What ChatGPT Actually Sees When You Link Your Bank Account

OpenAI shipped a personal finance feature on May 15 that lets ChatGPT Pro subscribers connect their bank accounts, brokerage accounts, and credit cards. Reactions split predictably: some people immediately saw the use case, others landed somewhere near the Tom's Guide headline asking what sane person would hand this level of access to OpenAI.

Both are legitimate starting points. Neither is a complete picture of what the connection actually does.

The Plaid layer

The connection doesn't run directly between OpenAI and your bank. It runs through Plaid, the financial data aggregator powering most of the US fintech stack: Venmo, Robinhood, Betterment, and thousands of other apps. When you set up the feature, you authenticate with your bank through Plaid's interface. OpenAI never receives your login credentials.

Plaid then issues a tokenized access grant: a limited, revocable permission slip. You can cancel it through Settings > Apps > Finances in ChatGPT without touching your bank login. This model replaced an older approach called screen scraping, where apps stored your actual credentials and logged in as you on a schedule. Tokenized auth is substantially more controlled, and it's been the industry standard in serious fintech applications for years.

The integration reaches over 12,000 US institutions, including Chase, Fidelity, Schwab, American Express, Capital One, and Robinhood.

Balances yes, transfers no

Once connected, ChatGPT has read access to four data categories: balances, transaction history, investment holdings, and outstanding liabilities. It cannot retrieve full account numbers. It cannot initiate transfers, payments, or any account changes. The data flow is strictly one-directional.

What that read access produces is a consolidated dashboard covering portfolio performance, spending patterns, active subscriptions, and upcoming payments. You can ask natural-language questions against your actual data: "Which subscriptions charged me in April?" or "How has my restaurant spending changed since January?" The answers come from your real transaction history, not from generic advice or hypothetical numbers you've typed in.

OpenAI built this through its April 2026 acquisition of Hiro, a personal finance startup that brought both financial data modeling and the existing Plaid relationship. The reasoning runs on GPT, which at this point means: a capable language model gets access to structured data and does what capable language models do with structured data. Surface patterns. Flag anomalies. Answer questions. The novel piece is live financial data context rather than whatever you paste in manually.

What the launch documentation leaves open

Plaid's tokenization addresses the credential-security concern. It doesn't address the data-use concern.

OpenAI says synced financial data deletes within 30 days of disconnecting your accounts. You can also clear stored financial memories from the Finances settings page before that window closes. These are specific, verifiable controls.

What's ambiguous: whether financial conversation content feeds into model training. OpenAI's standard terms permit using conversation data to improve its models, with an opt-out available. The finance feature's help documentation does not state an explicit exemption from that policy. That matters because OpenAI is currently facing a class action lawsuit alleging it shared ChatGPT conversation data with Google and Meta. The lawsuit hasn't been adjudicated. But the context makes the question of what happens to conversations containing your account data a reasonable one to answer before connecting a brokerage account.

Decrypt published a detailed breakdown of this gap on launch day if you want the full policy analysis.

Who this makes sense for right now

The feature is in preview for US Pro subscribers at $200 per month. OpenAI plans to expand to Plus after the initial feedback phase.

At that price point, the early adopters are almost certainly people already using ChatGPT for serious analytical work who want it operating on real numbers. If you're modeling a major purchase, working out debt payoff order, or trying to understand where your spending has drifted, having live account data is substantially better than manually entering balances. The feature also makes sense for anyone who already uses Plaid-connected apps and understands how the authorization model works.

For casual ChatGPT users, this feature isn't urgent. Worth noting: business adoption data from April showed Claude surpassing ChatGPT in enterprise spend for the first time, suggesting the two tools are diverging along consumer and professional lines. This personal finance feature is squarely a consumer play.

For anyone with ongoing concerns about OpenAI's data practices given the litigation backdrop, waiting for clearer policy language around conversation content is sensible.

The call

Plaid's tokenization makes this more secure than the initial reaction suggests. Read-only access, no account numbers, no ability to make changes: those are real constraints. The open question is whether financial conversation content flows into model training, and OpenAI hasn't explicitly closed that gap in its launch documentation.

Connect it if you have a real use case and you've read the relevant parts of the data policy. Skip it if you don't know why you'd use it.

---

About.chat Weekly covers AI chatbot developments every week. Subscribe for free to get the most useful stories delivered to your inbox.