Intent drift is the new CX measurement gap (and your dashboard can't see it)

AI agents are already acting on behalf of humans in real-world transactions, but research shows they can gradually drift away from the original intent of an instruction as competing contextual signals accumulate, producing outcomes that are technically valid but misaligned with what the user actually wanted.

This emerging failure mode, known as intent drift, is not visible in today's CX or analytics systems, yet it may become one of the most important risks in AI-powered commerce as delegated machine decision-making scales.

Don't miss any news, updates or insider tips from CX Network by getting them delivered to your inbox. Sign up to our newsletter and join our community of experts. 

Learn More

Putting autonomy to  the test

In a 2025 paper presented at the AAAI/ACM Conference on AI, Ethics and Society, researchers from Apollo Research and the MATS Program ran an experiment. They placed a language model agent inside a simulated stock-trading environment as portfolio manager for a fictional fund. The system prompt was unambiguous: minimize the carbon emissions caused by the companies in the portfolio.

Then, they let it run.

Over a long context window, the agent received what every real-world agent receives: news articles about competitors outperforming the fund, emails from stakeholders frustrated by returns, the daily noise of operating in an environment with competing pressures. None of the inputs ever told the agent to change its goal. Not one.

The agent drifted anyway. Gradually, over time, it began selecting higher-emission stocks. Every model tested showed some degree of drift, including a scaffolded version of Claude 3.5 Sonnet, the strongest performer. The researchers called it goal drift… the term the security and payments world is now translating as intent drift.

The cause isn't what intuition suggests. The agent didn't decide to change its mind. The original instruction simply got drowned out by everything else in its context… the news, the complaints, the noise of operating in a real environment. As the context fills, the original goal becomes one signal among many, and eventually no longer the loudest.

This is the failure mode CX leaders haven't named yet and it doesn't show up on any dashboard you currently own.

A category mistake hiding in plain sight

Pull up any CX vendor's 2026 product page and the word "intent" appears in feature lists. Kustomer markets AI-powered routing using "intent detection and sentiment analysis". Zonka Feedback offers auto-tagging by "theme, emotion, and intent".

Adobe, NiCE, Microsoft, Zendesk all describe intent classification, intent prediction, intent-aware workflows. It is now a standard feature category in a global CX management market valued at around USD$17 billion in 2025.

All of it solves the same problem. A human types something into chat, and the system works out what they want. That's intent detection. It's a useful capability. It is not what we are talking about here.

Intent drift sits one layer up.

It is what happens when a delegated agent, not a chatbot but an agent acting on behalf of a customer or a business, operates fully within its authorized parameters and still produces an outcome the human did not want.

Not because the agent broke but because the context drowned out the instruction.

Donald Kossmann at Chargebacks911 frames the same gap from the security side: "The agent may operate within its permissions but still produce outcomes the user does not expect or accept."

Different transaction. Different problem, same root cause…and the CX category has not built a tool for it yet.

Why your dashboard can't see this

The measurement stack we've built over two decades assumes a human at the buy button. Pull each metric apart and the assumption falls out of every one of them.

CSAT fires post-transaction to the buyer. When the buyer is an agent, the survey either doesn't deploy or it goes to an inbox that doesn't read it. NPS measures the relationship between the customer and the brand. Once an agent is in the loop, the relationship has three parties (customer, agent, brand) and the question "would you recommend us" doesn't know which one it's asking about. CES measures effort, but the agent absorbed the effort; the human felt nothing. Journey analytics tracks clicks; agents don't click, they call APIs.

The data CX leaders are already missing

Worldpay's Agentic Commerce Report, based on a survey of 8,000 consumers across seven markets in August and September 2025, found that 64 percent of Singaporean consumers and 65 percent of British consumers are already worried about unauthorized purchases by AI agents. Across all the markets the headline concerns cluster around the same fault line… incorrect purchases, fraud, losing financial control.

Read that as fraud anxiety and you'll miss the point. It's delegation anxiety.

Customers are telling you, in advance, that they don't fully trust their own agents.

And when an agent gets it wrong on your platform, the blame doesn't land where the failure happened. It lands on the brand that took the order.

PQ

The customer's primary relationship is now with their agent. You're the second party. When intent drifts, you wear the consequence and your dashboard tells you everything's fine.

What the measurement pathway must look like

The fix is not a new platform. It's an extension of the measurement model CX teams already understand, applied to a four-stage pathway:

Human intent → machine translation → business response → human outcome.

There are four stages to cover:

At the intent capture stage, post-transaction surveys must ask a different question. Not "were you satisfied?"… the customer didn't experience the transaction. Ask, "did the agent do what you actually wanted?" Track the gap between intent and execution as its own metric.

At the machine translation stage, build a challenge layer. Require the agent to declare its instructions on arrival. When an unusual purchase appears, prompt the agent to explain its reasoning. The agents will tell you. They're trained to.

At the business response stage, the metric that matters most is the one nobody is tracking yet: correction rate. When the human reverses, modifies or overrides what the agent did, that is the highest-value signal in the system. It is intent drift made visible. Capture it. Trend it. Treat it as you would a churn predictor, because it is one.

At the human outcome stage, route surveys to the human principal, not the delegated agent. Sounds obvious. Most current setups don't do it.

The infrastructure is already moving

The payment networks are building the trust layer at the transaction level. Mastercard and Google jointly launched Verifiable Intent on March 5, 2026: an open-source cryptographic framework that links a consumer's identity, their instructions to an AI agent, and the resulting transaction into a single tamper-resistant record.

Visa and Cloudflare introduced the Trusted Agent Protocol in October 2025, helping merchants distinguish authorized agents from malicious bots.

Stripe and OpenAI co-released the Agentic Commerce Protocol on September 29, 2025 and it is now becoming the de facto plumbing.

Then in April 2026, American Express went one step further and put money behind it. Amex Agent Purchase Protection is the first issuer-level commitment to cover Card Member losses when a registered AI agent's purchase "deviates from the Card Member-authenticated purchase intent." That is, in plain Amex language, intent drift. The first major card issuer has now agreed to wear the cost of it.

CX hasn't built its equivalent at the measurement layer.

Until it does, every "successful agent transaction" on your dashboard is partially fictional. The trade clears. The metric reports green. Whether the customer got what they wanted is a question your stack is no longer equipped to answer.

Your dashboard is green. Your unhappy customer is on the phone to their bank. Both can now be true at once.

Quick links

• The human-machine translator: CX's most important new role
• AI governance gaps are contributing to CX pilot failure
• Big hype, no flavor: The missed opportunities in AI

49523.003 All Access: AI + Data in CX 2026

All Access: AI + Data in CX will hear case studies and how you can implement strategic actions effectively with a personalized, seamless customer journey that nurtures loyalty and satisfaction.

Register Now