← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

Run This Onboarding
Check Before AI
Touches Anything Sensitive

ChatGPT, Claude and Perplexity can all connect to your bank accounts now. Here’s the exact onboarding check I run before any of them gets near my money.

This spring, the biggest AI apps all shipped ways to plug straight into your bank accounts, credit cards and investments.

  1. ChatGPT Finances. OpenAI launched it on May 15, 2026 for Pro users in the US and has since opened it up to more plans. It links accounts through Plaid.
  2. Claude. Era is a third-party personal finance connector in Claude’s connector directory, under Customize (being renamed Plugins), then Connectors.
  3. Perplexity. On April 9, 2026 it added Plaid so its Computer agent can analyze your bank accounts, cards and loans.

This is fine with the right guardrails and risky without them. Below is the prompt I run before I turn on any sensitive connector, whether that’s a bank, a brokerage, email, a calendar, health records, or anything else that would cost more than time if it went sideways.

Read Vs Act

Two years ago, the biggest AI risk was a wrong answer. Annoying, but recoverable. Now AI can act. Connected tools can send email for you, change your calendar, and agents that control a browser can fill forms and check out.

The bank features above are read-only today. They can see balances and transactions but can’t move money. Some connectors can still make changes, though. Era’s paid tiers let Claude recategorize transactions and create rules, for example. Read access and action access are completely different levels of risk, and this check makes the AI spell out which one you’re agreeing to before you click Approve.

The Prompts

01

Before you connect

Paste this into Claude or ChatGPT with web search on, and fill in what you’re about to connect. You get a one-page check you can actually read.

Onboarding Check
You are my AI onboarding officer. Before I connect an AI tool to something sensitive, write me a one-page onboarding check so I know exactly what I'm agreeing to. The AI feature or connector: [e.g. ChatGPT Finances / the Era connector in Claude / Perplexity's Plaid integration / a custom MCP connector] The account I'm connecting: [e.g. my checking account / my Gmail / my Google Calendar / my QuickBooks] Link to its help page or privacy page, if I have it: [PASTE URL OR "NONE"] Search the web for the provider's official help center, privacy policy and announcement posts. Use those first, and cite them. Then give me this exact structure: 1. WHAT IT CAN READ. The specific data it gets once connected. Be granular: not "transactions" but "every debit and credit, merchant name, amount, date and memo." 2. WHAT IT CAN DO ON MY BEHALF. Every action it can take that changes my account or the world: send, transfer, pay, cancel, schedule, delete, post, share, recategorize. If it's read-only, say "READ-ONLY" in capitals and cite where the provider says so. 3. WHAT GETS STORED, AND FOR HOW LONG. Where the data lives after it's pulled, whether my chats with it are used for training, and how long it's kept. Name the setting that controls each one. 4. THE SAFEST SETUP. Specific settings, permission scopes and habits that shrink my downside. Not "be careful," but things like "pick only the one account you need at the bank login screen" or "set this tool to ask before acting." 5. HOW TO DISCONNECT AND DELETE. The exact steps to revoke access and delete stored data, inside the AI tool AND at the data provider (Plaid, MX, my bank, Google, etc.). 6. 3 REALISTIC WORST CASES. For example a breach at the provider, a prompt hidden in an email or web page that tricks the AI, or an action I didn't expect. For each one, what I'd do to recover. 7. MY CALL. One line: connect now, connect with limits (say which), or skip it. Rules: if the official docs don't say something, write "NOT STATED" instead of guessing. No marketing language. Assume I'm not a security expert, but I'm not clueless either.

Point 2 is the line that matters most. Read-only (it can see your balance) is low risk. Action access (it can move or change things) is a totally different conversation.

02

Right after you connect

The docs tell you what a tool should do. This shows you what it can actually see in your account. Run it in a chat where the connector is turned on.

Scope Check
I just connected [ACCOUNT OR CONNECTOR NAME] to you. Before I use it for anything real, answer these from what you can actually see in this chat, not from general knowledge: 1. List every tool or action this connection gives you, by its exact name, and mark each one READ or WRITE. 2. For each WRITE action, tell me whether you would ask me before using it, or whether it can run on its own. 3. Show me a small sample of what you can read right now (for example the last 3 transactions, with amounts but no full account numbers) so I can confirm the scope matches what I approved. 4. Tell me anything you can see that I probably didn't mean to share. Then follow this rule for the rest of our chats: never take a WRITE action on this connection without showing me exactly what you're about to do and waiting for me to type "yes."
When To Run It

01

Banks and brokerages

Any account with a balance, including anything linked through Plaid or MX.

02

Email

Especially if the tool can send, draft or delete for you, not just read.

03

Calendar

Less about money, more about who knows where you’ll be and when.

04

Health data

Apple Health, patient portals, wearables, anything biometric.

05

Cloud storage

Especially folders with tax returns, contracts, IDs or passport scans.

06

Browser agents

Anything that can click, fill out forms and complete a checkout for you.

Think of every new connector like hiring an assistant and handing them your card on day one. You’d want to know what they can see, what they can spend, and how to walk them out the door if it goes wrong. Same standard here.