Direct answer: use rules to remove channels that are not permitted, feasible, or operationally supported, then use AI only to rank the remaining options. Combine declared preference, recent reachability, inquiry context, time, cost, team capacity, and likelihood of a useful conversation. Optimize cost per useful conversation and mature commercial outcome; opens, reads, and technical connections are intermediate states.
This is a commercial problem because many companies buy a multichannel stack while first contact is still chosen by seller habit, a sequence default, or the channel with the most attractive open rate. The result mixes permission, availability, and selection: a channel looks better because it received easier leads.
Write a First-Contact Channel Contract
- Eligibility — purpose, applicable authority, opt-in or opt-out, preference, market, hours, and channel policy.
- Addressability — valid email, usable number, active provider, and sufficiently reliable identity.
- Context — source, request, urgency, language, stage, history, and customer-initiated channel.
- Capacity — queue, SLA, human coverage, cost, and current volume ceiling.
- Choice — initial channel, reason, confidence, fallback, and human-review trigger.
- Outcome — delivered, connected, reply, useful conversation, next step, and mature commercial result.
Preference is not propensity. Preserve a declared choice and its scope. A prior reply is observed evidence. A predicted response is inference. Keep those classes separate using https://makinai.co/insights/en/ask-observe-infer-customer-preferences-ai-crm-personalization and enforce permission with https://makinai.co/insights/en/ai-permission-layer-consent-preferences-crm. Never treat an open or the presence of a phone number as consent.
Use the Reach-to-Conversation Ladder
- Permitted — the purpose may use that contact point.
- Addressable — the channel is configured and the contact is technically usable.
- Attempted — execution has an ID, time, owner, and cost.
- Delivered or answered — provider confirms delivery or technical connection.
- Replied — a human returned contact, not merely a read, IVR, or voicemail.
- Useful conversation — the right person discussed the need and produced a disposition or next step.
A native CRM can already combine email, SMS, phone, conditions, and A/B tests. That is execution capability, not evidence that the right channel was selected. Normalize events as described in https://makinai.co/insights/en/crm-sales-service-evidence-phone-messaging-ai and compare only leads eligible for the same channels. Keep a small exploration cohort or randomize among permitted options; otherwise the model learns the old policy and confuses selection with performance.
A five-week pilot
- Week 1 — choose one offer, queue, two channels, and one market; reconcile 30 days of attempts, delivery, replies, conversations, and opportunities.
- Week 2 — define contract, states, eligibility, preference, costs, capacity, reason codes, and outcome window.
- Week 3 — integrate minimum events, run ranking in shadow mode, and create an exception queue; AI sends nothing autonomously.
- Week 4 — release comparable groups, cap exploration, and monitor opt-outs, delay, workload, quality, and cost.
- Week 5 — compare cost per useful conversation and mature outcome; keep a simple rule, expand, or stop.
Cost, trade-offs, and accountability
Cost rises with channels, providers, countries, identity, consent, delivery events, telephony, numbers, templates, history, real-time needs, experimentation, and human review. A sophisticated ranker cannot compensate for missing outcomes. Marketing owns promise and purpose; privacy validates use; RevOps owns states and rules; sales defines useful conversation and capacity; technology integrates and monitors; analytics owns the comparator; finance confirms mature value.
Measure what changes the economics
- Coverage of eligibility, preference, and usable address by channel.
- Attempt, delivery, connection, reply, and useful conversation with explicit denominators.
- Time to conversation and opportunity, including p50/p95 and after-hours effects.
- Provider and labor cost per useful conversation and mature opportunity.
- Opt-out, complaint, repeated contact, queue, abandonment, and human override.
- Mature revenue or margin by cohort; association does not prove incrementality.
First-channel choice does not replace pressure limits; connect it to cross-channel governance at https://makinai.co/insights/en/ai-cross-channel-contact-pressure-crm-frequency. If your company cannot explain why a lead receives email, a call, or a message first, bring MAKINAI at https://makinai.co/services/en/crm-ecommerce-commerce-transformation one offer, current rules, and 30 days of anonymized events. We can scope a two-channel pilot without replacing the full stack.
Limitation: laws, platform policies, and expectations vary by state, country, sector, purpose, and contact type; this is not legal advice. Higher response can reflect selection, time, or capacity. Preserve comparable groups, honor revocation, and wait for the commercial cycle to mature before attributing revenue to AI.