Don’t Let the Page Whisper to the Agent
Plus: When AI helps customers research, compare, and choose, your website stops being a brochure and starts becoming operating evidence.
Your daily signal on AI and CX — minus the hype.
DCX Stat of the day: Zscaler ThreatLabz tested 26 AI models against malicious web pages designed to trick AI agents. In one payment-scam test, 4 models didn’t handle the attack correctly. In another, 2 models treated a fake DeBank page as legitimate..
In this issue:
→ Customers may arrive after AI chooses
→ Web content becomes a trust boundary
→ Prompt injection hits the buying journey
→ AI concierge moves into complex support
→ Banks get a resilience warning
🔍 DEEP DIVE
Your Customer May Send a Scout First
Customers are starting to send AI into the buying journey before they show up themselves.
That agent may research, compare, filter, and recommend before the customer ever reaches a brand’s site.
That changes the job of customer understanding.
Surveys, complaints, reviews, click paths, and sales calls still matter. But if an assistant is shaping the shortlist, the customer journey starts inside a system you don’t fully see. The assistant may read product pages, support docs, comparison sites, reviews, policies, pricing pages, FAQs, and whatever else looks authoritative enough to answer the customer’s question.
That creates a practical problem. Your content can’t be marketing fog. It has to be clean enough for a machine to parse, specific enough to prove, and current enough to trust.
If the agent gets the wrong answer, the customer still blames the experience.
Fair? Maybe not.
But that’s the job now.
Bottom Line: If AI becomes the customer’s scout, CX leaders need to treat public content, policy pages, proof points, and recovery paths as journey infrastructure.
📬 Copy-Paste Take
Before customers use AI to choose us, we need to inspect what AI can actually see. Which claims are provable? Which policies are clear? Which answers are stale? Which pages contradict the sales promise? If the customer’s assistant builds the expectation, our messy content becomes the handoff.
🧭 OPERATOR PLAYBOOK
Audit the Journey the Agent Reads
Before trying to influence what AI recommends, audit the customer decision path your human buyer and their assistant both inherit.
Start with one high-intent journey: product comparison, renewal, return, booking, eligibility, support escalation, or payment.
Then check four things:
Source truth: Is there one current page for pricing, availability, terms, exclusions, and eligibility?
Proof: Are customer claims backed by specific evidence, or by soft positioning that sounds good until someone asks for details?
Policy clarity: Are return, cancellation, privacy, escalation, and complaint rules written like someone might need them under stress?
Recovery ownership: Is there a named human path when an AI-shaped expectation is wrong?
Then test whether an assistant can answer a real customer question without inventing the connective tissue you forgot to provide.
Ask your team: If a buyer asked an AI agent to compare us with two competitors, which page would make us sound trustworthy, and which page would make us sound slippery?
Signal: The best AI work here may look like boring CX hygiene. Annoying. Also where the money is.
📊 MARKET REALITY CHECK
Hidden Instructions Are Now Journey Design
A malicious page can now talk to the agent without talking to the customer.
That’s the nasty customer-journey wrinkle in Zscaler’s research. Web content can carry instructions meant for agents, not people. Malicious sites used search poisoning and hidden page elements to influence AI agents that were trying to complete tasks.
Across 26 LLMs, 4 failed to respond correctly in a payment-scam test, and 2 misclassified a fraudulent DeBank impersonation page.
That doesn’t prove every shopping or service assistant is fragile. It does prove the trust boundary has moved onto the page.
Why it matters: If agents start browsing, comparing, recommending, booking, or paying for customers, CX teams inherit a security and content-quality problem. The journey now includes what the agent reads, what it ignores, what it trusts, and who fixes the outcome when a hidden instruction bends the decision.
Bad content + delegated action = customer-impacting risk.
🧰 TOOL WORTH KNOWING
Lorikeet
What it does: Lorikeet is an AI customer concierge for complex support, built to help companies resolve complicated cases across channels and around the clock.
CX use case: Useful for teams where support work isn’t a simple FAQ lane: insurance claims, healthtech coordination, fintech account questions, travel changes, telecom troubleshooting, or any service moment where the customer needs a real resolution across systems and policies.
Worth watching because: It’s a good test of the next support promise. If AI is going to shape customer expectations before the customer reaches you, the support layer has to do more than deflect tickets. It has to understand the case, use the right source of truth, take the right next step, and know when a human needs to own the outcome.
Bottom line: The useful question isn’t “Can AI answer?” It’s “Can AI resolve the messy part without making the customer pay for the mess?”
The DCX AI Today - AI Tool Directory - If you lead a CX team and want a curated shortlist of tools worth evaluating, this is your starting point.
📡 90-SECOND CX RADAR
Bank of England flags AI as an operational-resilience risk
The Bank of England’s Financial Policy Committee warned that rapid frontier AI advances have increased cyber and operational-resilience risks, because bad actors may be able to create shocks and outages at lower cost and larger scale.
Why it matters: For banks and other high-trust services, AI risk shows up as access, outage communication, fraud response, remediation, and who owns the customer when the system fails.
✅ YOUR MOVE
The customer journey is starting earlier than your analytics may show.
It may start when a customer asks an AI assistant what to buy, who to trust, what policy applies, or whether your promise is credible.
Pick one journey this week and walk it as a customer’s AI assistant would.
Check five things:
What sources would the assistant read?
Which answer would it probably give?
Which claim would it struggle to prove?
Which policy would it misunderstand?
Where does a human recover the relationship?
If an AI assistant is going to carry your promise into the room, make sure it isn’t carrying a blurry copy.
Until tomorrow,
👥 Share This Issue
Think of one person who’s wrestling with AI in CX right now
and forward this to them.
I’m obsessed with Wispr Flow Pro! Get a Free Month on me.
If someone forwarded this to you, they thought you needed to see it before your next AI planning meeting. Get your own copy.