
Ready when: A reviewer can see and correct the draft; AI failure does not prevent CRM capture.
04 / Pilot and reconcile
Pilot with a small approved cohort. Review missing handoffs and observed response times before widening access.
Ready when: Receipts reconcile to CRM records and unresolved exceptions have accountable owners.
Measurement
Handoff completeness
Distinct eligible submission IDs with a CRM enquiry and owner / all distinct eligible received submissions, by submission week.
Guardrail: Show invalid, test and unresolved submissions separately. Do not remove missing records from the denominator.
Time to reviewed response
Median and 90th percentile time from receipt to a recorded human response, using the same weekly cohort.
Guardrail: Report unanswered requests separately; do not calculate only on fast completed cases.
Qualified enquiry progression
Count human-reviewed enquiries reaching an agreed qualification stage within a stated observation window.
Guardrail: Document stage definitions and sales-cycle lag. An AI label is not sales qualification.
Failure planning
Duplicate delivery after a timeout
Reuse the submission ID, inspect receipt state and make writes idempotent before retrying.
Free text manipulates the AI step
Treat all visitor text as data. Give the model no CRM write tools, validate output and require human review.
Contact details reach the model
Redact before generation and minimise logs. Establish retention and provider processing terms before production use.
Unbounded AI costs
Enforce token ceilings and an atomic server-side spend reservation before each request. Stop when the allowance is exhausted.
Open questions
- Which CRM fields, API scopes and routing features are available on your current plan?
- Who owns exceptions, and what response-time policy is realistic?
- What constitutes a qualified enquiry, and which cohort window fits the sales cycle?
This tool does not access your accounts, change your systems or verify results. Review any AI-assisted draft before implementation. Send a summary to Akshay only after reviewing and submitting an enquiry.
Continue exploring
Forward-deployed marketing: scope and methods