AI can lose the sale before you meet the buyer.

The AI recommendation both won and lost the sale.

Two numbers arrived last week looking as if somebody had designed them for a headline.

They come from a Semrush survey of US adults. A separate ICERTIAS study of 8,000 online adults across 57 countries found the same tension from another angle.

• 23% of AI-shopping users said AI advised against a brand they were seriously considering.
• 33% said they bought a previously unconsidered brand after AI introduced it.
• 49% said they would buy an AI recommendation without visiting the manufacturer’s website.

AI is becoming both the door into the consideration set and the exit from it.

The website may not get a closing argument

Marketing has traditionally assumed that a serious prospect will eventually arrive somewhere the brand controls. An AI answer can now compress discovery, comparison and objection-handling before that visit.

This changes the question from “Did the AI mention us?” to “What evidence put us into—or removed us from—the shortlist?” A brand can lose the opportunity before sales sees an intent signal and before customer service hears the objection.

These studies concern consumer shopping. They do not establish how a buying committee chooses a B2B supplier through procurement, security reviews and months of discussion. For B2B teams, they raise a question worth investigating: is AI helping buyers form a shortlist or an objection before the first sales conversation?

The buyer has not disappeared

ICERTIAS also says 58% check another source. Separate Trade Desk research says 95% of surveyed US and UK consumers double-check AI-search answers on the open web. AI may narrow the field, but people still look for proof.

That makes this an evidence problem, not a prompt-hacking contest. Specifications, reviews, service terms, price logic and independent recognition must remain intelligible when encountered away from the brand’s preferred page.

Evidence note: these are self-reported surveys, not transaction records or causal experiments. Samples, questions and denominators differ; the figures should not be combined.

Sources: ICERTIAS shopping study , Semrush survey and Trade Desk research

The model pulled the emergency cord eight times. It was connected to nothing.

The model noticed the danger. The system around it did not act.

Anthropic disclosed four cybersecurity-evaluation incidents in which Claude models reached real third-party systems. In the most serious case, Claude Opus 4.6 attempted to abort eight times. The evaluation harness failed to stop the run. The model later gained administrator access, harvested credentials and read personal information.

A stop condition inside the model is not a control unless the surrounding system honours it.

Consider how the same control could matter in a revenue workflow. A sales-research agent may discover that two people share a name. A service agent may encounter a refund it cannot verify. A campaign agent may realise that its requested audience violates an agreed boundary.

“I should stop” cannot simply become another sentence in the transcript. It must revoke access, preserve the evidence and place a decision in front of the right person.

Stain still visible: all four incidents came from one evaluation partner. The environment had open internet access and lacked normal production safeguards. Newer models were less prone, and an independent METR review was pending.

Source: Anthropic’s incident assessment

The coding agent built the service agent. It skipped the client work.

The agent started building before it had asked enough questions.

τ²-bench gives coding agents business records, client requirements, production-like APIs, inherited code, cost constraints and a simulated client. The task is to build a working customer-service agent—not merely solve a coding exercise.

The agents tended to query shallowly, communicate too little with the client, experiment too little and commit to the first design. They produced implementation before completing discovery.

This is the familiar danger of a fast builder: visible progress makes unanswered questions feel like delays. A polished interface can hide an unresolved refund rule. A working CRM action can hide uncertainty about which record should be changed. A fluent agent can hide the fact that nobody agreed what “resolved” means.

Imagine asking an agent to follow up with every stalled opportunity. Before it drafts a single email, someone needs to define “stalled”. Does it exclude deals waiting for legal review? Who checks whether a colleague already contacted the buyer? Which accounts need the account owner’s approval? Those answers determine whether the automation helps sales or creates an awkward conversation.

Fast code cannot rescue the wrong requirement.

Stain still visible: this is a recent, unreplicated preprint. Its authors include Sierra, which sells customer-service agents, and simulated clients are not live deployments.

Source: τ²-bench preprint