Direct answer: do not replace workflows with AI until the company can say what automations exist, who owns them, which decision each one executes, what it depends on, and whether it changes a business outcome. Build one evidence register, connect runs to cost, exceptions, and outcomes, then assign one of five dispositions: keep, repair, consolidate, retire, or qualify for AI. AI belongs only where real variability requires interpretation—not where duplicate rules, missing ownership, or weak process design created complexity.
A midsize company often has an accumulation problem rather than an automation shortage. Old nurtures, owner assignment, enrichment, notifications, syncs, tasks, webhooks, and exceptions were built by different teams and agencies. The CRM stays green while contacts enter overlapping flows, integrations consume capacity, and nobody can show which automation changed a sales conversation, opportunity, or customer revenue.
Build the Automation Evidence Register
- Identity — stable name, platform, environment, version, business owner, and technical owner.
- Decision — trigger, eligibility, action, exclusion, frequency, exit, and fallback.
- Dependencies — objects, fields, lists, models, APIs, credentials, channels, and other workflows.
- Operations — volume, success, failure, latency, retries, manual review, and incidents.
- Outcome — CRM state change and subsequent effect on acceptance, meeting, conversion, revenue, retention, or cost.
- Economics — license, usage, maintenance, agency, support, exception labor, and opportunity cost.
- Evidence — logs, version, reconciled sample, review date, and known limitation.
- Disposition — keep, repair, consolidate, retire, or qualify for AI, with an owner and due date.
This is not a spreadsheet of workflow names. It traces a path from trigger to outcome. Power Platform inventory, for example, centralizes flows, owners, connectors, state, and trigger data. Salesforce Flow Trigger Explorer shows flows competing on the same object and trigger, including run order. Start with native discovery. Add a cross-platform layer when the portfolio spans tools, environments, and agencies.
Do not mistake activity for value
No enrollment for 90 days can mean waste, seasonality, contingency coverage, or a product that stopped selling. HubSpot uses that threshold to surface unused workflows, but business context determines retirement. Thousands of successful runs can still update a field that no one consumes. Every automation needs an observable downstream change that justifies its existence.
Turn on and preserve metrics before deciding. HubSpot documents enrollments, errors, branch conversion, and version comparisons, but tracking is not retroactive and aggregate data can lag. Export what policy permits before logs expire. Missing history is uncertainty—not a reason to keep every workflow forever.
Make five decisions instead of inventing a composite score
- Keep — relevant outcome, clear owner, known dependencies, proportional cost, and controlled exceptions.
- Repair — the decision is still needed, but data, rules, observability, security, or operations are failing.
- Consolidate — multiple flows repeat a trigger, audience, action, or integration and can share one policy.
- Retire — the obligation ended, outcome is absent, cost exceeds value, or another flow assumed the function.
- Qualify for AI — language, context, or varied paths exceed simple rules; the action is bounded, evaluated, and reversible.
A workflow is not an AI candidate simply because it has many conditions. Complexity caused by poor naming, duplicate fields, or political exceptions should be removed. AI becomes plausible when inputs are unstructured, taxonomies change, or recommendations must combine evidence. Deterministic controls should still enforce consent, eligibility, limits, write-back, and rollback.
Scope a six-week portfolio project
- Week 1 — choose one domain such as inbound, nurture, or renewal; export inventory and pause new ownerless automations.
- Week 2 — map triggers, owners, dependencies, overlap, and run data; select 15 to 25 priority flows.
- Week 3 — reconcile samples to records and outcomes; measure exceptions, rework, and operating cost.
- Week 4 — assign dispositions, design the future state, and write acceptance and safety cases.
- Week 5 — repair and consolidate, run an AI candidate in shadow mode, and disable selected flows reversibly.
- Week 6 — observe effects, restore when a safety trigger fires, document runbooks, and approve the next tranche.
Retirement is a controlled change. Record dependencies, stop new entries, preserve the approved version and permitted logs, keep an alternative for a defined window, monitor effects, and name the reactivation authority. Deleting a workflow on day one destroys evidence and may break silent dependencies. Use the journey QA method at https://makinai.co/insights/en/ai-test-broken-crm-customer-journeys for the future state and the observability contract at https://makinai.co/insights/en/monitor-ai-marketing-workflow-crm-api-changes for dependencies after release.
Fund the portfolio decision, not the prompt
Cost is driven by platforms, environments, objects, branches, channels, countries, history, owners, permissions, APIs, integrations, and criticality. Separate discovery, repair, consolidation, retirement, monitoring, and any AI component. The first-workflow budget at https://makinai.co/insights/en/first-ai-marketing-workflow-cost-complete-budget separates build, run, and improve. Lower API calls or labor hours are not automatically realized value; finance validates what actually changed.
Hypothetical example: 60 automations include 18 notifications, 12 routing flows, 10 nurtures, eight syncs, and 12 exception flows. The team finds three competing routes, four ownerless flows, and two nurtures with no observable outcome. Consolidation and retirement can reduce incidents before any model is deployed. One AI candidate summarizes free-text requests in shadow mode and advances only if it beats the current rule without increasing review, inappropriate contact, or cost per completed case.
Measure outcomes and operational debt
- Inventory coverage and the share with an owner, dependency, version, and outcome.
- Workflows kept, repaired, consolidated, retired, and qualified for AI.
- Runs, failures, retries, incidents, calls, and cost per completed outcome.
- Maintenance, review, and exception hours by workflow and platform.
- Collisions, duplicate contacts, reopened records, and rollback events.
- Acceptance, meeting, conversion, revenue, retention, or cycle time versus baseline.
- Countermetrics: improper suppression, lost coverage, complaints, risk, and concentrated knowledge.
Marketing and sales define outcomes and priority. RevOps owns the register and portfolio policy. CRM admins know versions, limits, and dependencies. Data and analytics reconcile execution to effects. Technology owns APIs, service identities, and environments. Security and privacy govern access, purpose, and retention. Finance confirms economics. Agencies and integrators deliver documentation and transition; the sponsor resolves conflicts and accepts retirement.
NIST recommends testing before deployment and throughout operation, benchmark comparisons, and uncertainty documentation. Apply that discipline to every AI candidate: compare it with a simple rule, preserve abstention, and test regularly. Portfolio rationalization does not prove causal lift; it makes the next controlled test interpretable.
If your CRM accumulated automations across several programs, bring MAKINAI at https://makinai.co/services/en/crm-ecommerce-commerce-transformation an inventory export, known owners, and three recent incidents. We can bound one domain, build the evidence register, and define the smallest repair, consolidation, retirement, or AI pilot without replacing the whole platform.