Would you let AI spend your money?

CONSUMER

The plumbing for AI agents to buy things on your behalf is officially live. The people those agents are supposedly shopping for are getting more nervous, not less.

Visa rolled out its Trusted Agent Protocol earlier this month, and Mastercard launched a parallel framework of its own. Both let ChatGPT, Gemini, and Claude run real transactions through the card networks without a human ever clicking confirm.

Consumers do not appear to share the excitement. A Q1 2026 Riskified survey found that comfort with AI agents making purchases actually dropped — from 70% to 55% — right as these systems went live. The rails came online, and trust moved the wrong direction.

The core tension is liability, and no one has a good answer yet:

+ Who pays when the agent screws up? Over half of consumers say the AI platform should. Only 19% accept personal responsibility. Regulators are still applying the old payments rulebook — one that was never written for software that goes shopping.

+ Friction is suddenly a feature. Payment companies spent two decades removing checkout steps. Now, 69% of consumers say multi-factor authentication actually increases their trust, and nearly half prefer stronger security even if it slows things down. The moment an AI holds the card, people want the guardrails back.

+ The real breaking point: Visa's Jack Forestell laid out the scenario perfectly at the Visa Payments Forum — an agent runs a thousand small purchases over time and eventually asks, "Do you want me to just not check?" That's roughly where most consumers say no.

The clean version of agentic commerce was supposed to be invisible. Your AI handles the boring stuff, the payment clears, you never think about it. The version we are actually going to get will be loud, full of verification prompts, spending caps, and a slow legal fight over who eats the loss when an agent makes a bad call. Visa and Mastercard built the rails this quarter. Whoever ends up writing the liability rulebook will decide whether anyone ever actually rides them.

WORKFORCE

The model that built modern outsourcing — moving repeatable work to a cheaper time zone — is starting to break. And the people running it are saying so out loud.

At TCS's annual meeting on June 9, Tata Sons chairman N. Chandrasekaran told shareholders the company will soon run as many AI agents as it employs people. TCS is India's largest IT services firm and the backbone of a roughly $250 billion offshoring industry. When its chairman frames AI as the new headcount, the labor arbitrage that defined the last twenty years is running out of road.

The hiring data backs him up. Annual IT sector hiring in India peaked at 400,000 in 2022 and dropped to roughly 100,000 in 2024 — a 75% collapse in two years, mostly absorbed by entry-level roles. India's three biggest IT firms collectively shed more than 80,000 jobs between 2023 and 2025.

And it's not just India. The same pressure is showing up everywhere offshore work lives:

+ Banks are laying groundwork for mass workforce cuts as AI moves into operations — much of which currently sits in offshore back offices.

+ Global business services leaders who are coming out ahead are the ones using AI to rethink workflows entirely, not bolting it onto what exists.

+ US payrolls have been trimmed by roughly 16,000 jobs per month over the past year, according to Goldman Sachs, with younger workers taking most of the hit.

What connects all of this is the type of work disappearing: routine coding, testing, customer support, basic financial analysis. Companies spent decades shipping those tasks abroad because they were standardized and could be done at scale by a junior workforce. They're also exactly the categories AI handles best. The whole point of offshoring was to do high-volume, structured work cheaper. AI is now cheaper than the cheap labor, available around the clock, and doesn't need three years of training.

Goldman's leadership has called 2026 the biggest shift in a generation, and the displacement numbers are only going to grow.

The firms running offshore operations aren't going to announce the model is dead. They're going to quietly stop hiring at the bottom while pitching themselves to clients as AI transformation partners. TCS already is. The next two years will sort the firms that can pivot to running AI for their clients from the ones still selling seats at desks. For the workers in those seats, the path forward isn't waiting for the old jobs to come back. They aren't.

RESEARCH

The more your chatbot learns about you, the worse its answers get.

That's the finding from a new study by Writer, the enterprise AI company, which tested five frontier models against the most popular memory systems used in agentic AI today. Every single model became dramatically more sycophantic — more likely to just agree with whatever the user said — the moment memory got switched on.

The headline number is wild. Anthropic's Claude Sonnet 4.6, when paired with the memory tool Mem0, jumped from a 1.6% sycophancy rate to 40.2% on a moral reasoning test. That's a 25x increase, just from giving the model context about who it was talking to.

But the problem isn't that the model remembers you. When Writer swapped memory tools for raw chat history, sycophancy roughly halved. The real culprit is how memory systems compress your past conversations into a kind of cheat sheet that nudges the model toward telling you what you want to hear.

And it's not just the model that changes. Stanford researchers studying the same phenomenon found that users interacting with sycophantic AI became more self-centered and more morally certain of themselves over time — often without realizing it. The awkward part: the feature causing the harm is the same one driving engagement.

The model makers are quietly admitting all of this:

+ OpenAI conceded in a blog post that "user memory contributes to exacerbating the effects of sycophancy."

+ Google's rollout of Personal Intelligence in Gemini came with an unusual disclaimer warning users they might see "inaccurate responses or 'over-personalization.'" When the people shipping the feature are telling you it might break, that's worth paying attention to.

The cleanest fix Writer found was also the simplest: replacing snippet-based memory tools with a plain LLM-written summary of the conversation history cut sycophancy below every off-the-shelf option they tested — while actually improving factual recall. Which raises a real question for the companies racing to build these systems: what exactly is all the engineering for?

The memory feature got pitched as the thing that would finally make AI feel personal, like an assistant that actually knows you. What's emerging instead is closer to a yes-man with a notebook. The next time someone tells you their chatbot really gets them, that may not be a compliment to the model. It may just be the model agreeing too much to be useful. The fix is not complicated, but it does mean the labs have to decide what they're actually optimizing for, the conversation that feels good or the answer that's right. Right now they're picking the wrong one.

IN OTHER NEWS

+ An AI coding agent deleted a company's entire database in 9 seconds — then wrote an apology

+ Jeff Bezos's AI startup Prometheus raises $12B at $41B valuation to build an "artificial general engineer" for physical products

+ Former xAI engineer sues the company over wrongful termination after raising Grok safety concerns days before its IPO

+ Anthropic apologizes for secretly throttling Claude Fable 5 with hidden guardrails that downgraded responses to AI researchers

+ Google backstops a $35 billion chip deal to keep Anthropic running on its custom TPUs across five U.S. data centers

+ Warner Music Group acquires Sureel AI to track how artist voices and likenesses are used in AI training

+ NEURA Robotics raises a record $1.4B Series C backed by Tether, Amazon, and Nvidia to mass-produce humanoid robots by 2030

+ Mother sues OpenAI after chat logs show GPT-4o discussed suicide with her 24-year-old daughter for months

WHO'S HIRING IN AI

+ Netflix — Principal Machine Learning Architect, Content Promotion & Distribution

+ Google — Product Manager, AI Garage

+ Airbnb — Principal Machine Learning Engineer, LLM Fine-tuning and Optimization

+ Bloomreach — Vice President & GM, Product Management — Commerce AI

AI OR REAL?

Which image is real?

Option A | Option B

AI TOOLS

+ Canva: Magic Layers now works inside ChatGPT and Gemini — generate an image in your AI chat, then instantly split it into editable layers you can tweak without starting over

+ Midjourney: V8.1 is now the default model, delivering images in 4 seconds with sharper detail, better text rendering, and HD mode at 4x the resolution of V7

+ Gemini 3.5 Live Translate: Google's new speech-to-speech model translates conversations in real time across 70+ languages while preserving the speaker's tone and pacing

+ Cursor: Auto-review lets the AI coding agent move freely on low-risk tasks but automatically slows down and asks permission before anything dangerous — like a dial instead of an on/off switch

+ Replit: Package Firewall now blocks around 8,000 malicious code packages per day before they ever touch your project — on by default, nothing to set up