AI Customer Experience Guide: When to Personalize or Hand Off

An AI experience can sound personal while being wrong about the person, the moment, or the business's authority. Start with one customer job and decide whether current context is safe enough to help. If not, use a generic answer, do nothing, or hand the case to a person.

Check five things before personalizing

Copy these checks into a document or spreadsheet for one customer action. TRUST is only an optional memory aid.

Check	Decision
Task	Name the customer's immediate job and acceptable no-action result.
Record	Verify which records may be linked and how uncertainty is represented.
Up-to-date context	Define source, freshness, event order, permission, and service conflicts.
Safe action	Bound content, channel, claims, and consequential commitments.
Transfer	Specify human trigger, context packet, acceptance, and fallback.

Salesforce defines identity resolution as unifying customer data from multiple sources. That vendor definition explains a mechanism; it does not prove that a profile is complete or correctly matched. Review the source.

When personalization is appropriate

Use it for a bounded, useful choice where identity is sufficiently supported, context is current, action is reversible, and correction is easy. Avoid it when records conflict, permission is unclear, a service or safety issue is active, or the decision affects eligibility, legal rights, health, credit, pricing exceptions, or other material outcomes without an authorized person.

NIST's AI RMF supports explicit roles, monitoring, evaluation, and risk response. It does not prescribe a particular customer journey or prove benefit. Review the framework.

How to apply it

1. Choose one customer job, not “personalize everything.”
2. Map the minimum fields and their purpose, provenance, freshness, confidence, and permission.
3. Write the action ladder: generic response, contextual response, clarify, no action, human handoff.
4. Make service cases, opt-outs, corrections, and newer events override commercial actions.
5. Test mismatched identity, shared device, stale event, missing value, conflicting consent, sensitive inference, and handoff timeout.

Worked example

A fictional learning service wants to personalize renewal help. The system can see a completed course and a current support case, but two email records may belong to the same person. It does not merge them automatically or recommend a renewal. The support case suppresses commercial messaging, and an agent asks the customer to confirm the account. After the issue closes, the system offers a generic renewal-options page. No retention effect is claimed.

Risks and controls

Risk	Consequence	Control
False identity match	Another person's context exposed	Confidence threshold and no-merge review
Stale event order	Irrelevant or contradictory action	Freshness and precedence rules
Sensitive inference	Harm or unfair treatment	Approved fields and human decision
Handoff without acceptance	Customer repeats context	Packet, target, fallback owner

How to measure the result

Track context-valid action rate = actions with verified identity, purpose, freshness, permission, and precedence / contextual actions; customer corrections, false-link reviews, service-conflict prevention, accepted handoff, stale-action, opt-out, and no-action rates. The customer-journey owner is accountable; data and service owners supply their quality measures. Review weekly during pilot against a pre-pilot baseline and set pause/expansion thresholds in advance.

What to do next

Use the unified-data personalization use case for one decision, the human-handoff guide for transfer design, and the customer-data guide for the underlying identity contract.

Evidence and limitations

No match accuracy, personalization uplift, legal permission, customer-360 completeness, integration, or Easy AI capability is claimed.