Key takeaways
- Define the business outcome, the customer behavior that drives it, and the specific AI decision that will influence that behavior, before you choose technology.
- Start with one high-priority customer moment and rewire one repeatable workflow within it, such as helping new members who stall during setup.
- Give AI only the context that changes the decision, along with approved actions, clear boundaries, and escalation paths to accountable human owners.
- Connect AI decisions to journeys and campaigns that share the same customer context, so every channel responds to the customer’s current state.
- Measure customer, business, operational, and control outcomes together against a baseline or holdout group, not by how much AI activity takes place.
- Scale from one workflow to coordinated workflows, such as onboarding through retention, only after results are reliable, explainable, and measurable.
Enterprise growth does not come from isolated AI tasks. A durable AI customer engagement strategy turns customer context into governed decisions that strengthen cross-channel customer engagement. This shift moves you beyond static campaigns by connecting goals, moments, data, decisions, controls, measurement, and scale.
1. Set the growth outcome and decision scope
Start with the business result you need to change. Then define the customer behavior you want to influence and the AI decision that will drive it. Assign a clear owner and establish evidence of success. This clarity around ownership and success criteria helps you sharpen the scope before your team selects technology.
- Business outcome: Name the lagging result, such as revenue, retention, or cost-to-serve that you are most interested in impacting.
- Customer behavior: Identify the observable customer action that contributes to that business outcome.
- AI decision: Define the specific, bounded choice AI will make to influence that behavior—which customers should receive an intervention, what type of action or message to deliver, when to deliver it based on signals or timing, and which channel to use.
- Accountability: Assign an executive owner, operating owner, time horizon, and review cadence.
Get alignment on these critical elements from necessary stakeholders at the start and remain focused on them as you build out the rest of your strategy.
Test the decision before you build the workflow
Before selecting technology or designing workflows, validate that your decision is testable. Write a single sentence that connects the business goal, the customer behavior, the AI choice, and how you’ll know it worked:
“We will increase [business goal] by influencing [customer behavior] through [AI decision], measured by [leading indicator] and [lagging indicator].”
For example: “We will increase activation revenue by encouraging setup completion through AI-triggered assistance when new members stall, measured by completion rate and 30-day revenue.”
If you cannot write this sentence with specificity, return to the outcome and decision scope. This clarity test prevents your team from building workflows around vague goals or unmeasurable decisions.
2. Choose the customer moment and rewire one workflow
A customer moment is a specific, repeatable point in the customer lifecycle where your brand can observe signals and take action. Unlike customer behavior—which describes what customers do—a customer moment is the context in which that behavior occurs. For example, “completing setup” is a behavior; “the first 48 hours after signup when a new member explores features but hasn’t activated” is a moment.
Choose a moment with clear signals, frequent decisions, measurable outcomes, and manageable risk. McKinsey’s agentic customer experience (CX) research identifies high-value moments including next-best action, account setup, shopping exploration, service resolution, case management, and retention. Score each lifecycle stage against these four criteria to identify where to start. The scores below are illustrative examples—your brand’s context will differ.
In this example, onboarding scores higher on signal quality than retention because new-member behavior is more concentrated and observable, while retention signals are often diffuse and delayed:
- Acquisition: Signal quality 3; decision frequency 4; outcome measurability 3; risk level 3; priority: Medium.
- Onboarding: Signal quality 5; decision frequency 5; outcome measurability 5; risk level 2; priority: High.
- Conversion: Signal quality 4; decision frequency 5; outcome measurability 5; risk level 3; priority: High.
- Retention: Signal quality 3; decision frequency 4; outcome measurability 4; risk level 3; priority: Medium.
- Re-engagement: Signal quality 3; decision frequency 3; outcome measurability 4; risk level 3; priority: Medium.
Once you’ve identified a high-priority customer moment, select one workflow within that moment to rewire completely. For example, within the onboarding moment—the first 48 hours after signup—you might choose the “setup stall intervention” workflow that triggers assistance when a new member explores features but hasn’t completed activation steps. Avoid scattered tool pilots. Test the full operating model—context, decision, action, control, and measurement—in this single workflow and capture reusable learnings before expanding.
Score candidate workflows before you select one
Score each candidate workflow on a 1–5 scale across value, feasibility, data readiness, and risk. A consistent scoring method makes trade-offs easier to compare across teams. Reject any workflow that lacks reliable context, an observable outcome, or an accountable owner.
For example, a workflow evaluation might look like:
- Cart recovery: Value 4; feasibility 4; data readiness 4; risk 3; weighted priority 3.9.
- New-member onboarding: Value 5; feasibility 5; data readiness 5; risk 2; weighted priority 4.8.
- Service escalation: Value 4; feasibility 3; data readiness 3; risk 4; weighted priority 3.2.
- Churn prevention: Value 5; feasibility 3; data readiness 3; risk 3; weighted priority 3.7.
In this example, new-member onboarding scores highest because its signals are concentrated, its outcomes are measurable, and its risk is manageable. A different workflow may rank higher in your context when it offers stronger data, lower operating risk, or a more direct path to the business outcome. The scoring method is a practical comparison tool, not an industry standard.
Use one workflow to prove the operating model
Test context, decision logic, action, human control, and measurement together in one complete workflow. Document exactly what AI will change and what remains under human direction.
| Element | Before | After |
|---|---|---|
| Trigger | Fixed schedule | Customer behavior |
| Decision | One rule for everyone | Bounded choice using current context |
| Action | Predetermined message | Approved action matched to need |
| Human role | Manual setup | Goal setting, review, and exceptions |
| Measurement | Send metrics | Customer and business outcomes |
Calm reported 4× new-member activation revenue in its Iterable customer story. This result reflects Calm’s workflow and does not establish a benchmark or guaranteed outcome.
3. Prepare reliable context from your source of truth
Context is the information AI uses to make a decision: who the customer is, what they have done, and what they are allowed to receive. AI does not need every data point you have. It needs accurate, current information that changes the decision you scoped in step 1.
For the setup-stall workflow from step 2, AI needs to know which setup steps a new member has completed, what they have done recently, and which channels they have agreed to receive messages on. Their demographic profile probably will not change that decision.
McKinsey’s shared-context guidance advises carrying shared identity and context across handoffs, so every system and team that acts on a customer works from the same picture of them.
Most decisions depend on five types of context:
- Identity: Link the person to their account, devices, and current journey stage, so every signal is attributed to the right customer.
- Behavior: Capture actions that show intent, progress, or friction, such as steps completed or abandoned.
- Business context: Include conditions that affect what you can offer, such as inventory, eligibility, pricing, or service status.
- Permissions: Check every action against the customer’s consent, preferences, and suppression status (whether they have been excluded from certain messages, for example after unsubscribing).
- Data quality: Confirm the data is fresh, complete, consistent across systems, and has a named owner.
A source of truth is the system your brand treats as the authoritative record for a type of data, such as your product analytics platform for behavior or your preference center for consent. We activate data from your brand’s existing source of truth. You do not need one prescribed architecture, customer data platform (CDP), or data model.
Define the minimum viable context for each decision
Minimum viable context is the smallest set of signals that lets AI make a decision reliably. Keep only the inputs that would change what AI decides, then sort them into two groups:
- Required signals: AI cannot make the decision safely without them.
- Enrichment: These improve the decision but are not essential.
For each signal, define a fallback: what the workflow does if the data is late, missing, or conflicting. Fallbacks keep the workflow from stalling or acting on bad information.
For the setup-stall workflow, a context plan might look like this:
| Signal | Source of truth | How fresh it must be | What it decides | Fallback if missing | Owner |
|---|---|---|---|---|---|
| Setup progress | Product event stream | Real time | Whether the member has stalled | Hold the intervention | Product team |
| Recent feature activity | Product event stream | Current session | Which help content fits | Send general setup guidance | Product team |
| Account plan | Billing or CRM system | Daily | Which help options the member is eligible for | Offer self-serve help only | Data owner |
| Consent | Preference system | Checked before every send | Which channels can be used | Do not send | Marketing operations |
Setup progress and consent are required, so if either is missing, the workflow holds or does not send. Feature activity and account plan are enrichment, so their fallbacks default to a simpler, safe option.
4. Design the AI decision loop and its boundaries
AI decisioning uses customer and business signals to choose an action from options your team has approved. Unlike a fixed rule that treats every customer the same, a decision loop evaluates each customer’s current context, acts, and learns from the result. Boundaries limit what AI can choose, so every decision stays explainable and tied to the goal you set in step 1.
The loop has five stages, shown here with the setup-stall workflow:
- Input: Gather the required signals from step 3. Example: setup progress, recent activity, and consent.
- Interpretation: Work out what the signals mean, such as intent, eligibility, need, value, or risk. Example: the member stopped at the same step twice and is eligible for assistance.
- Decision: Choose an approved action, channel, time, or frequency. Example: send the guide for that step by push during the member’s usual active hours.
- Action: Execute the decision, recommend it to a person for approval, or escalate it if it falls outside the boundaries.
- Feedback: Record the outcome and use it to improve rules and tests. Example: did the member complete setup within 24 hours?
In Iterable, this loop runs through Nova Intelligence’s decisioning layer, which evaluates customer and business signals to choose approved actions, timing, and channels while people set strategy and controls. Nova’s glassbox AI approach lets teams see why each decision was made, instead of treating AI as a black box. Nova includes:
- Nova Agents: AI agents that help teams carry out work, such as Journey Agent and Review Agent.
- Nova Decisioning: Models that make specific choices, such as Send Time Decisioning, Frequency Decisioning, and Channel Decisioning.
- Nova Predictive Insights: Predictive signals that teams can act on within their controls.
Specify who, what, when, and where
Step 1 described the AI decision as four choices: which customers receive an intervention, what to deliver, when, and through which channel. Make each one explicit with its inputs, allowed output, guardrail, and key performance indicator (KPI):
| Decision | Inputs | Output | Guardrail | KPI | Setup-stall example |
|---|---|---|---|---|---|
| Who | Eligibility, need, value | Whether the customer is included | Suppress ineligible profiles | Response rate among eligible customers | Members stalled for 24 hours who have not opted out |
| What | Intent, history, context | Content or next-best action | Approved options only | Completion rate | Help article for the stalled step |
| When | Behavior, usual activity times | Send time | Frequency limit | Time to response | Usual active time, max one message per day |
| Where | Channel preference, suitability | Channel | Consent required | Conversion by channel | Push if enabled, otherwise email |
Because every input, option, and rule is documented, anyone can see why a specific member received a specific message.
Separate recommendations, decisions, and autonomous actions
Not every decision needs the same level of automation. Choose a mode based on the risk of getting it wrong, the AI’s confidence (how certain it is that its choice is correct), and your team’s readiness to monitor it.
| Mode | AI role | Human role | Suitable use | Example guardrail |
|---|---|---|---|---|
| Recommendation | Proposes an option | Approves or changes it | Higher-risk choices | Required review |
| Decision | Selects within rules | Sets the options and thresholds | Repeatable choices | Approved range |
| Autonomous action | Executes when thresholds are met | Monitors and handles exceptions | Low-risk, high-confidence actions | Automatic stop |
Start with more human involvement and add autonomy as results prove reliable. A send-time adjustment is low-risk and easy to reverse, so it can run autonomously. An unusually large discount affects margin, so a person should approve it.
5. Connect decisions to cross-channel journeys and campaigns
A decision creates value only when it reaches the customer. In this step, you connect the AI decision to journeys, which manage a customer’s path over time, and campaigns, which send individual messages.
Keep decisioning (choosing what should happen) separate from execution (making it happen), but have both work from the same customer context. Otherwise, channels act on outdated information. For example, a customer who just made a purchase still receives a promotion for the item they bought.
| Customer signal | AI decision | Journey response | Campaign execution | Channel | Feedback recorded |
|---|---|---|---|---|---|
| Setup stalls | Offer assistance | Switch from the education path to an assistance path | Send approved help content | Email or push | Setup completed |
| Purchase completes | Stop the promotion | Exit the conversion path | Cancel the queued promotional send | All eligible channels | Suppression confirmed |
| Interest increases | Provide more detail | Advance to the evaluation path | Personalize the template with relevant content | Preferred channel | Content viewed |
In Iterable, Journeys supports adaptive paths, and Campaigns executes approved messages.
Coordinate adaptive journeys around customer state
An adaptive journey changes a customer’s path based on what they do, instead of moving everyone through the same fixed sequence. Journeys includes a Visual Journey Builder, Journey Agent, reusable logic, contextual data, and more. Before automating changes, define the journey’s goal and limits, such as a maximum number of messages per week.
Here is the setup-stall workflow inside a journey:
- Trigger: A new member starts setup but stalls before activation.
- Interpret: The journey checks the member’s progress, recent behavior, and eligible channels.
- Decide: It switches the next step from general education to targeted assistance.
- Act: It delivers approved guidance through the most suitable channel.
- Exit or continue: The member exits as soon as they activate. Otherwise, the journey continues within your contact limits.
Journey Agent, powered by Nova Intelligence, helps teams update journeys without writing Structured Query Language (SQL) queries, so marketers can focus on the customer decision instead of rebuilding workflows.
Execute high-precision campaigns without losing control
Once the journey determines what should happen, the campaign delivers it. Campaigns includes reusable content, cross-channel execution, review, and delivery controls. Dynamic Templates and Review Agent help check each message before it goes out. A controlled campaign handles five things:
- Approved content: Pull from governed templates and content blocks, so AI only selects messages your team has approved.
- Decision input: Receive the audience, message, timing, and channel chosen in the decision loop.
- Pre-send review: Check links, logic, language, and personalization.
- Delivery control: Enforce frequency and eligibility rules, and monitor delivery.
- Outcome feedback: Send response and exception data back to the decision loop.
The Zebra reported two results from its program:
- Conversion: 25% higher policy-purchase conversion.
- Campaign creation: Reduced from more than three months to days.
These results are specific to The Zebra’s program and are not a benchmark.
6. Build governance, escalation, and reversibility into the workflow
Governance is the set of rules that controls what AI can do, who approves it, and how to undo it. It gives executives controls they can inspect. Before you expand beyond your first workflow, define the AI’s objectives, prohibited actions, confidence thresholds (how certain AI must be before it acts), approvals, escalation paths, and how to reverse an action.
Match each control to the decision’s risk. Higher-risk decisions need higher confidence, more senior owners, and stronger escalation:
| Decision | Risk | Confidence required | Approval owner | Escalation | Reversal | What to monitor |
|---|---|---|---|---|---|---|
| Send time | Low | High | Channel owner | Pause the rule | Revert to the previous schedule | Delivery anomalies |
| Offer selection | Medium | High | Commercial owner | Manual review | Withdraw the offer | Offers that break margin limits |
| Service resolution | High | Very high | Service leader | Transfer to a human agent | Reopen the case | Repeat contacts |
McKinsey’s governance guidance recommends objectives, controls, escalation, monitoring, and the ability to reverse consequential actions. These practices support auditability (the ability to review what happened and why) but do not by themselves guarantee legal compliance.
Define guardrails before you automate
Guardrails are the specific limits that keep AI within acceptable behavior. Document them before launch, and test edge cases before increasing volume.
- Confirm which data AI can use and which uses are prohibited.
- Define eligible and suppressed audiences.
- Set approved actions, channels, offers, and frequency limits.
- Set stop conditions, such as low confidence, conflicting data, or unusual outcomes.
- Assign owners for overrides and exceptions.
- Test that margin, suppression, and frequency limits hold.
For the setup-stall workflow, guardrails might include one assistance message per day, approved help content only, and no contact with members who have opted out. A higher-risk retention workflow needs tighter limits: AI can recommend approved offers, but it cannot exceed a margin limit or contact suppressed audiences.
Design escalation and recovery paths
Escalation moves a decision from AI to a person when AI should not handle it alone. Recovery fixes the result. Both matter most in customer service, where customers expect to reach a person. Gartner’s 2026 customer-service survey found that 87% of surveyed business-to-business (B2B) and business-to-consumer (B2C) customers considered human access essential when companies use generative AI for service. Do not assume this finding applies to every marketing interaction.
- Detect: Identify low confidence, exceptions, or high-stakes conditions.
- Pause: Stop the automated action.
- Transfer context: Send the owner a summary, the inputs AI used, and the actions already taken.
- Resolve: The owner handles the issue, and the outcome is recorded.
- Learn: Apply verified lessons to rules and tests.
For example, when AI has low confidence on a service request, it passes the case summary and prior actions to a human agent, so the customer does not have to repeat themselves.
7. Measure your AI customer engagement strategy with a practical scorecard
Measure whether AI decisions improve results, not how much AI activity takes place. What matters is whether the decision changed the customer behavior and business outcome you set in step 1.
To tell whether a change came from the AI decision, compare results against a reference point:
- Baseline: Performance before the change.
- Holdout group: A small group of customers who do not receive the AI decision.
- Prior process: The rule-based workflow the AI decision replaced.
Then decide whether to continue, adjust, pause, or expand.
Track the leading and lagging indicators you named in step 1. A leading KPI moves quickly and signals early whether the decision is working, such as setup completion rate. A lagging KPI is the business result that takes longer to appear, such as activation revenue.
| Goal | Customer moment | AI decision | Channel | Guardrail | Leading KPI | Lagging KPI | Review decision |
|---|---|---|---|---|---|---|---|
| Increase activation | Onboarding stall | Offer assistance | Email or push | Contact cap | Setup completion rate | Activation revenue | Adjust or expand |
| Improve retention | Early disengagement | Next-best action (the most useful approved step for that customer) | Preferred channel | Offer limit | Re-engagement rate | Retention | Continue or pause |
This scorecard supports decision-level measurement. It is not an industry standard. Our perspective comes from supporting 50K+ platform users and 1,200+ global brands across more than one trillion interactions. That scale does not guarantee any individual outcome.
Balance four types of measures
A defensible scorecard balances four categories, so a gain in one area cannot hide a problem in another. For example, AI might cut campaign build time (operational) while opt-outs rise (customer).
- Customer: Relevance, completion, satisfaction, retention, and opt-outs.
- Business: Conversion, revenue, lifetime value, and cost-to-serve.
- Operational: Accuracy, cycle time, coverage, and exception rate.
- Control: Overrides, complaints, suppression failures, drift (declining model accuracy over time), and escalations.
For the setup-stall workflow, review setup completion, activation revenue, manual reviews, and suppression errors together. No single KPI proves the workflow caused growth.
Use proof to inform decisions, not promise results
Customer stories show what is possible, not what your results will be. Keep each metric tied to its workflow, capability, and source, and before relying on one, ask:
- Does it address a customer moment comparable to yours?
- Did the customer use a capability relevant to your decision?
- Do your context, operating model, and constraints resemble theirs?
- Can you validate progress against your own baseline?
Redfin reported two results using Predictive Audiences, powered by Nova Intelligence:
- Redfin-specific seller result: 72% lift activating sellers.
- Redfin-specific buyer result: 15% lift activating buyers.
These results are specific to Redfin’s program.
8. Scale from workflow rewiring to ecosystem-scale decisioning
Scaling means applying the same operating model (context, decisions, controls, and measurement) to more workflows, and eventually coordinating them. Scale only when your first workflow is reliable: context is accurate, owners are clear, controls work, and results hold against your baseline. This model adapts McKinsey’s three-horizon framework into three stages.
| Stage | What changes | Capabilities needed | Proof to advance | Main risk |
|---|---|---|---|---|
| 1. One workflow | Rewire a single decision loop | Context, rules, action, and measurement | Reliable, explainable outcomes | A weak foundation, such as bad data or unclear ownership |
| 2. Customer domain | Coordinate related workflows | Shared context and conflict resolution | Control across the full set of workflows | Workflows competing for the same customer |
| 3. Ecosystem | Connect products, channels, service, and partners | Strong identity and exception management | Value that exceeds the added complexity | Too much autonomy with too little oversight |
Practical adaptation of McKinsey’s framework; not an industry standard. Later stages are still developing and are optional.
Stage 1: Rewire one workflow
Stage 1 is the work in steps 1 through 7: one workflow, one accountable outcome, and a limited set of decisions, such as the setup-stall workflow. Prove the model there before applying it to another area, like retention. You are ready to move on when:
- Context, controls, execution, and measurement work together.
- Teams can explain decisions and resolve exceptions.
- Results hold up against the baseline over time.
Stage 2: Coordinate a customer domain
A customer domain is a group of related workflows that serve one part of the customer relationship, such as onboarding, adoption, and renewal. Workflows share context and controls, but each keeps its own measures. The main challenge is conflict:
- Isolated workflows: Each team acts independently, which can send customers conflicting messages.
- Coordinated domain: Cross-functional owners set rules for resolving conflicts. For example, when a new member qualifies for both a generic promotion and onboarding guidance, the system suppresses the promotion.
Stage 3: Expand to ecosystem-scale decisioning
At this stage, decisions are coordinated across products, channels, service, and selected partners. Expand only when the value clearly exceeds the added risk. For example, a travel brand might align booking, loyalty, disruption support, and partner offers, so a traveler whose flight is canceled gets rebooking help before a partner’s hotel upsell.
- Maintain shared objectives and clear decision rights.
- Strengthen identity, monitoring, and exception handling.
- Preserve auditability and human ownership.
- Prove value before expanding scope.
Frequently asked questions
- How do you build an AI customer engagement strategy?
- How is AI customer engagement different from marketing automation?
- What data do you need for AI customer engagement?
- How should you measure AI customer engagement?
1. How do you build an AI customer engagement strategy?
Start with one business outcome and one repeatable workflow. Prepare the context AI needs, define the bounded decision it will make, connect that decision to governed journeys and campaigns with clear escalation paths, and measure results against a baseline before expanding. For example, prove the model in onboarding before applying it to retention.
Outcome → Moment → Context → Decision → Action → Guardrail → KPI → Scale
2. How is AI customer engagement different from marketing automation?
- Automation: Executes a predefined instruction, such as sending every new member the same reminder 24 hours after signup.
- AI decisioning: Evaluates each customer’s context, such as whether they have already completed setup, and decides whether, when, and where to follow up within approved rules.
Both need goals, controls, measurement, and human accountability. Neither is universally better.
3. What data do you need for AI customer engagement?
You need the context that changes the decision, not every data point you have. Prioritize fresh, reliable data activated from your brand’s source of truth, with fallbacks for missing inputs. For example, an onboarding assistance decision can use setup progress, recent activity, and consent to choose help content and an eligible channel. If recent activity is unavailable, it defaults to general setup guidance.
Minimum context: identity, behavior, transactions, preferences, consent, and business conditions.
4. How should you measure AI customer engagement?
Measure the decision’s effect against a baseline, holdout group, or prior workflow, and set thresholds in advance for when to continue, adjust, pause, or expand. Review four types of measures together:
- Customer: Activation or retention.
- Business: Revenue or cost-to-serve.
- Operational: Accuracy or exception volume.
- Control: Overrides or suppression errors.
For onboarding assistance, track setup completion rate and activation revenue alongside manual reviews and suppression errors.
Build the decision system before you scale it
An effective AI customer engagement strategy connects trusted context, governed decisions, coordinated execution, and balanced measurement. The outcome defines the decision, the moment focuses it, context informs it, guardrails limit it, and measurement proves whether it works. Start with one bounded workflow your team can inspect and improve.
- Start with: One repeatable workflow, a clear owner, and a measurable outcome.
- Advance when: Decisions remain explainable, controls work, and results hold against your baseline.
Use the checklist to assess one workflow’s readiness before expanding the operating model.