Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
AI just BROKE the ENTIRE INDUSTRY...
TL;DR
Wes Roth’s core claim is that finance is the next coding-sized AI wave — he argues the same pattern that telegraphed coding’s takeover is now showing up in finance benchmarks, product launches, and enterprise deals from OpenAI, Anthropic, and Perplexity.
OpenAI is pushing straight into consumer finance with Plaid and 12,000 institutions — ChatGPT Pro users can now connect bank, credit card, retirement, and investment accounts so the model can answer spending, subscription, cash-flow, and portfolio questions like a personal financial OS.
Anthropic is going after Wall Street’s plumbing, not just chatbots — Roth cites a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs plus an FIS partnership touching about 12% of the global economy to build finance-specific agents, starting with anti-money-laundering work.
The important benchmark isn’t perfection, it’s beating humans cheaply — Roth highlights Reed Hoffman’s point that Anthropic’s AML agents were around 64% accurate versus human-only compliance in the mid-70s, arguing that if models move to 75-85% in 12-18 months they become economically dominant.
Roth says AI agents already do real CPA-adjacent work shockingly well — in his own tests, an agent cleaned up messy transaction matching in about 5-10 minutes and automated a workflow he says could otherwise cost a CPA roughly $200-$300 per month.
He ties the finance opportunity to an active cybersecurity crisis — after citing Google’s claimed prevention of an AI-written zero-day attack, Apple vulnerabilities, Vercel’s supply-chain breach, and OpenAI’s npm incident, he warns that AI may both power finance infrastructure and expose it to new systemic risks.
The Breakdown
The setup: coding got hit first, finance is next
Roth opens with a hard swing: AI is about to “kill fintech” and upend finance the way it already started reshaping software. His framing is that frontier labs always telegraph the next target through benchmarks and demos, and after coding, the next obvious battleground is money.
Cybersecurity is already melting down in public
Before he gets to finance, he lays out the chaos already underway: Google reportedly stopped what it called a mass exploitation event involving an AI-written zero-day, researchers warned Apple about Claude Mythos-enabled vulnerabilities, Vercel suffered a supply-chain breach, and OpenAI disclosed fallout from the 10-stack npm attack. The vibe is not theoretical risk — “we’re in the thick of it right now.”
OpenAI turns ChatGPT into a personal finance operating system
The big product news is OpenAI letting ChatGPT Pro users connect accounts through Plaid, with support for roughly 12,000 financial institutions. Roth’s point is that once an AI can see your spending, subscriptions, portfolio performance, and upcoming payments, it stops being a chatbot and starts feeling like the interface layer for your financial life.
His own agent already replaced tedious accounting work
Roth says he’s been doing a version of this himself by exporting financial data into a database and querying it with an agent, though he’s gotten more cautious on security. The memorable anecdote: a messy transaction-matching task he thought needed a CPA got done in minutes, then automated going forward — something he says might otherwise run $200-$300 a month.
Sponsored aside: Ghost as “disposable scratch paper” for agents
In the middle, he detours into a sponsored segment for Ghost, a Postgres database built for AI agents rather than humans. The hook he likes isn’t just the 1TB free tier or unlimited databases, but the mindset shift: let agents spin up, fork, and destroy databases like scratch paper instead of treating one precious database like glass.
Anthropic goes after banks, insurers, and financial crime
Back on track, Roth says Anthropic is attacking finance from the enterprise side with 10 finance-specific agents, plus a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs. He leans on Reed Hoffman’s framing that AI is becoming “the plumbing of the financial system,” especially through Anthropic’s FIS partnership for money-laundering investigations.
The real benchmark: better than humans, cheaper than humans
He spends time on one controversial stat: AI agents being correct in about 64% of AML cases. Roth’s answer to the obvious criticism is that human-only compliance isn’t perfect either — Hoffman places it in the mid-70s — so the real race is whether models can close that gap in the next 12 to 18 months while costing an order of magnitude less.
How the labs are embedding themselves into finance
Roth says OpenAI and Anthropic are copying Palantir’s forward-deployment model by planting engineers inside firms like Goldman Sachs instead of selling from a distance. He piles on signals: CNBC’s Jamie Dimon–Dario Amodei event, Anthropic saying “coding has changed forever, finance is next,” Reuters reporting 40% of Anthropic customers are financial institutions, and PwC planning to certify 30,000 people on Claude.
The Microsoft question, the systemic risk, and who owns the workflow
One of his more provocative detours is investor Christopher Hohn selling most of TCI’s Microsoft stake, which Roth interprets as a bet that AI agents could erode the old Office/Excel workflow. He closes by widening the lens: finance is attractive because it’s structured, high-value, regulated, and deeply embedded, but also dangerous because AI-enabled cyberattacks against banks could trigger panic, liquidity stress, and broader contagion — meaning the real contest is no longer best chatbot, but who controls the world’s most important workflows.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
