An AI agent that cannot connect to financial data is not necessarily suffering from a bad password. The failure may be an expired consent, an institution security step, an unsupported account product, a temporary outage, a rate limit, or data that is not ready yet. The fastest fix is to identify the exact failure class before changing code or telling a user to retry.
This guide organizes the most common connection failures into three practical problems: authentication and consent, unsupported or unavailable data, and API throttling. It then shows how to capture useful diagnostics, protect credentials, verify data freshness, and decide when a provider-side tool can help.
As an Amazon Associate I earn from qualifying purchases.
Start with the exact operation and error
“I can’t connect my account” is too broad to troubleshoot. Write down the operation that failed: linking an institution, exchanging an OAuth authorization, listing accounts, reading balances, refreshing transactions, or asking the agent to summarize an existing dataset. Record the timestamp, provider, environment, endpoint, HTTP status, and complete error code and message. Keep request IDs, but remove API keys, access tokens, account numbers, and other personal data before storing or sharing logs.
OpenAI’s API guidance treats keys as secrets that must not appear in client-side code. Its Agents API documentation also recommends structured error codes and messages, rather than a single generic “connection failed” string. A diagnostic record should let an engineer answer four questions:
- Was the failure in the person’s institution connection or in your application’s API call?
- Which endpoint and product were requested?
- Did the provider identify a user-action, permission, support, availability, or limit problem?
- Is the returned data current enough and permitted for the agent’s intended use?
Problem 1: The login works, but consent or permissions do not
Credential errors are only one member of a larger authentication group. A connection can fail because credentials are invalid or expired, an account is locked, the institution requires an additional verification step, OAuth consent was revoked, a requested permission is missing, or the application asks for a data product the user did not authorize.
What the user sees
- An “invalid credential,” “username,” or “password” message even though the user can sign in through the institution’s own website.
- A repeated request to reauthorize after a password change, consent expiration, or institution policy change.
- A successful sign-in followed by a permissions or additional-consent screen.
- A connection that works for one account but not for a particular product, such as investments or detailed transactions.
- A locked account or a required one-time security action, such as multifactor authentication or an institution message that must be acknowledged.
OpenAI’s Finances in ChatGPT help documentation lists these causes alongside temporary institution unavailability. Therefore, “wrong password” is an unsafe default diagnosis.
Fix the user-connection path
- Verify direct access. Ask the person to sign in to the financial institution through its official channel. If direct sign-in fails, the institution’s recovery or security process must be completed before the agent can reconnect.
- Complete every institution-required step. Multifactor prompts, device approval, terms acceptance, account unlocks, and security questions can leave an OAuth flow incomplete even when the initial credentials were accepted.
- Relink after a credential or consent change. Password changes and revoked or expired consent commonly invalidate a previously stored connection. Start a fresh authorization rather than repeatedly submitting the old token.
- Inspect the requested OAuth permissions. Confirm that the authorization includes the data products your agent requests. A token can be valid while lacking permission for balances, transactions, holdings, or another restricted product.
- Check institution and account support. A user may have a supported institution but an unsupported account type, region, or data product. Treat that as a coverage problem, not an authentication problem.
- Retry only after the cause changes. Repeating the same request without a new token, completed security step, or corrected scope only creates noise and may trigger additional protections.
The control point for this class of failure is usually the end user, the institution, or the integration’s consent configuration. Application code should surface the provider’s specific message and a safe next action instead of asking the user to guess.
Recommended Free Tools
Problem 2: The account is connected, but the data is unsupported, not ready, or temporarily unavailable
A valid connection does not guarantee that every requested dataset is available immediately. Providers distinguish between an institution outage, an unsupported institution-product combination, data that is still being prepared, and a permission that was never granted. These states require different responses.
Separate the failure states
| Provider response or symptom | Likely meaning | Correct response |
|---|---|---|
| Institution temporarily unavailable | The provider cannot reach or complete work with the institution at this moment. | Show the outage state, preserve the connection context, and retry later according to provider guidance. |
| Unsupported institution or product | The requested account type, geography, or data product is not covered. | Offer a supported path or explain that this dataset cannot be retrieved through the current connection. |
| Data not ready | The connection exists, but processing or synchronization has not completed. | Display a pending state and check readiness; do not relabel it as a bad credential. |
| Missing permission or additional consent | The token is valid but does not authorize the requested data. | Send the user through the appropriate consent update, then request the product again. |
| Successful response with an old timestamp | The source returned data, but its update schedule is delayed. | Show the source time and delay to the user before the agent summarizes or acts. |
Use the provider response as a decision tree
- Check readiness and support before changing credentials. Read the endpoint’s error code and message. Terms such as “not ready,” “unsupported,” or “institution unavailable” point away from authentication.
- Confirm the exact product. “Accounts” may be available while balances, holdings, or transaction history are not. Verify that the requested product is supported for that institution and account.
- Check the institution’s status and retry policy. During an outage, repeated immediate retries add load without restoring access. Preserve the user’s connection and communicate that availability is the blocker.
- Handle asynchronous preparation explicitly. A pending synchronization should have a status in your application. Let the agent say that data is not ready instead of inventing a complete answer from an empty or partial response.
- Record source timestamps with every result. The time associated with source data is part of the answer’s meaning, especially for balances and market information.
Do not confuse coverage with interchangeability
A bank-data aggregator, a direct institution API, and a specialized financial dataset are not automatically interchangeable. Coverage depends on geography, institution, account type, consent flow, and the data products exposed by that connection. If the current path does not support a required dataset, switching connectors may solve the problem only if those specific requirements are verified.
Problem 3: The integration hits endpoint-specific rate limits
Once authentication and coverage are correct, an agent can still fail because it asks for data too often. Rate limits apply to particular endpoints and scopes; they are not a universal allowance for an entire integration.
Rank #4
Plaid’s current production documentation, checked in 2026, gives these examples per Item: /accounts/get allows 15 requests per minute, while /accounts/balance/get allows 5 requests per minute and 30 per hour. Plaid says the table is not exhaustive, limits can differ by customer, and thresholds can change. Treat those numbers as dated documentation examples, not permanent guarantees.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsRecognize throttling
- The same endpoint succeeds at low volume but returns a limit or too-many-requests error during bursts.
- One user’s Item fails while another user’s Item continues to work.
- Account listing succeeds, but balance refreshes fail because each endpoint has its own threshold.
- A scheduled job and an interactive agent compete for the same per-Item allowance.
Fix the request pattern
- Identify the endpoint and scope. Log the endpoint, Item or equivalent connection identifier, client environment, and response headers or error details that describe the limit.
- Reduce bursts. Remove duplicate reads, avoid refreshing unchanged data on every conversational turn, and serialize or stagger jobs that target the same Item.
- Use controlled retry behavior. Honor the provider’s retry information when supplied. Otherwise, apply increasing delays and stop after a bounded number of attempts; an immediate loop makes throttling worse.
- Cache within the data’s validity window. Reuse a recent successful response when the user’s request does not require a new fetch, while clearly labeling its retrieval time.
- Recheck live limits before release. Because thresholds and customer-specific rules can change, confirm the current provider documentation and configure limits rather than hard-coding an assumption.
Rate limiting is an application and provider-control problem, not a reason to ask the user to reset a password. The agent should explain that a refresh is temporarily limited and return the most recent permitted data only when its timestamp is visible.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.A diagnostic workflow that covers all three problems
- Capture a safe failure record. Save the operation, timestamp, provider, environment, endpoint, status, error code, message, and request ID. Redact secrets and financial identifiers.
- Classify the control point. Decide whether the next action belongs to the user or institution, your consent and integration configuration, the provider API, or the agent’s own error handling.
- Test direct user access when linking fails. A direct sign-in check distinguishes an institution lock or security step from an application defect.
- Check support and readiness. Verify the institution, account type, geography, requested product, OAuth scope, and synchronization status.
- Check availability and limits. Inspect provider status information and endpoint-specific throttling details before retrying.
- Validate the result before generation. Confirm that the response is complete enough for the requested action, attach source and retrieval times, and enforce any entitlement or sharing restrictions.
Using diagnostics without creating a second failure
Plaid documents a Dashboard MCP server for Production diagnostics and analytics. It requires Production access and an OAuth token carrying the mcp:dashboard scope; the token should be acquired through a backend rather than exposed to a browser or model prompt. Plaid also states that the server is under active development, may undergo breaking changes, and currently has limited support. Treat it as an operational aid with those prerequisites and limitations, not as a guaranteed fix for every connection.
Freshness, delay, and entitlement checks
An agent can produce a technically valid answer that is still unsuitable for a financial decision if the underlying data is delayed or restricted. OpenAI’s Financial Services Terms warn: “The time shown for source data does not guarantee that an AI response reflects the latest market conditions.” Show the source timestamp, identify any stated delay, and avoid implying real-time coverage when the provider does not promise it.
In the listed ChatGPT for Financial Services offering, OpenAI documents a 24-hour delay and a limit of 3,000 Daloopa datapoints per user per month, while Nasdaq data supplied through FMP is described as having a 15-minute delay. Those figures are product-specific and may change; they should not be generalized to every financial-data connection.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Before storing, copying, sharing, training on, or redistributing connected data or generated outputs, check the applicable partner terms and user entitlements. A connection that technically works may still prohibit a planned downstream use.
What a reliable error message looks like
Replace “Could not connect” with a message tied to the diagnosed class. For example: “Your institution requires a security step. Sign in directly, complete it, then reconnect.” For a support issue: “This account type is connected, but transaction data is not supported for this institution.” For throttling: “The balance endpoint is temporarily rate-limited; the last successful balance was retrieved at 14:05 UTC.” Each message gives the reader a cause, a safe action, and the data’s time context without exposing credentials.
Quick Recap
Decision table: choose the fix before changing code
| Failure class | Primary control point | First fix | What not to do |
|---|---|---|---|
| Credential, consent, or security step | User, institution, or OAuth configuration | Verify direct sign-in, complete required action, and relink with the correct scope. | Assume every failure is a bad password or loop retries. |
| Unsupported or not-ready data | Provider coverage or synchronization state | Check product support and readiness; communicate pending or unavailable status. | Invent a result from partial data or rotate credentials. |
| Institution outage | Provider availability | Preserve context, show the outage, and retry later under the provider’s policy. | Delete a valid connection or repeatedly hammer the endpoint. |
| Rate limit | Application request pattern and provider policy | Identify endpoint limits, reduce bursts, and use bounded backoff and valid caching. | Apply one global limit or present stale data as current. |
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
