Direct answer: before using AI to multiply campaigns, subject lines, segments, or send times, implement an Inbox Delivery Contract. It should state who sends through which domain and provider, with what permission, authentication, and message class; decide who is eligible; constrain volume and frequency; ingest bounces, complaints, and unsubscribes; and assign an action and owner. AI can detect anomalies and optimize inside those boundaries. It cannot create permission, authenticate a domain, or turn server acceptance into guaranteed inbox placement.
This is a revenue operating problem, not merely an infrastructure issue. CRM selects the audience, automation builds the journey, an ESP sends, DNS establishes identity, mailbox providers decide treatment, service receives complaints, and analytics sees conversion. When each system knows only its own state, marketing can celebrate more sends while reputation erodes, valid customers miss useful messages, and personalization revenue declines without a visible cause.
Write the Inbox Delivery Contract
- 1. Identity — brand, domain, subdomain, return path, IP or pool, and accountable provider.
- 2. Permission — source, purpose, market, date, preference, and available evidence.
- 3. Authentication — SPF, DKIM, and DMARC aligned with the visible domain.
- 4. Eligibility — reconciled identity, valid address, suppression, frequency, and customer state.
- 5. Message class — transactional, service, relationship, or marketing, each with its own rule and expectation.
- 6. Volume — baseline, cap, ramp, cadence, timing, and kill switch.
- 7. Signals — delivery, response code, bounce, complaint, unsubscribe, reputation, and downstream conversion.
- 8. Action and owner — deliver, reduce, pause, repair, suppress, or review, with evidence and accountability.
The contract is a control layer between CRM and providers, not a standalone dashboard. A person may be eligible for a receipt but not an offer. A domain may authenticate and still carry poor reputation. A message can be accepted and land in spam. A conversion can occur despite a structural fault. Preserve these states instead of reducing them to sent and opened.
Separate hard rules from AI-assisted decisions
Explicit rules should control permission, unsubscribe, suppression, authentication, alignment, message classes, frequency caps, ramps, invalid addresses, complaints, and stop conditions. AI can prioritize investigation, detect an unusual domain shift, recommend timing inside an approved window, and choose among approved content. It should not bypass suppression, repurpose consent, hide poor reputation behind fresh domains, or keep increasing volume during an incident.
Google requires authentication and additional practices for bulk senders, including DMARC and easy unsubscribe for eligible marketing mail, and advises senders to keep reported spam rates low. Yahoo also sets authentication, easy-unsubscribe, and complaint-control expectations for bulk senders. Microsoft published SPF, DKIM, and DMARC requirements for high-volume domains sending to consumer Outlook accounts. Treat published thresholds as changeable operating floors, not comfortable performance targets.
Reconcile identity, permission, and purpose
Map every form, checkout, import, event, partner, and service interaction that creates or updates a contact. Retain source, stated purpose, timestamp, market, and preference version. Remove synthetic, duplicated, and unexplained records before personalization. US federal and state requirements, industry rules, and contracts require their own review; an AI score does not repair a list with no provenance.
Separate transactional and commercial streams when architecture, risk, and customer expectation warrant it, while maintaining coordinated identity and suppression. A necessary receipt should not depend on a newsletter opt-in; an offer should not masquerade as an operational update. New subdomains do not grant permission to ignore aggregate brand reputation or customer experience.
Build a usable feedback loop
Capture send event, campaign and person identifiers, provider, domain, IP or pool where relevant, SMTP response, bounce class, complaint feedback, unsubscribe, and commercial outcome. Normalize codes without deleting the original. Yahoo's Complaint Feedback Loop can return abuse reports for eligible authenticated senders; that signal should reach suppression quickly rather than wait in a monthly report.
Google Postmaster Tools exposes spam, reputation, authentication, delivery error, and compliance dashboards for qualifying traffic. Combine those views with ESP logs and CRM outcomes. Missing dashboard data or insufficient volume does not prove good health. Open pixels are also weak evidence on their own: examine delivery, meaningful clicks, sessions, conversion, and complaints together with appropriate privacy controls.
Run a six-week pilot
- Week 1 — choose one domain, market, provider, and message class; reconcile sources and establish a baseline.
- Week 2 — validate SPF, DKIM, DMARC, links, return path, unsubscribe, preferences, and suppression.
- Week 3 — connect events, normalize failures, and alert by domain, provider, and campaign.
- Week 4 — operate in shadow mode; compare current rules, the proposed contract, and AI recommendations without increasing volume.
- Week 5 — release a limited cohort with a holdout or stable baseline and an effective kill switch.
- Week 6 — evaluate delivery, complaints, conversion, contribution, and operational load before expansion.
The pilot should produce an identity inventory, source map, class matrix, eligibility policy, volume limits, authenticated configuration, response-code catalog, exception view, suppression routine, incident runbook, and impact report. Test invalid addresses, blocks, delays, soft and hard bounces, complaints, unsubscribe, duplicates, provider failover, and campaign shutdown.
Measure the funnel in the right order
- Foundation — records with reconciled source, preference, and identity.
- Authentication — mail covered by aligned SPF, DKIM, and DMARC.
- Eligibility — people remaining after suppression, frequency, and customer-state checks.
- Transport — accepts, rejects, bounces, and response codes by provider.
- Experience — spam, complaints, unsubscribes, and service contacts.
- Outcome — meaningful click, session, conversion, and contribution after delivery.
- Incrementality — difference versus a holdout or comparable baseline, adjusted for seasonality and mix.
Do not optimize solely for opens or attribute every post-send sale to AI. Higher conversion among delivered recipients can coexist with fewer eligible customers, worse reputation, or more complaints. Compare by recipient domain, contact source, message class, provider, market, and cohort; expose small samples and data delays.
Scope cost and accountability
Cost grows with brands, domains, IPs, ESPs, CRMs, countries, data sources, legacy templates, volume, velocity, permission requirements, warm-up needs, monitoring, and incident coverage. The minimum solution may be rules, events, and an exception view. A predictive model is justified only when signals are reliable, volume is sufficient, and the recommendation changes a real decision.
Marketing owns purpose and experience; CRM owns eligibility and pressure; engineering and infrastructure maintain DNS, identity, and events; the ESP executes; service handles complaints; privacy and legal validate uses; analytics designs the comparison; security responds to abuse; and the sponsor can reduce or pause volume. Agencies should not scale sends without the same barriers and signals.
Connect deliverability to the rest of the operation
Control cross-channel pressure with https://makinai.co/insights/en/ai-cross-channel-contact-pressure-crm-frequency, test broken automations with https://makinai.co/insights/en/ai-test-broken-crm-customer-journeys, pause messaging when service reports a customer problem through https://makinai.co/insights/en/service-aware-ai-crm-pause-marketing-support-issue, and return commercial outcomes to CRM through https://makinai.co/insights/en/ai-crm-revenue-feedback-loop-ad-platforms. Deliverability becomes a revenue capability when those states work together.
If CRM reports more sends, the provider reports more failures, and leadership cannot reconcile the business effect, bring MAKINAI an anonymized sample of domains, sources, preferences, events, codes, and outcomes. Through https://makinai.co/services/en/crm-ecommerce-commerce-transformation, we can scope the smallest deliverability control pilot worth integrating before funding more complex personalization.
The sources document provider requirements and tools; they do not guarantee inbox placement or define a universal policy. The Contract, pilot, metrics, and accountability model are MAKINAI editorial recommendations. Requirements change, delivery data arrives with delay, and causality requires a comparator. Permission, privacy, and communication rules must be validated for each market and use.