CHANGES 01
CX becomes agentic
Customers express goals. Systems predict, converse and act to meet them.
Ask about Skip
Prepared answers · no data sent
From attention to judgment — why the next advantage goes to businesses that turn customer context into trusted decisions at the moment of need
2026 · Seventeen original figures; every external number sourced.
A wealth of information creates a poverty of attention.
For twenty years companies competed to capture attention. Attention is now abundant and cheap to buy. What customers cannot get enough of is good judgment — a business that understands what they are trying to do and acts on it correctly, the first time. That is what the intelligence economy prices.
| Dimension | Attention economy | Intelligence economy |
|---|---|---|
| Scarce resource | Customer attention | Trusted, relevant judgment at the moment of need |
| Core asset | Reach, content, audience data | Decision loops that combine context, expertise and AI |
| Customer role | Navigates menus, pages and funnels | States a goal; the system helps complete it |
| Design object | Fixed journeys and interfaces | Adaptive systems that sense, decide, act and learn |
| Innovation cadence | Annual research, quarterly releases | Continuous — every interaction is a signal |
| Primary metric | Engagement, time on site, clicks | Outcomes resolved, decisions corrected, trust extended |
Read the last row twice. Engagement rewards keeping customers busy. Outcomes reward letting them go — with the thing done.
Leaders who own the customer relationship — chief customer, experience, marketing and digital officers — now asked to put intelligence at the center of it.
It offers one operating model (a five-step loop), a design framework for AI-first journeys, and the trust controls that keep both from backfiring.
Not a model or vendor survey. The winners will not be the companies with the most AI. They will be the ones that build the most useful, trustworthy and improving link between human intent and organizational action.
External figures are attributed where they appear and listed at the back.
of service organizations now use AI agents, up from 39%
Salesforce, 2026executives vs. consumers who say loyalty has grown
PwC, 2025of executives say expectations move faster than they can adapt
PwC, 2025of consumers would lose trust in a brand that mishandles data
PwC, 2025Digital automation handled the predictable: rules, forms, scripts. Intelligence handles the rest — messy, unstructured requests, decided in real time, including problems nobody wrote a rule for.
In the intelligence economy, AI is not a feature bolted onto channels. It is the layer that connects datasets to predictions, and predictions to coordinated action across service, operations, finance and product.
That changes where value sits. McKinsey estimated generative AI could add $2.6–4.4 trillion a year, with roughly three-quarters of it in four functions — customer operations and marketing and sales among them. The customer edge is where the economy is being built.
When customers stop navigating and start delegating, the interface stops being the experience. The decision is the experience. Whoever designs how the business decides now designs how it feels to be its customer.
How should we redesign customer experience and innovation when intelligence — not interfaces — becomes the main layer through which customers engage with us?
Customers express goals. Systems predict, converse and act to meet them.
Every interaction is a signal. Unmet needs surface in days, not survey cycles.
Advantage comes from combining people, context and AI — not owning a model.
An intelligent enterprise runs one loop, thousands of times a day. The speed, quality and trustworthiness of that loop set the quality of both the customer experience and the innovation pipeline.
Capture signals across channels and systems
Resolve identity, history and intent in context
Choose a response within policy and authority limits
Execute across channels, or hand to a person
Measure outcome, capture corrections, feed innovation
Trust controls at every step
| Step | What it takes | Typical failure | What the customer feels |
|---|---|---|---|
| Sense | Consented, connected signals from every touchpoint | Channel silos; signals nobody owns | “I already told you this.” |
| Understand | One customer, one identity, full history in session | Wrong record, missing context | “You don’t know who I am.” |
| Decide | Written policy, clear authority tiers, explainable logic | Opaque or inconsistent decisions | “Why did it do that?” |
| Act | Scoped write access; graceful handoff to people | Dead ends; answers without action | “So nothing actually happened.” |
| Learn | Outcome data, correction capture, a route to product teams | Loop stops at resolution | “Same problem, every time.” |
The loop is only as fast as its slowest step. Most organizations have invested in Decide and Act. The binding constraints are usually Understand and Learn.
Measure the loop, not the model. A better model inside a broken loop produces confident mistakes faster.
72% of service teams name data readiness as a major blocker (Salesforce, 2026). That is a Sense and Understand problem.
Menus, forms and page flows made customers do the translating. In the intelligence economy, the customer states a goal and the system does the translation — then acts.
| Model | What the customer does | Cost |
|---|---|---|
| Navigated | Log in → find policy → edit address → error: quote → call center → repeat story | 6 steps · 2 days |
| Delegated | States the goal — “moving to Austin on the 1st” → system re-rates, explains the change, asks one question → customer approves, undo stays available | 1 turn · 1 min |
Illustrative example. What changed is not the channel or the tone. It is who does the work — and that the customer keeps the final say.
Systems shape the experience around each customer’s context instead of forcing everyone through the same static menus.
Agentic AI holds multi-turn conversations, recalls past interactions and resolves issues without human scripts.
AI triages and resolves the simple work instantly, so people spend their time on complex, emotional situations.
Demand for continuity is clear: 81% expect a conversation to pick up where it left off, and 74% are frustrated by repeating themselves (Zendesk, 2026).
Comfort with delegation is not. 58% of consumers report limited or no comfort using AI to engage with brands, and 86% still call human interaction important (PwC, 2025). They will use AI to track an order (49%) far sooner than to make a payment (29%).
Delegation is earned by risk tier. Start where the stakes are low and the outcome is checkable. Widen autonomy only as evidence and customer trust grow.
People are not afraid of AI. They are afraid of being stuck with it. A visible path to a person is part of the product, not a fallback.
Every conversation holds evidence of an unmet need. In the intelligence economy, the organization can detect it, test a response and ship the fix while the need is still fresh.
| Cycle | How the work moves | Learning rate |
|---|---|---|
| Traditional | Survey and research → business case → build and release → measure | Once a year |
| Intelligence economy | Detect → cluster → test → ship — in weeks, then repeat | Twenty times a year |
Speed is not the point; feedback is. A short cycle means the organization learns from customers twenty times a year instead of once.
Service has long been budgeted as a cost to minimize. When each interaction is sensed, clustered and routed to product owners, service becomes the organization’s richest source of demand insight.
Real-time feedback and behavioral signals let teams spot friction, size an opportunity and test a fix with a small cohort — before a quarterly roadmap would have heard about it.
Adding isolated AI tools does not make an intelligent enterprise. Intelligence has to be designed into operations, finance and product from day one — so the loop can reach the systems that need to change.
| Signal | Becomes |
|---|---|
| Repeated “how do I…” | Product or policy redesign |
| Escalation reasons | Next autonomy candidate |
| Human corrections | Decision-logic updates |
| Requests you can’t serve | New offer or partner |
| Abandoned goals | Journey friction backlog |
Organizations that learn faster make better decisions, which earn more delegation, which produces more signal. The loop feeds itself — or starves.
of customer-facing AI deployments fully scale — 3.5× the rate elsewhere
McKinsey, 2026faster resolution where the process was reinvented around AI
McKinsey, 2025of consumers have stopped using a brand over a poor experience
PwC, 2025Models are converging and widely available. Data is necessary but not sufficient. Durable advantage comes from how an organization combines four ingredients into reliable decision loops.
Judgment, empathy, exceptions and the rules worth writing down.
Hardest to copyIdentity, history, intent and consent — resolved in session.
Policies, inventory, pricing, processes and the systems that act.
Reasoning, language, prediction — often from more than one provider.
Most availableModel choice matters and it changes every quarter. A capability survives those changes. Design decision loops so models can be swapped, compared or combined without redesigning the experience.
The organizations that pull ahead write down their expertise as explicit policy, resolve customer context before they automate, and connect AI to the systems that actually act.
Humans do not leave the loop. They move to the parts machines do poorly: setting policy, handling exceptions, reviewing edge cases and correcting decisions — which is how the loop learns.
77% of leaders say AI adoption is outpacing governance (IBM Institute for Business Value, 2026). A capability without control is a liability at scale.
The loop is the same everywhere. The first moment worth rebuilding, and the place a person must stay, differ by industry. Start where volume is high, the rule is written, and the outcome can be checked.
| Industry | Moment that matters | What the loop does | Human stays for |
|---|---|---|---|
| Insurance | First notice of loss; mid-term policy changes | Captures the claim once, checks coverage, sets expectations and triggers next steps | Complex or contested claims, bodily injury, fraud flags |
| Healthcare | Access, scheduling, benefits and billing questions | Matches patient to the right slot, explains cost and coverage, prepares the visit | Anything clinical; vulnerable patients; consent decisions |
| Financial services | Disputes, payments and life events | Explains a charge, opens a dispute, flags the next best action with the reason why | Credit decisions, hardship, suitability of advice |
| High tech | Onboarding, adoption and support | Detects stalled usage, diagnoses issues from telemetry, fixes or guides in product | Enterprise escalations, renewals and roadmap commitments |
| Professional services | Scoping, status and knowledge requests | Assembles prior work, drafts proposals and reports status across engagements | Advice, pricing and client relationships |
| Industrial & manufacturing | Service, parts and uptime | Predicts failure from sensor data, orders parts and schedules technicians | Safety-critical calls, warranty exceptions, field judgment |
Patterns, not prescriptions. Each row is a starting hypothesis. Score your own candidates on volume, reversibility and whether the policy is written down.
In every industry, the highest-value first loop is information plus one action: explain the status and do the next step. It is where customers feel the change most and risk is lowest.
Regulated industries do not move slower because AI is riskier there. They move slower because the policy that governs the decision is often not written in a form a system can follow. Writing it down is the first project.
Across sectors, the same moments belong with people: vulnerability, power imbalance, irreversible loss and nuanced judgment. Publish the list before the first build.
A high-volume, irreversible decision is the most dangerous thing to automate first — and exactly what an ROI-ranked backlog puts at the top.
Past a point, knowing more about a customer makes the experience worse, not better. In the intelligence economy, trust is infrastructure — not a brand attribute.
90% of consumers will share some personal data for personalization, yet 93% say a brand that mishandles it loses their trust (PwC, 2025).
95% expect an explanation for decisions an AI makes (Zendesk, 2026). Prediction without explanation reads as surveillance.
Use what the customer would expect you to know. Disclose what they would not. Never act on what they would object to.
| Control | What it means in the experience | If it is missing |
|---|---|---|
| 01 Consent | Customers choose what is used, for what, and can change their mind | Personalization feels like surveillance |
| 02 Transparency | Customers know when AI is involved and why it decided | Regulatory exposure; trust erodes quietly |
| 03 Fairness | Decisions are tested for bias across customer groups | Some customers get a worse business |
| 04 Privacy | Data is minimized, protected and scoped to the task | One incident resets years of loyalty |
| 05 Oversight | A named owner can see, pause and reverse what the system does | Nobody can say stop |
| 06 Recourse | Customers can correct, refuse or undo an AI decision — and reach a person | Customers leave instead of complaining |
These are product features, not policies in a drawer. If a customer cannot see or use a control, it does not build trust.
Only 5% of task-specific enterprise AI tools reach production (MIT NANDA, 2025). The reasons are rarely the model. They are data, ethics, integration and ownership.
| Barrier | How it shows up | Design response | Owner |
|---|---|---|---|
| Data privacy | Legal blocks the pilot late; consent is unclear for new uses | Map data to purpose before the build. Minimize, scope and log every use. | Privacy + CX |
| Ethical AI | Inconsistent outcomes across customer groups; no one can explain a decision | Write decision logic as policy. Test for bias. Publish an exclusion list. | Risk + Design |
| Integration | AI answers but cannot act; identity breaks across systems | Resolve identity first. Grant scoped write access, one system at a time. | Engineering |
| Ownership | Pilots with no business owner; success defined as “it works” | Name an accountable owner and an outcome metric before day one. | Executive |
Every barrier is cheaper in week one than in week six. Each is usually discovered after a demo has created expectations the data cannot meet.
of task-specific enterprise AI tools reach production
MIT NANDA, 2025of service teams name data readiness as a major blocker
Salesforce, 2026of agentic AI projects expected to be canceled by end of 2027
Gartner, 2025say AI adoption is outpacing governance
IBM IBV, 2026A system that can only read is an interface project. A system that can write — cancel, refund, reschedule — is an accountability project. It needs identity resolution, scoped credentials, a replayable log and an owner who can turn it off.
Teams that budget for the first and build the second stall in pilot. Price the controls from the start.
Disclosure rules are arriving — including the EU AI Act’s transparency obligations and California’s bot disclosure law. Treat them as requirements for the experience, not a legal review at the end.
Trust controls are a material share of the build. Priced at zero, they get cut — and the program reaches customers without a stop button.
An AI-first journey is not a traditional journey with a chatbot added. It is designed around the customer’s goal and the decisions needed to meet it. Six steps, each with an artifact someone owns.
Start with one journey, not a department
| Step | Key question | Artifact | Done when |
|---|---|---|---|
| 1 Intent | What is the customer trying to get done, in their words? | Intent inventory | Top intents ranked by volume and value |
| 2 Context | What must the system know to act well — and is it reliable? | Context and data readiness map | Identity resolves in every system touched |
| 3 Decisions | Which decisions can AI make alone, with approval, or never? | Decision and authority tiers | Policy written; exclusion list published |
| 4 Handoffs | When does a person step in, and what do they receive? | Handoff design | No customer repeats themselves |
| 5 Trust | How does the customer see, question and reverse what happened? | Trust control spec | All six controls visible in the experience |
| 6 Learning | Where do outcomes and corrections go, and who acts on them? | Loop scorecard and insight route | A product owner receives signals weekly |
The artifacts are the deliverable. A journey that cannot produce all six has not been designed for intelligence — it has been automated and hoped for.
Every intent ends with something done or a clear, owned next step — never with instructions.
Set it per decision, not per product. Raise it only when evidence supports it.
A fast, informed handoff builds more trust than a clever answer.
Engagement metrics were built for the attention economy. The intelligence economy needs measures of decision quality, speed and trust — read together, weekly.
| Measure | What it tells you | Trend to watch | Sample |
|---|---|---|---|
| Outcome resolution rate | Share of intents completed without the customer returning | Up, steadily | 68% |
| Loop latency | Time from signal to action, and from insight to shipped change | Down | 3 weeks |
| Correction rate | Share of AI decisions reversed by people or customers | Low and falling — never zero | 4% |
| Delegation rate | Share of customers who choose the AI path when a person is on offer | Up — the truest trust signal | 54% |
| Insight yield | Product, policy or journey changes sourced from interaction signals | Up | 9 / qtr |
A correction rate of zero is a warning, not a win. It usually means nobody is checking. Deflection is deliberately absent: it names the goal as getting rid of the customer.
Most are here → stage 2
Pick one high-volume journey. Map its intents, resolve its context and stand up the five measures. Let evidence decide the next step.
Put the loop scorecard next to revenue and cost in the executive review. What leadership reads weekly is what the organization builds.
Attention is an input, not an advantage. The scarce resource is trusted judgment at the moment of need.
Measure the loop, not the model. A better model inside a broken loop produces confident mistakes faster.
Trust is infrastructure. Consent, explanation, oversight and recourse are product features with a budget line.
Autonomy is earned per decision. Start where volume is high, the policy is written, and the outcome is reversible.
The loop compounds or starves. Organizations that learn faster earn more delegation, which produces more signal.
Every figure in this paper is original. External data points are attributed where they appear and listed here. Vendor-published data is self-interested and is paired with independent sources where a claim rests on it. Verify regulatory dates against the primary legal text before relying on them.
The economy and the shift
Adoption, scaling and the pilot gap
Customer trust and sentiment
Governance and regulation
Industry views, examples and sample values inside figures are illustrative. Frameworks are working instruments and should be revised against your own evidence.
Where this goes next
They will build the most useful, trustworthy and continuously improving relationship between human intent and organizational action. The conversation worth having first: which customer decisions should intelligence make, how much authority it should have, and what must be true before it gets more.
Skip Vanderburg · CX Strategy, Design & Innovation · skip.vanderburg@gmail.com · skipvanderburg.com
