CRM and AI Operating Model for B2B Marketing Checklist

CRM and AI automation can make an undefined process faster without making it better. Before automating qualification or reporting, use this checklist to make fields, evidence, human responsibility, routing, access, and rollback explicit.

1. State the automation decision

Name the workflow: lead capture, enrichment, qualification, prioritisation, routing, reporting, or renewal signal. Describe the current manual decision and the proposed automated action. Record the owner who can approve, pause, or reverse it.

Define what success means operationally. “Save time” is incomplete. Specify which review is reduced, which error must not increase, which handoff should become more visible, and which outcome remains outside the system’s claim.

2. Sign the CRM dictionary

List every field used by the workflow, its definition, allowed values, source, owner, refresh rule, and disposition when it becomes stale. For stage and qualification fields, use the Salesforce lead implementation guide to prompt explicit ownership, conversion, assignment, and disqualification rules. Local definitions remain authoritative.

Hold if two teams use the same field to mean different things, or if a model will infer a stage that nobody is willing to own.

3. Map inputs and provenance

For each automated input, record source system, timestamp, consent or permission basis, transformation, confidence, and failure behavior. Separate first-party observations from enrichment, inference, free text, and vendor scores. Preserve the original value where a reviewer may need to understand a decision.

If events feed a CRM, define event name, parameters, identity, and retention before implementation. Google Analytics event guidance can structure the record, but an event is not automatically a qualified lead, opportunity, or revenue outcome.

4. Define the human review lane

Write which decisions require a person: high-value account routing, sensitive attributes, rejection, customer-facing claims, pricing, regulated language, or unusual combinations of evidence. Name the reviewer, response time, sample rate, escalation path, and record of the decision.

Do not describe review as “monitoring.” A review lane needs authority to correct a field, override a route, pause automation, and record why the override occurred.

5. Test quality and drift

Create a representative sample containing clear cases, edge cases, missing fields, duplicates, language variation, and recently changed stages. Measure field completeness, disagreement with a human reviewer, false routing, unresolved exceptions, and time to correction. Set a review date and a drift trigger before launch.

For AI-assisted content or decision explanations, apply the IAB AI Transparency and Disclosure Framework as a disclosure and accountability prompt, not as a substitute for local law, contract terms, or internal policy.

6. Govern routing and exceptions

Map the output to the next owner, queue, service level, and fallback. Record what happens when a score is missing, a field conflicts, a duplicate is detected, or a human does not respond. Make the exception queue visible in the same report as the successful routes.

Hold if an automated route can create customer contact without a responsible owner, or if a failed integration silently drops a lead.

7. Separate access from administration

List roles that can view, edit, export, train, configure, and approve. Use least-privilege access where practical. Record vendor subprocessors, retention, deletion, audit logs, and account-recovery ownership. A model that works only through one person’s credentials is not an operating model.

8. Reconcile outcomes

Link the automated decision to downstream disposition: accepted, rejected, nurtured, opportunity, closed, or unknown. Compare the system record with the owner’s decision and the commercial record. Never call a model successful from activity volume alone.

Add an “unknown” state. Forcing every record into a positive or negative outcome creates clean reporting that cannot explain what happened.

9. Check reporting semantics

Write which dashboard questions the workflow is allowed to answer and which it is not. Separate operational health, model quality, human correction, pipeline progression, and financial reconciliation. Give each view a refresh owner and a visible data-lag note. A report that combines today’s automated activity with last quarter’s closed revenue can look precise while comparing incompatible periods.

10. Set automation levels

Choose the least powerful level that can answer the current question: recommend, draft, queue, execute with review, or execute automatically. Advance only when definitions, sample quality, review, routing, access, and outcome reconciliation pass together.

| Gate | Required evidence | Hold if | | — | — | — | | definition | field dictionary and owner | stage meaning differs by team | | input | source, timestamp, consent, provenance | inference replaces the source record | | quality | sample, error and drift rule | only ideal cases were tested | | human control | reviewer, override, escalation | nobody can pause the action | | routing | queue, SLA, fallback, exception | failure can disappear silently | | access | roles, retention, vendor boundary | credentials or deletion are unclear | | outcome | disposition and reconciliation | activity is treated as revenue | | rollback | trigger, owner, restore step | the pilot cannot be reversed |

11. Run a bounded pilot

Use one segment, one workflow, one review period, and a known manual baseline. Keep a parallel sample where a human performs the decision. Record differences, corrections, escalations, and unknown outcomes. Define the stop rule before the first automated action.

At the end, choose continue with the same level, reduce automation, repair the data, expand to a new cohort, or stop. Preserve the model version, prompt or rule set, sample, overrides, and decision note so a later reviewer can reproduce the result.

Keep this checklist local and non-indexable until current source, privacy, legal, technical, overlap, and editorial review are complete. It is a control sequence, not a claim that AI improves qualification, reporting, or revenue in every business.