BriefTechNews
Anthropic Ties Finance to Its IPO as Visa Proves AI Agents Can't Shop

Anthropic Ties Finance to Its IPO as Visa Proves AI Agents Can't Shop

6 min read · 13 sources

TL;DR
  • Anthropic launched Claude Personal Finance for US Pro, Max, and Ultra subscribers, connecting up to 16 banks via Plaid.
  • Visa's benchmark of 22 AI models on shopping tasks found a 57.2% pass rate, with 54% of failures still ending in a purchase.
  • Enterprise AI deployment spend reached $177B by September 2026, up from $33B a year prior.
  • Revolut completed a fundraising process establishing a $115B valuation with Coatue, T. Rowe Price, and Baillie Gifford.
  • Goldman Sachs restricted Claude Code and GitHub coding tools for bankers but exempted quantitative strats.

Anthropic filed for its Nasdaq IPO under the ticker ANT ten days ago, and today it shipped the product that is supposed to make the subscription story stick: Claude Personal Finance. Paid US subscribers on Pro, Max, and Ultra can now link up to 16 bank accounts through Plaid and ask Claude about their spending, subscriptions, and balances. It is read-only, sees no investments or trades, carries no ads, and Anthropic promises it will not train on finance chats.

The timing is not subtle. Anthropic is going public in a market where only 4.5% of consumers pay for AI assistants, and finance is the most obvious wedge for retention. The bet is that a model that can see your bank balance is harder to cancel than one that just writes code. Nearly two-thirds of US consumers connecting accounts to AI assistants through Plaid had never used a Plaid-powered personal finance app, which suggests Claude is reaching people who never bothered with Mint-style budgeting tools in the first place.

Visa’s benchmark found that 54% of failed AI shopping runs still ended in a purchase recommendation.

Visa's Benchmark Says AI Agents Are Confident, Wrong, and Still Buying

Source: usa.visa.com ↗

Visa released the first industry benchmark for AI shopping agents, and the results are a warning shot for anyone planning to let agents spend money. The test ran 22 models from 12 labs against 15 realistic shopping scenarios, and the overall pass rate was 57.2%. The best models hit 82.2%; the worst scraped 28.9%.

The number that matters is buried deeper: 54% of failed runs still ended in a purchase. The agent did not uncover the customer’s requirements, did not realize it, and recommended a purchase anyway. Models that uncovered more hidden buying factors performed better (70.2% vs. 44.4%), but consistency was terrible - all three attempts passed only 34.2% of the time. Visa plans to extend the benchmark to checkout and order placement, which is where the actual money moves. For anyone building agentic commerce, the takeaway is that the models are not the bottleneck; the failure modes are.

Revolut Is Worth $115B and Still Has a Lending Problem

Source: revolut.com ↗

Revolut closed its fundraising at a $115B valuation, bringing in Coatue, T. Rowe Price, and Baillie Gifford. The amount raised was not disclosed, but employees got another chance to sell shares. The valuation rides on international bank launches and conditional US charter approval.

The gap is on the balance sheet. Reuters reports that lending was just 6% of deposits at year-end 2025. Revolut has the customer base and the regulatory momentum, but it is still basically a payments and FX business that happens to hold deposits. The upside is obvious - 94% of deposits are not being lent out - but the credit risk infrastructure to deploy that capital does not build itself. The $115B price assumes Revolut figures out how to become a real bank.

Enterprise AI Spend Hits $177B and Agents Own the Traffic

Source: menlovc.com ↗

Menlo Ventures’ state of enterprise AI report puts deployment spend at $177B by September 2026, up from $33B a year prior. Agent traffic on OpenRouter is up 37x since overtaking human traffic in February, and agents now generate 80% of token usage.

The concentration is stark. Frontier labs command 69% of deployment dollars, with Anthropic taking 55% of enterprise API spend and OpenAI 31%. Google fell to 9%. Coding ($34.0B) and chatbots ($25.0B) dominate the $77.3B app market. But the most interesting number is that 55% of AI spending is net new - AI is expanding the software economy, not just eating SaaS budgets. For engineers, that means the “AI is a bubble displacing existing spend” argument is not holding up in the data.

Mercury Puts Guardrails on Agents That Touch Your Bank Account

Source: lex.substack.com ↗

Mercury launched controls for AI agents connecting to its business bank accounts, and this is the template every fintech will copy. Agents never get more access than their human counterpart, cannot send payments without human approval, and roles determine who can even connect an agent via Mercury’s MCP server.

This is the boring infrastructure that makes agentic finance possible. Plaid is doing the same job on the consumer side, connecting bank data to Claude’s Personal Finance feature. Visa’s benchmark showed why these guardrails matter - about four in ten AI runs fail - and Nubank published its own testing methodology for support agents. Permissions and governance are becoming the actual product in fintech, not the model.

Remote Bakes Expense Cards Into the Payroll System

Source: prnewswire.com ↗

Remote launched Expense Cards inside its Employer of Record platform, and it kills the reimbursement workflow. Virtual cards go to eligible employees, typically within one business day, and the flow shifts from “pay → approve → reimburse” to “control → pay → review → record.”

The point is that spend data now lives in the same system as payroll and employment data. Admins define who gets a card, what they can buy, and the limits. No separate reconciliation system, no expense report archaeology. For finance teams running distributed workforces, this is the difference between closing the books in days versus weeks. It is only for eligible EOR customers in supported countries, with no platform or per-card charge.

Google Wants the Wallet to Be the Agent's Credit Card

Source: paymentsdive.com ↗

Google’s wallet ambitions go far beyond tap-to-pay. The company is building receipts, order tracking, refunds, and merchant offers into the wallet, plus constrained credentials that let third-party agents pay without exposing private data.

The infrastructure argument is simple: Gemini and Google Pay already exist, so the product experience is the only thing left to build. That puts Google squarely against Stripe’s Link and the agentic commerce protocols from OpenAI, Visa, and Mastercard. The article gives no adoption figures, so treat this as a strategic positioning move rather than a shipped product.

Goldman Bans Claude Code for Bankers, Except the Quants

Source: thenewstack.io ↗

Goldman Sachs restricted Claude Code and Microsoft’s GitHub coding tools for bankers and traders, but made one exception: the quantitative strats keep their access. The bank is weighing productivity against security risk, and the quants won.

Replit CEO Amjad Masad calls this a systemic security risk, pointing at supply chain attacks and credential exposure. Replit published a security advisory recommending AI agents stay off employee endpoints entirely, aligning with a broader shift toward short-lived tokens and cloud-isolated agent execution. For anyone running a financial institution, the lesson is that AI coding agents are a supply chain problem, not just a productivity tool.

BMO Uses Agents to Kill the 14-Minute Call

Source: pymnts.com ↗

BMO launched AI Agent Assist for its contact centre, which handles over 500,000 annual calls about credit card transactions at 12-14 minutes each. The tool targets a 20-30% reduction in resolution time and automatically processes about 40% of eligible credit refund requests, potentially enabling same-day credits.

It is human-in-the-loop by design - AI does transaction matching, policy retrieval, and case preparation, but a person makes the final call. This reflects the industry’s cautious posture: generative AI use for cash flow management surged to 80%, but agentic AI dipped to 5%. Banks want the efficiency, just not the autonomy.

The Rest of the Day

Source: paymentsdive.com ↗

Marqeta CEO Mike Milotich says cards are the strongest customer engagement tool, and the numbers back him: Q3 processing volume rose 29% to $126.5B, revenue up 8% to $176.2M. Marqeta is bundling card issuing with stablecoin-backed cards to win the embedded finance layer.

KCB Tanzania integrated PayPal into its mobile app, giving customers global payments without leaving the banking app. Pliant raised $80M in a Series C led by Seaya Nova and Speedinvest for modular B2B payments. And Fiserv doubled its cost-savings target to $1B, having already cut about 3,000 employees and contractors, while exploring asset sales. The consolidation story in payments infrastructure is not slowing down.

Get the brief

Liked this one? Pick the beats you care about and get just those, five minutes a morning. Free, no hype, one click to leave.

About Me Author

My name is

BriefTechNews

A daily digest of what actually moved in AI, tech, crypto and fintech, assembled and written with AI, with every item linked to its source. Read More
Tags

You May Also Like