The direct answer: do not ask AI to decide which system is right. Assign authority by fact, preserve every event, and calculate explicit commercial states—payment, fulfillment, return, subscription, service, and opportunity—before an action is released. Use AI to extract signals from notes, emails, and documents, classify exceptions, and prioritize review. Use auditable rules and transitions for money, consent, entitlement, and suppression.
One person can be active in CRM, have a paid order in commerce, a pending refund at the gateway, and an open service case. No single label answers whether to send an upsell, survey, win-back, or nothing. After resolving who the customer is, the company must agree on what happened, when it happened, and which action remains permitted.
Write the Commercial State Contract
- Object — customer, order, subscription, case, opportunity, location, or relationship.
- Source facts — raw event, identifier, producer, version, and minimum payload.
- Authority by fact — the system that confirms payment, delivery, return, consent, service, or sale.
- State dimensions — never compress financial, fulfillment, relationship, and service into one field.
- Time — when it occurred, when it arrived, and how long it remains correctable.
- Transition — allowed prior states, valid next state, and reversal condition.
- Precedence — which fact wins a conflict and why.
- Idempotency — how a retry or duplicate is prevented from executing twice.
- Correction and replay — how adjustments are recorded, state recalculated, and downstream reproduced.
- Exception — threshold, queue, owner, SLA, and review evidence.
- Downstream permission — the states that release, wait, suppress, or terminate each action.
- Evidence — lineage, rule version, decision, and outcome remain auditable.
The Contract prevents the seductive customer_status field. An order may be paid but unshipped; a return may be requested before the item is received; a subscription may be active while an invoice is past due; an opportunity may be closed-won while provisioning was canceled. Keep dimensions separate and derive only the state required for the decision.
Where rules end and AI begins
- Rules validate transitions, authority, precedence, correction windows, eligibility, consent, and idempotency.
- Integration captures events, retrieves current objects, normalizes keys, and preserves lineage.
- AI extracts a return reason from free text, detects likely resolution in case notes, clusters exceptions, and summarizes context.
- Humans decide material conflicts, changes without a trusted source, financial exceptions, and strategic-account impact.
- Abstention is mandatory when evidence is missing; state remains unknown and a risky action does not run.
- Amount paid, completed refund, valid opt-out, or entitlement must never be invented by a model.
Stripe documents that webhook events can arrive out of order, repeat, and retry; it recommends tracking event IDs and retrieving objects when needed. Shopify exposes separate financial and fulfillment states and treats refunds, returns, and closing as related operations. Design consumers for late, duplicate, and corrective events without sending twice or preserving an obsolete action.
A six-week pilot around one decision
- Week 1 — select one high-friction decision, such as pausing cross-sell during a return or blocking win-back while a subscription remains active.
- Week 2 — write the Contract and sample normal, late, duplicate, out-of-order, and corrected sequences.
- Week 3 — connect sources in shadow mode; preserve facts and compare projected state with the operating system.
- Week 4 — add AI only to unstructured exceptions; require evidence and calibrated confidence.
- Week 5 — release a reversible cohort with human queue, kill switch, and replay.
- Week 6 — compare error, latency, avoided contact, eligible conversion, and cost; choose batch, near-real-time, or stop.
The smallest useful project is not a customer 360. It is one object, one decision, and one critical sequence: return_requested → item_received → refund_pending → refund_succeeded. The action waits until the defined point. If refund_succeeded arrives before item_received because of delay, retrieve current truth and recalculate; do not move the customer backward.
Measure operational truth before uplift
- Quality — correct state by dimension, unresolved conflict, invalid transition, latency, and unknown state.
- Reliability — duplicates absorbed, missing events recovered, successful replay, and repeated actions.
- Operations — exceptions per thousand events, review time, backlog, corrections, and manual hours.
- Customer — contradictory messages, offers during problems, premature surveys, and complaints.
- Business — eligible revenue, conversion after confirmed state, retention, refund leakage, and avoided contact cost.
- Economics — integration, storage, observability, model, review, support, and cost per trusted decision.
First reduce critical false positives: promotion after cancellation, authorization treated as revenue, win-back with an active agreement, or upsell during a complaint. Then estimate incremental value with a holdout or staged rollout. Gross conversion can rise because eligibility changed, not because AI created demand.
Ownership, architecture, and cost
The domain producing a fact owns its semantics; CRM or the data product maintains the Contract; engineering operates ingestion, idempotency, and replay; marketing defines actions; service, sales, and commerce validate exceptions; data measures; privacy and counsel review purpose and retention. Cost grows with sources, volume, cadence, history, versions, corrections, SLAs, destinations, observability, and review. Full event sourcing may be excessive; an append-only fact table plus a versioned projection can be enough.
Connect identity, state, and decision
Resolve who the customer is at https://makinai.co/insights/en/ai-customer-identity-resolution-crm-duplicates, protect campaigns during service at https://makinai.co/insights/en/service-aware-ai-crm-pause-marketing-support-issue, choose latency at https://makinai.co/insights/en/real-time-ai-crm-when-batch-is-enough, and only then orchestrate next-best action at https://makinai.co/insights/en/ai-next-best-action-crm-personalization. MAKINAI structures CRM, commerce, integration, and automation at https://makinai.co/services/en/crm-ecommerce-commerce-transformation. Bring a source diagram and twenty problematic sequences; we can bound a pilot before rewriting the full architecture.
The sources document event behavior, commerce states, state reconstruction, and risk management. They do not prove that a particular model will increase revenue or that AI is necessary. The Contract, pilot, boundaries, and measures are MAKINAI editorial recommendations to validate against systems, accountable owners, data, and local requirements.