Ethan Bloch runs product for ChatGPT Finances at OpenAI. He founded Digit (the SMS money-saving chatbot, acquired 2021) and a second company that OpenAI acquired in April 2026, and he walks Peter Yang through six ways to actually use the product — then talks about how his team ships.

Why he built it

  • A credit card in his mom’s name at 10, day-trading his bar mitzvah money at 13 (1998), net worth wiped out by 2001.
  • Digit in 2012 was an AI financial chatbot over text message. It helped millions of Americans save, but the tech wasn’t good enough — it kept answering “I’m sorry, I’m afraid I can’t do that.”
  • He thought that dream died when Digit sold in 2021. ChatGPT’s arrival in 2022 convinced him the same product was finally buildable.

1. Make ChatGPT pay for itself

  • The prompt: review the last 90 days for subscriptions — price increases, duplicate charges, refundable fees, unusually expensive household bills.
  • Then escalate: “draft an email to self storage, tell them I’ve been billed twice” and send it. He was double-billed on a hotel points booking, had ChatGPT handle customer service through the browser, and got reimbursed.
  • Peter’s version: he cancelled hotels for his parents, told ChatGPT to schedule a task to monitor replies and keep going back and forth with the hotel until the money came back. It worked.
  • The logic for a $20/month Plus plan is finding $20/month you didn’t know you were losing.

2. Know where the money is

  • Connect accounts through the Finances tab (or just ask in chat). Bank connections run on Plaid.
  • Dashboard: spending by category with month-over-month toggle and drill-downs, net worth, daily movers for investment accounts, upcoming cash flow, bills and subscriptions.
  • Uncertain transactions land in a category-review widget; you can also fix them in chat (“that’s hotels”).
  • A weekly finances update arrives proactively by email and/or push once you sign up.

3. Taxes

  • The prompt: “help me figure out the best way to optimize taxes based on my current situation.” His run took about two minutes.
  • His own pattern for anything long: flip through for the big buckets, then ask what it would actually save, then ask for an image visualizing the options ordered by impact — including what isn’t worth bothering with.
  • Then: “make yourself a to-do list on what I need to do to get these next steps done.” The list doubles as an artifact he and ChatGPT can both return to.
  • His framing: for most people the biggest money is in taxes, not portfolio alpha — it compounds every year and carries no investment risk. “If you’re using ChatGPT to its max, you’re getting abilities that even people with $5, 10, 20 million aren’t getting from their service providers.”
  • For CPAs: have ChatGPT read the accountant’s email, pull the documents, draft the reply, and either send it or hand it over for review.

4. Fund fees

  • Connect brokerage and IRA accounts, then ask: analyze my ETFs and mutual funds — are there replacement funds that could be cheaper?
  • It looks for like-for-like funds with lower fees and estimates annual savings from switching. Worth running for relatives parked in expensive mutual funds.

5. Credit card rewards

  • The prompt: analyze how I’ve been using my credit cards — am I maximizing rewards?
  • It reads transactions, card reward structures, points value, and member benefits (it can spot benefits you never activated because the activation email never arrived).
  • His own case: a 4% reward card that only pays up to $10,000 of spend — he’d been blowing past the cap and losing rewards. Asking for a visualization showed where to move spend, with a roughly $2,900 potential gain on the table.

6. Big decisions as interactive sites

  • “I’m thinking about taking six months off. Build me an interactive site to help me think through this. I want three scenarios.”
  • The output has toggles: a savings floor he won’t go below, scenarios for funding a startup idea, and quick what-ifs (seven months? eight? nine?). It’s built to sit down with a spouse and argue with, not to read.
  • Same shape for rent-vs-buy, and he’s had it build a “when can I fat FIRE” dashboard. You can paste the link to the actual house you’re considering.

Privacy and hard limits

  • Plaid handles the bank login; credentials never pass to ChatGPT, and ChatGPT never receives routing or account numbers. The connection is read-only — it cannot move money.
  • Raw transactions and balances are never used for training. Caveat he volunteers: what enters a conversation can be, so consumers can turn training off in settings.
  • Categorization is the best he’s seen out of the box, but one-offs still miss (Peter’s $4,000 mattress order got filed as an Italian restaurant).
  • Most US financial firms block agents at some step for good reason, and OpenAI is working to whitelist ChatGPT in some cases — so browser-use investing is hit or miss.
  • Finances is US-only today; elsewhere you can still connect Google Drive and email and lean on model intelligence plus web search.

Scheduled work beyond the weekly update

  • The app auto-creates one weekly finances update on signup because there’s a cap on scheduled tasks; more is up to you.
  • His own running jobs: a monthly subscription check, a monthly spend-and-investments email to him and his wife (they reply on that thread), and a cash monitor that flags idle cash — a forgotten employer account, or a few thousand too much sitting in checking instead of savings at 3.5%.

How the team ships

  • 47 shipped items in the last seven days from a small team — mostly small fixes, latency, polish, plus features broken into shippable pieces.
  • Collaboration moved higher up the stack: not arguing over a color or writing a seven-page PRD, but working the genuinely tricky next feature. He calls it coarser-grain collaboration.
  • The talk-to-work ratio shifted hard: “maybe it used to be like 30% talk about the work, 70% do the work… now it’s like 5% talk about the work, 95% do the work.”
  • Artifacts vary: meeting notes into Slack, Linear, strategy docs — all referenceable by Codex. But often there is no written artifact at all, just a quick conversation and the engineer spinning up Codex sessions until a dog-food version is ready to look at together.

For most people, the most money that they could be saving or impacting their life is all in their taxes.

What’s next: nothing announced. Current focus is making existing features easier to find and use, and a voice plugin for finance is in the works.