Next Best Experience: How AI Can Orchestrate Every Customer Interaction, Predict Customer Needs, and Turn Fragmented Touchpoints into Long-Term Relationships

Executive Summary

A next best experience system uses data, predictive analytics, machine learning, generative AI, and journey orchestration to determine the most useful interaction a company should provide to an individual customer at a particular moment.

Traditional marketing asks:

What product should we promote to this customer?

Next best experience asks:

What does this customer most need from us right now? The answer may not be an offer. It may be assistance, reassurance, education, an apology, a service recovery action, a fraud warning, a billing explanation, a product recommendation, or a temporary pause in communication. This distinction matters because customers do not experience companies as separate departments. They experience one brand. Yet many companies operate through disconnected marketing, sales, billing, customer service, loyalty, product, and operations systems.

As a result, one customer may receive:

A promotional email while waiting for a complaint to be resolved An upgrade offer immediately after a service failure A satisfaction survey during an unresolved billing dispute A retention discount after already deciding to cancel Repetitive requests for information the company already possesses Different answers from the website, mobile app, chatbot, and call center A next best experience engine is designed to coordinate these interactions. It can combine signals from customer relationship management systems, transaction platforms, websites, applications, call centers, product telemetry, billing records, loyalty programs, surveys, and operational systems. Predictive models estimate customer needs, risks, preferences, and potential outcomes. A decision engine then selects the most appropriate action, message, channel, timing, and level of human involvement. McKinsey reports that properly implemented next best experience capabilities can potentially increase customer satisfaction by 15 to 20 percent, increase revenue by 5 to 8 percent, and reduce cost to serve by 20 to 30 percent. These figures should not be treated as automatic outcomes, but they illustrate the scale of opportunity available when customer interactions become coordinated and predictive. The most important principle is that next best experience is not simply an AI marketing project. It is an enterprise operating capability.

It requires:

A trusted customer-data foundation Real-time or near-real-time customer signals Predictive and prescriptive models A centralized decisioning and orchestration layer Generative AI for controlled content creation Integration with customer-facing channels Experimentation and measurement systems Privacy, security, fairness, and human-oversight controls Cross-functional governance Continuous learning from customer responses Organizations should begin with one high-value use case, such as preventing churn, resolving billing confusion, reducing service calls, improving onboarding, or recovering dissatisfied customers. They can then expand the system across channels, journeys, and business units. The ultimate goal is not to maximize the number of customer interactions. It is to maximize the usefulness of every interaction.

1. The Customer Experience Problem Is Not a Lack of Communication

Most companies communicate with customers constantly. They send emails, push notifications, text messages, invoices, reminders, surveys, advertisements, product recommendations, loyalty offers, and service updates. They also operate websites, mobile applications, call centers, chatbots, stores, social accounts, and account-management teams. The problem is rarely that the customer hears too little.

The problem is that the communication is often:

Poorly timed Repetitive Inconsistent Irrelevant Uncoordinated Too promotional Disconnected from the customer’s recent experience Optimized for one department rather than the customer relationship The marketing department may be rewarded for conversions. Customer service may be rewarded for reducing handling time. Sales may be rewarded for closing new business. Billing may be rewarded for collecting overdue balances. Product teams may be rewarded for adoption. Loyalty teams may be rewarded for engagement. Each department can be performing well against its own metrics while the customer receives a confusing and frustrating experience. Consider a customer whose internet connection has failed several times in one month.

The company may simultaneously:

Send a premium-service upgrade offer Request a satisfaction survey Promote a referral program Warn about an upcoming payment Recommend a new entertainment package Ask the customer to renew a contract Route the customer through an automated support menu Each message may be technically valid. Collectively, they reveal that the company does not understand the customer’s situation. A next best experience system would recognize the active service problem and suppress most promotional communication. It might instead: Acknowledge the outage. Explain what happened. Provide an estimated resolution time.

Offer a service credit where appropriate. notify a frontline representative if the customer is at high risk of leaving. Resume normal commercial communication only after the issue has been resolved. This is the difference between automated communication and intelligent orchestration.

2. What Is a Next Best Experience?

Next best experience, or NBE, is an AI-enabled capability that determines the most appropriate experience a company should provide to a customer based on the customer’s current circumstances, history, needs, predicted behavior, and potential long-term value.

The experience may include:

Providing information Solving a problem Preventing a problem Making a recommendation Adjusting a service Sending a reminder Escalating to a human Offering compensation Recommending a product Changing the communication channel Delaying or suppressing a message Taking an authorized action automatically

The word “experience” is important. Earlier systems frequently focused on the next best product or next best offer. Their central objective was commercial: What should we sell next?

Next best experience uses a broader objective:

What action will create the best combined outcome for the customer and the company? Sometimes the best action generates immediate revenue. Sometimes it protects future revenue by reducing frustration, resolving an error, preventing churn, or strengthening trust. McKinsey describes next best experience as a capability that allows organizations to deliver the right interaction at the right time and through the right place, rather than relying on indiscriminate promotional outreach. The system is therefore not merely a recommendation engine.

It is a decision system that balances multiple possible outcomes, including:

Customer satisfaction Retention Conversion Revenue Margin Lifetime value Cost to serve Risk Compliance Customer vulnerability Operational capacity Brand trust

A mature implementation may evaluate hundreds of possible actions and choose the one with the highest expected value under established business and ethical constraints.

3. From Next Best Offer to Next Best Experience

The evolution toward next best experience can be understood through five stages. Stage One: Mass Communication Every customer receives approximately the same message.

Examples include:

National television advertising Mass promotional emails Generic discount campaigns Standardized service announcements This approach is easy to operate but produces limited relevance. Stage Two: Customer Segmentation Customers are grouped into categories based on common characteristics.

Examples include:

New customers High-income customers Small-business customers Frequent travelers At-risk subscribers Loyalty-program members Segmentation improves targeting, but every person inside the segment is still treated similarly. Stage Three: Next Best Offer Predictive models estimate which product or promotion a customer is most likely to accept.

Examples include:

A credit-card recommendation A mobile-plan upgrade A streaming bundle A complementary retail product A software subscription upgrade This can increase conversion, but it remains primarily sales-oriented. Stage Four: Next Best Action The system chooses from a wider range of possible actions.

The recommendation could be:

Contact the customer Provide education Resolve a service issue Offer a discount Send a renewal reminder Escalate to a specialist Recommend a product Take no action This stage introduces broader decisioning. Stage Five: Next Best Experience The company coordinates a sequence of interactions across departments and channels.

Instead of selecting one isolated action, the system asks:

What happened before this moment? What is happening now? What is likely to happen next? What does the customer need? Which company objective is relevant? Which channel is appropriate? Should the interaction be automated or human? What should happen after the customer responds? Which other communications should be suppressed? This creates a continuous, adaptive journey rather than a collection of individual campaigns.

4. Why Next Best Experience Is Becoming Possible Now

The idea of personalized customer treatment is not new. Companies have attempted one-to-one marketing for decades. What has changed is the combination of technological capabilities now available.

4.1 Larger and More Diverse Customer Data

Companies can collect information from:

Transactions Website behavior Mobile applications Customer-service conversations Billing platforms Subscription systems Loyalty programs Product usage Connected devices Surveys Social interactions Delivery systems

Fraud and risk platforms The challenge is no longer merely collecting data. It is connecting, governing, interpreting, and activating it. Adobe’s research involving thousands of marketing and customer-experience professionals and consumers found that fragmented data continues to prevent many organizations from delivering real-time, individual-level personalization.

4.2 Better Predictive Analytics

Machine-learning models can estimate:

Churn probability Purchase likelihood Contact-center demand Product affinity Complaint risk Payment difficulty Fraud probability Customer lifetime value Channel preference Propensity to respond Likelihood of accepting an intervention These predictions do not determine the action alone. They provide inputs to a broader decision process.

4.3 Generative AI

Generative AI can transform a recommendation into customer-facing communication.

It can help create:

Personalized explanations Product comparisons Service summaries Follow-up emails In-app messages Chat responses Call-center scripts Account reviews Onboarding guidance The value is not simply producing more content. The value is producing relevant variations quickly enough to support individualized experiences.

4.4 Real-Time Decisioning

Modern event-streaming, cloud, customer-data, and marketing-automation platforms allow companies to respond shortly after an important event occurs.

Examples include:

A declined payment An abandoned application A delivery delay Repeated failed login attempts A sudden decline in product usage An unusually large transaction A negative customer-service interaction A subscription cancellation attempt A major increase in data consumption A customer entering a store or application Real-time interaction matters because relevance deteriorates rapidly. A billing explanation delivered immediately may prevent a complaint. The same explanation delivered two weeks later may have little value.

4.5 Agentic AI

Agentic AI can go beyond generating content by planning and performing sequences of authorized actions.

A customer-service agent could potentially:

Review the customer’s account. Identify the cause of a problem. search approved internal knowledge. Recommend a solution. Apply a permitted credit. update the customer record. Send confirmation. Schedule a follow-up. Escalate unusual cases to a human. Such systems require strict permissions, monitoring, testing, and escalation boundaries. Nevertheless, they move next best experience from message selection toward operational execution.

5. How a Next Best Experience Engine Works

A next best experience engine can be understood as a continuous decision loop. Step 1: Observe The system receives a customer event or detects a change.

Examples:

The customer visited a cancellation page. A payment failed. Product usage declined. A shipment was delayed. A support ticket was opened. The customer completed onboarding. A high-value customer experienced another service failure. Step 2: Understand The system builds context from relevant data.

It may consider:

Customer history Current products Recent interactions Open complaints Lifetime value Channel preferences Consent status Sentiment Risk indicators Operational conditions Eligibility Past responses to similar actions

Step 3: Predict Models estimate possible future outcomes.

For example:

Probability of cancellation Probability of contacting support Probability of accepting an offer Expected revenue Expected service cost Probability of payment recovery Probability of dissatisfaction Likely effectiveness of each channel Step 4: Generate Candidate Actions The engine produces a list of possible interventions.

These may include:

Send educational content Route to a human specialist Provide self-service instructions Offer a temporary discount Recommend a different plan Issue a service credit Request additional information Delay a promotional message Schedule a follow-up Take no action Step 5: Apply Rules and Constraints

The system removes actions that are:

Legally prohibited Inconsistent with consent Inappropriate for the customer Outside operational capacity Financially unjustified Duplicative Conflicting with another journey Potentially discriminatory Too frequent Inconsistent with brand policy Step 6: Rank the Remaining Actions Each permitted action can be scored according to expected impact.

A simplified decision score might consider:

Customer benefit Revenue potential Retention effect Cost Risk urgency confidence Channel suitability The highest-scoring action is not necessarily the most profitable immediate offer. It should represent the best balanced outcome. Step 7: Deliver The experience is delivered through the most suitable channel.

Examples include:

Email SMS Mobile push notification In-app message Website personalization Chatbot Voice assistant Call-center representative Account manager Retail employee Connected device Step 8: Learn

The customer’s response is captured.

The system records:

Whether the customer saw the message Whether the customer responded Which option was selected Whether the issue was resolved Whether satisfaction changed Whether churn occurred Whether revenue increased Whether the model’s prediction was accurate This information improves future decisions.

6. The Four Core Layers of a Next Best Experience Platform

McKinsey identifies data engineering, advanced analytics, generative AI, and campaign delivery as central components of an NBE engine. For implementation purposes, these can be expanded into four interconnected layers. Layer One: Customer Data and Identity The platform needs a reliable representation of the customer.

Relevant systems may include:

Customer relationship management Customer data platforms Data lakes Data warehouses Billing systems E-commerce platforms Loyalty databases Product telemetry Contact-center systems Identity-resolution services The central challenge is identity resolution.

A company may have separate records for:

The customer’s email address Mobile number Website cookie Store loyalty card Application account Service contract Household Business account Without appropriate identity matching, the system cannot confidently coordinate interactions. The goal is not necessarily to create one enormous physical database. It is to create a trusted and governed customer profile that approved systems can use consistently. Layer Two: Intelligence and Prediction This layer estimates customer states and potential outcomes.

Common models include:

Propensity models These estimate the probability of a specific behavior.

Examples:

Purchase Upgrade Renewal Churn Complaint Payment Application completion Customer-value models These estimate the customer’s current or future economic value.

Measures may include:

Historical revenue Expected future revenue Margin Retention probability Cost to serve Referral potential Channel models These estimate which channel is most likely to produce a successful outcome.

A customer may respond best through:

Email Text Mobile application Website Phone Human adviser Sentiment and intent models

These interpret customer language and behavior to identify:

Frustration Confusion urgency purchase intent cancellation intent need for assistance Causal and uplift models

A propensity model asks:

Who is likely to churn?

An uplift model asks:

Whose probability of churning can actually be reduced by this intervention? This difference is critical. Some high-risk customers may leave regardless of the offer. Some low-risk customers may remain without intervention. A company should focus resources on customers whose outcomes can be meaningfully improved. Layer Three: Decisioning and Orchestration This is the central brain of the system.

It combines:

Model predictions Business rules Eligibility criteria Customer consent Contact policies Journey status Operational capacity Strategic priorities Experiment assignments

It then determines:

Whether the company should act What the action should be Which message should be used Which channel should deliver it When it should happen Which other actions should be suppressed Whether human approval is required Without this layer, companies may simply create more personalized messages without fixing organizational fragmentation. Layer Four: Engagement and Execution The chosen action must be activated through customer-facing systems.

These may include:

Marketing-automation platforms CRM systems Contact centers Chatbots Mobile applications Websites Point-of-sale systems Sales tools Service-management platforms Notification infrastructure The decision should also be visible to employees.

For example, when a customer calls, the representative might see:

Recent problems Current sentiment Recommended resolution Approved compensation range Relevant products Actions already taken Communications that should not be repeated This creates continuity between automated and human interactions.

7. The Most Valuable Next Best Experience Use Cases

Companies should not begin by attempting to orchestrate every interaction.

They should select a use case with:

Clear customer pain Strong economic value Accessible data Frequent customer events Measurable outcomes Manageable risk The following use cases are particularly promising.

7.1 Churn Prevention

A subscription company can detect signals such as:

Declining usage Repeated complaints Payment failures Negative sentiment Competitor research Cancellation-page visits Reduced account activity

The system can determine whether the customer needs:

Technical help Education A lower-cost plan A temporary pause A loyalty benefit Human outreach No intervention The objective is not to distribute discounts to every at-risk customer. It is to diagnose the probable cause and offer the most relevant remedy. McKinsey describes a payments organization that used operational, financial, and customer information to predict merchants at risk of reducing business, categorize their underlying issues, and match them with appropriate commercial or service interventions. The company estimated that the capability could reduce annual merchant attrition by as much as 20 percent.

7.2 Proactive Customer Service

The company identifies a problem before the customer contacts support.

Examples:

Shipping delays Service outages Incorrect charges Failed installations Product malfunctions Account-access problems Potential duplicate payments

A proactive system can:

Detect the issue. Notify the customer. Explain the situation. Offer a solution. Estimate resolution. Provide compensation when appropriate. Prevent unnecessary support calls. This can improve satisfaction while reducing the cost to serve.

7.3 Billing and Payment Assistance

Billing confusion is a major source of dissatisfaction.

An NBE engine can distinguish between:

A forgotten payment A declined card A disputed charge A sudden bill increase Financial difficulty A subscription the customer no longer uses A customer who needs a different payment schedule

Possible interventions include:

A simple reminder A billing explanation A payment-plan option A new payment-method request A fee waiver Human assistance A lower-cost plan Treating every case as a collection problem can damage valuable relationships.

7.4 Onboarding

The first days or weeks of a customer relationship often determine long-term adoption.

A next best experience system can monitor:

Setup completion Feature usage Training participation Integration status Login frequency Early errors Support requests Team invitations

It can then recommend:

A tutorial A guided setup A webinar A human onboarding call A product template An integration A simplified workflow This is especially valuable for software companies whose customers may purchase a product but fail to adopt it fully.

7.5 Product and Plan Optimization

Rather than pushing a more expensive plan, the system can evaluate actual customer needs.

For example:

A customer consistently exceeds data limits. Another pays for unused capacity. A business customer is adding employees rapidly. A household has overlapping subscriptions. A banking customer keeps excess cash in a low-yield account. A software customer needs a security feature available in another tier. The recommendation should explain why the change is useful. This turns upselling into advisory service.

7.6 Service Recovery

A service failure creates both risk and opportunity.

The system may consider:

Severity of the failure Customer value Previous failures Customer sentiment Cost of inconvenience Churn risk Available remedies

The experience might include:

An apology A refund A credit Priority support A replacement A loyalty benefit A call from a senior representative A major US airline example described by McKinsey used machine-learning recommendations to differentiate compensation based on customer history and risk rather than giving every passenger the same treatment.

7.7 Customer-Service Agent Assistance

The AI does not need to communicate directly with the customer to create value.

It can assist human representatives by:

Summarizing the customer relationship Identifying the likely issue Retrieving approved knowledge Recommending solutions Drafting responses Predicting escalation risk Identifying retention opportunities Completing administrative work Salesforce’s service research indicates that AI-enabled service organizations are increasingly using agents to handle more cases and free employees for higher-value work.

8. Why “Do Nothing” Must Be a Valid Recommendation

One of the most valuable outcomes from an NBE system may be deciding not to contact the customer. Companies often assume that every available opportunity should trigger a message.

This creates:

Notification fatigue Email unsubscribes Application deletion Reduced trust Promotional blindness Customer irritation A mature system considers the cost of contact.

It should ask:

Has this customer already received several messages? Is there an open complaint? Is another department contacting the customer? Is the predicted benefit meaningful? Could the message create confusion? Is the customer likely to act without intervention? Would waiting produce a better result? Contact suppression should be treated as an intelligent decision rather than a failure to market.

A unified contact policy can establish limits on:

Frequency Channel Message priority Promotional eligibility Complaint-related suppression Quiet hours Vulnerable-customer treatment Consent This is one of the clearest ways next best experience differs from campaign optimization.

9. Generative AI’s Role in Next Best Experience

Generative AI can make individualized customer communication economically practical.

Traditional personalization often inserts a few fields into a template:

Hello [First Name], based on your recent purchase of [Product], you may like [Product B]. Generative AI can create a fuller explanation based on the customer’s situation.

For example:

Your monthly bill increased because your promotional period ended on June 30. Based on your average usage during the past six months, the Standard Plan would provide the same core features and reduce your monthly cost. You can compare both plans before making any change.

This is more useful because it explains:

What changed Why it changed What the company recommends How the recommendation relates to the customer’s actual behavior What the customer can do next However, generative AI should not be allowed to improvise without boundaries.

A production system should generally use:

Approved facts Controlled data access Brand guidelines Legal language Content templates Retrieval from approved knowledge Forbidden-claim lists Output validation Human review for sensitive cases Audit logs The model may determine how to express an approved recommendation, but it should not independently invent prices, benefits, eligibility, policies, or compensation.

10. The Trust Problem

AI makes personalization more powerful, but it can also make customers feel watched, manipulated, or unfairly treated. Salesforce’s consumer research found that advances in AI are making trust more important to customers, while overall trust in companies remains under pressure.

A next best experience may fail when the customer thinks:

How does this company know that? Why am I being treated differently? Is the company exploiting my vulnerability? Was this decision made by a person or an algorithm? Can I correct inaccurate information? Is my conversation being used to train an AI model? Why did another customer receive a better offer? Can I reach a human? Trust must therefore be designed into the system. Principles for Trustworthy Next Best Experience Use data customers reasonably expect the company to use A bank using account activity to detect fraud is expected.

A retailer inferring sensitive personal conditions from unrelated purchasing patterns may create discomfort or harm. Explain consequential recommendations

The company should be able to communicate why:

An offer was made A transaction was blocked A service level changed A case was escalated A recommendation appeared Provide customer control

Customers should be able to:

Manage preferences Change channels Opt out where applicable Correct inaccurate information Request human assistance Appeal important decisions Avoid manipulative optimization An AI system should not exploit customers’ emotional, financial, or cognitive vulnerabilities merely because doing so improves conversion. Protect customer data The Federal Trade Commission has emphasized that AI companies and businesses using AI must honor their privacy and confidentiality commitments rather than quietly expanding data use beyond what customers were promised. Monitor for unequal outcomes Different treatment may sometimes be appropriate, but organizations must evaluate whether models produce unfair or discriminatory results.

Preserve meaningful human escalation Automation should not trap the customer inside an experience that cannot resolve the issue.

11. Responsible AI Governance for Customer Decisioning

The NIST AI Risk Management Framework provides a useful foundation for organizations developing or operating AI systems. It emphasizes incorporating trustworthiness into the design, deployment, use, and evaluation of AI. A practical NBE governance model should include four functions. Govern

Define:

Ownership Accountability Policies Risk appetite Human-oversight requirements Approved data uses Prohibited actions Documentation standards Map

Understand:

The customer use case Affected populations Potential harms Data sources Model dependencies Operational context Legal obligations Failure scenarios Measure

Evaluate:

Accuracy Reliability Fairness Privacy Security Hallucination rates Customer outcomes Model drift Channel effectiveness Employee override patterns Manage

Respond through:

Risk controls Human review Action limits Incident procedures Model retraining Rollback mechanisms Customer appeals Vendor oversight Continuous monitoring The organization should also classify use cases by risk. A product-recommendation email may require relatively lightweight controls. An AI system affecting access to credit, insurance, health services, housing, or essential utilities requires far greater scrutiny.

12. Measuring the Business Impact

Next best experience programs should not be evaluated only through clicks or campaign conversion. They should be measured through customer, financial, operational, and risk outcomes. Customer Metrics Customer satisfaction Net Promoter Score Customer effort Complaint rate Resolution rate Repeat-contact rate Retention Product adoption Digital self-service success

Trust indicators Financial Metrics Incremental revenue Customer lifetime value Churn reduction Cross-sell revenue Upsell revenue Margin Cost per interaction Cost to serve Payment recovery Return on investment

Operational Metrics Average handling time First-contact resolution Number of avoidable contacts Employee productivity Escalation frequency Automation success Decision latency Delivery failure rate AI Metrics Prediction accuracy Precision and recall

Calibration False-positive rate False-negative rate Recommendation acceptance Employee override rate Model drift Content-error rate Unsupported-claim rate Trust and Risk Metrics Opt-out rate Privacy complaints Fairness disparities

Escalation complaints Human-review frequency Incorrect personalization Unauthorized-data exposure Customer appeals The most credible measurement method is controlled experimentation.

A company can compare:

Customers receiving the NBE treatment Customers receiving the existing experience Customers receiving no proactive intervention Different recommended actions Different channels Different message formulations McKinsey recommends using universal control and target groups, alongside ongoing A/B testing and transparent performance dashboards. The organization should measure incremental impact, not merely correlation. A customer who renews after receiving a message may have renewed anyway. The system creates value only when the intervention changes the likely outcome.

13. A Step-by-Step Implementation Roadmap

Phase One: Select the Business Problem Choose one use case with measurable value.

Good starting examples include:

Churn reduction Payment-failure recovery Billing-call reduction Onboarding completion Delivery-delay management Service recovery Subscription renewal

Define:

Customer problem Business problem Target population Possible actions Success metrics Risk level Phase Two: Map the Existing Journey Document every touchpoint.

Identify:

Which teams contact the customer Which systems they use Where delays occur Where messages conflict Which data is missing Which decisions are manual Where customers repeat information Which interactions create avoidable cost Do not automate a broken journey without redesigning it. Phase Three: Establish a Minimum Data Foundation Begin with the data required for the selected use case.

For churn reduction, this might include:

Account status Usage Payments Complaints Support interactions Contract details Product history Engagement

Assess:

Completeness Accuracy Timeliness Consent Ownership Accessibility A narrowly scoped, high-quality dataset is more useful than a massive unreliable one. Phase Four: Create the Initial Models Start with understandable models.

Possible outputs include:

Churn risk Value score Contact likelihood Channel preference Intervention response Compare model performance with simple baselines. The system does not need the most complex algorithm. It needs a model that is sufficiently accurate, stable, explainable, and deployable. Phase Five: Build the Action Library List the actions the company can actually perform.

For each action, define:

Purpose Eligibility Cost Channel Required approval Customer benefit Expected outcome Legal restrictions Operational owner An AI engine cannot create value if the organization has no useful interventions available. Phase Six: Build Decision Rules Combine model scores with business constraints.

Example:

Open complaint: suppress all upsell messages. High churn risk plus unresolved technical issue: route to service recovery. High churn risk plus low usage: offer onboarding help. Payment failure plus strong payment history: send a friendly update request. Repeated payment failure plus hardship signal: offer payment assistance. Low predicted intervention impact: take no action. Phase Seven: Integrate One or Two Channels Avoid launching everywhere at once.

Begin with channels such as:

Email and contact center Mobile application and SMS Website and CRM Ensure the customer receives a coherent experience across the selected channels. Phase Eight: Run a Controlled Pilot

Use:

Target groups Control groups Clear thresholds Human review Defined duration Daily monitoring Incident escalation Track both expected and unexpected outcomes. Phase Nine: Improve the Operating Model

Create a cross-functional team involving:

Customer experience Marketing Sales Service Product Data science Engineering Privacy Legal Security Compliance Finance

Assign one accountable product owner for the customer decisioning capability. Phase Ten: Scale Gradually

Expand by:

Adding customer segments Adding actions Adding channels Adding business units Improving real-time data Introducing generative content Introducing limited agentic execution Creating shared enterprise contact policies McKinsey recommends a two-speed approach in which organizations launch a focused lighthouse use case while building the larger data, technology, and governance foundation in parallel.

14. Common Failure Modes

Failure One: Treating NBE as a Marketing Campaign This limits the system to promotions and ignores service, billing, product, and operational interactions. Failure Two: Buying Technology Before Defining the Decision

A company may purchase a customer-data platform, AI model, or orchestration tool without deciding:

Which customer problem it will solve Which actions it can take Which outcome it will measure Failure Three: Poor Data Quality Incorrect or delayed data produces irrelevant recommendations. Failure Four: No Action Library The model identifies a problem, but the company has no operational response available. Failure Five: Departmental Competition Marketing, service, sales, and product teams continue to optimize independent contact strategies. Failure Six: Overpersonalization The message reveals information in a way that makes the customer uncomfortable. Failure Seven: Optimizing Only for Revenue

The engine repeatedly selects offers while ignoring trust, satisfaction, and service needs. Failure Eight: No Control Group The company reports activity but cannot prove incremental value. Failure Nine: Frontline Rejection Employees do not trust the recommendations or find them difficult to use. McKinsey emphasizes that implementation is not only a technology challenge. Adoption depends on workflow integration, training, enablement, and organizational change. Failure Ten: Excessive Autonomy The organization gives an AI agent permission to perform sensitive actions before establishing reliable monitoring and limits.