Finance integration guide
Plaid + UbiGrowth workflows
Plaid is a financial data network that provides account verification and transaction data through user-permissioned connections. This guide covers the records that matter, how the connection should be scoped, and what the first bounded workflow should be.
Introduction
Make Plaid part of the workflow, not another silo.
Validate connector availability for your workspace
This guide covers how a team designs a finance workflow around Plaid with UbiGrowth: which records stay authoritative, how the connection should be scoped, what the first bounded workflow should be, and how to tell whether it worked.
The records that matter are Item, Account, Transaction, Institution, Access Token, and Webhook. Plaid identity is the Item — one user's connection to one institution — and Items break when a user changes credentials or the bank changes its login. A stale Item is the normal failure mode, not an exception.
Plaid is not currently on UbiVibe's verified connector list. This page is an implementation design reference: use it to specify the workflow, then validate whether the connection is available and correctly scoped for your workspace before you make it a dependency. The verified UbiVibe connections today are Salesforce, HubSpot, Gmail, Google Drive, Slack, and GitHub.
The platform layer is the usual destination for this connection, because the value shows up as governed context and execution shared across more than one team.
Why teams evaluate this connection
Integrations create value when they remove operating friction.
The first design decision is not which API endpoint to call; it is which system owns the record, what event should trigger work, who owns the exception path, and what successful completion means.
Finance integrations carry a different risk profile from the rest of the stack. A duplicated CRM record is annoying; a duplicated payment, a double-posted invoice, or an unauthorized write into Plaid is a real financial and control problem.
That is why most finance teams end up doing the work manually. The available automation is fast but cannot demonstrate what it did, and a number without a traceable basis is unusable in a function where everything has to be defensible.
You're likely here because
- Recurring finance analysis is mostly export, match, and reformat
- Approval trails live in email rather than in a system
- AI answers cannot be traced back to a source anyone will accept
Record model
What a Plaid integration actually reads and writes.
Integration design starts from the objects the system really exposes, not from a generic connector diagram. These are Plaid's.
Identity and matching
Plaid identity is the Item — one user's connection to one institution — and Items break when a user changes credentials or the bank changes its login. A stale Item is the normal failure mode, not an exception.
Start here
Consume the Item error webhook and route re-authentication to the customer before the data gap affects anything downstream.
What this will not do
It will not categorize spend authoritatively. Plaid's categories are inferred; finance decisions should use the ledger.
The constraint to plan around
Transaction data arrives with institution-specific delay and can be revised after posting, including changes to amount and date. Anything computed from transactions needs to be recomputable.
Build notes
What you actually have to reason about in Plaid.
The fields that carry meaning, how the connection authenticates, and whether the event surface can be trusted. This is the part that decides whether the integration works in month three.
| Field | Why it matters |
|---|---|
| item_id | one user's connection to one institution, and the thing that breaks |
| access_token | per item and per environment, non-transferable between them |
| transaction pending flag | pending transactions are revised, including amount and date, after posting |
| account subtype | depository, credit, loan — semantics differ enough to change the maths |
| ITEM_LOGIN_REQUIRED error | the re-authentication signal, and the normal failure mode rather than an exception |
Authentication
Client id and secret per environment, with Link handling the user-facing institution credentials so they never reach your system. That boundary is the product; anything routing around it is a serious mistake.
Events and delivery
Webhooks per item cover transaction updates and error states. The error webhook is the important one, because a broken item silently stops producing data otherwise.
Workflow
How the Plaid workflow runs.
The operating sequence, from reading the source system through to the result landing back where it belongs.
Step 01
Read the source records
Connected Plaid records replace the recurring export step, so the analysis starts from current data.
Step 02
Match and reconcile
Records are matched against the other systems involved, and anything that does not match becomes an explicit exception rather than a silent adjustment.
Step 03
Present the working queue
Exceptions and items needing attention are presented with owners, so the review is a queue rather than a highlighted spreadsheet.
Step 04
Stage the action for approval
Where an action is required, it is prepared with its supporting records and held for explicit human authorization.
Design decisions
The finance decisions this connection forces.
Each of these has to be settled before the Plaid workflow is allowed to write anything.
Step 01
Start read-only and stay there longer than feels necessary
Read access to Plaid delivers most of the reporting and reconciliation value with none of the write risk. Add write paths only when a specific, bounded workflow requires them.
Step 02
Keep authorization human
Anything that moves money, changes billing state, or affects a closed period requires an explicit human authorization step that is recorded, not inferred.
Implementation path
How to implement the Plaid workflow.
- 01
Define the data boundary: which Plaid records the workflow may read, and what is explicitly out of scope.
- 02
Start with a reporting or reconciliation workflow you already produce manually, so the output can be checked against a known-good answer.
- 03
Reconcile the connected output against the last two closed periods before anyone relies on it.
- 04
Once re-auth prompting works, add coverage reporting: what share of connected users have live data, which is the number that governs everything built on top.
Governance
Controls that matter.
Control 01
Payment, billing, and period-affecting actions require explicit human authorization; automation prepares, people approve.
Control 02
Credentials are least-privilege and reviewed, with read and write scopes separated wherever the platform allows it.
Control 03
Every automated action leaves a record that can be reconciled against the source system.
Failure modes
How a Plaid integration breaks in production.
Not generic integration advice. These follow from how this system actually behaves, which is why they look nothing like the list on the next guide over.
Symptom 01
Data stops arriving for a subset of users with no error.
Cause
Items entered a login-required state after credential or bank changes.
Fix
Consume the item error webhook and prompt re-authentication promptly; a stale item never recovers on its own.
Symptom 02
Computed balances drift from the bank.
Cause
Pending transactions were treated as final and later revised.
Fix
Recompute from current data rather than storing derived figures as final.
Symptom 03
An access token stops working after an environment change.
Cause
Tokens are environment-specific.
Fix
Store the environment with the token and never share tokens across environments.
What changes at scale
Item count drives cost and webhook volume. Re-authentication rate is the operational metric that matters, because it determines what fraction of the data is actually current.
Examples
What a working Plaid workflow looks like.
Bounded scenarios rather than a feature list. Each one can be verified against work the team already does.
Financial data workflows
With Plaid connected read-only, the recurring view is assembled from live records rather than a fresh set of exports each cycle.
Reconciliation exception queue
Unmatched or unexpected items from Plaid become a working queue with owners, instead of highlighted rows in a spreadsheet emailed around the team.
Limitations and considerations
What to validate before you depend on this.
- Transaction data arrives with institution-specific delay and can be revised after posting, including changes to amount and date. Anything computed from transactions needs to be recomputable.
- Payment initiation products move money. For data products the risk is interpretive: transactions are revised after posting, so anything computed from them must be recomputable rather than stored as final.
- When the ledger already holds the answer. Plaid categories are inferred, and finance decisions built on inference will be contradicted by the books.
- Nothing produced here is accounting, tax, or audit advice, and no output should be treated as a professional opinion.
- Write access to Plaid carries real financial risk. Duplicate protection, idempotency, and an approval step are requirements rather than refinements.
FAQ
Plaid integration questions.
What records does a Plaid integration actually work with?
The primary records are Item, Account, Transaction, Institution, Access Token, and Webhook. Plaid identity is the Item — one user's connection to one institution — and Items break when a user changes credentials or the bank changes its login. A stale Item is the normal failure mode, not an exception.
What should the first Plaid workflow be?
Consume the Item error webhook and route re-authentication to the customer before the data gap affects anything downstream.
What will a Plaid integration not do?
It will not categorize spend authoritatively. Plaid's categories are inferred; finance decisions should use the ledger.
What is the main constraint to plan around?
Transaction data arrives with institution-specific delay and can be revised after posting, including changes to amount and date. Anything computed from transactions needs to be recomputable.
What changes about a Plaid integration at scale?
Item count drives cost and webhook volume. Re-authentication rate is the operational metric that matters, because it determines what fraction of the data is actually current.
How does authentication work for Plaid?
Client id and secret per environment, with Link handling the user-facing institution credentials so they never reach your system. That boundary is the product; anything routing around it is a serious mistake.
Does Plaid support webhooks, and can they be trusted?
Webhooks per item cover transaction updates and error states. The error webhook is the important one, because a broken item silently stops producing data otherwise.
What is the risk of writing to Plaid?
Payment initiation products move money. For data products the risk is interpretive: transactions are revised after posting, so anything computed from them must be recomputable rather than stored as final.
When is connecting Plaid the wrong call?
When the ledger already holds the answer. Plaid categories are inferred, and finance decisions built on inference will be contradicted by the books.
What should a Plaid integration automate first?
Start with one bounded workflow that removes a measurable handoff, duplicate-entry step, reporting delay, or follow-up gap. Expand only after the first workflow is reliable.
Does UbiGrowth require Plaid to be replaced?
No. The operating model is designed around connecting to systems that should remain authoritative and building workflows around them rather than forcing a wholesale replacement.
Is connector availability identical for every workspace?
No. Availability can depend on provider configuration, authentication, scopes, workspace setup, and deployment state. Validate the required connection before treating it as an operational dependency.
Can the workflow write into Plaid?
Only where a specific bounded workflow requires it, with duplicate protection and an explicit human approval step. Most of the value is available read-only.
Is the output auditable?
Execution state and results return to the product surface, and analysis is grounded in connected records rather than an uploaded file, so an answer can be checked against its source.
Is this a replacement for the accounting system?
No. Plaid stays authoritative. The workflow layer sits around it to remove the export, match, and reformat cycle.
How this access is governed
What ARIA is allowed to do in Plaid, and who decides.
Connecting Plaid is a permission decision, not just a setup step. These are the controls that decide what ARIA can reach, what it can change, what gets recorded, and how you take the access back.
Required permissions
ARIA works through the scopes the connection was granted, and no others. Authorization happens at the provider, so the permissions being requested are shown by the system itself before anything is connected.
What it can reach
Reachable systems are the intersection of what your organization approved in the connector registry and what the requesting identity is permitted to use. Identity resolves before execution, not after.
What it can do
Actions run through explicit execution paths with state, spend, and failure boundaries — a bounded worker path rather than an open-ended agent loop with a credential.
Credential handling
Credentials live in the governed connection layer and are resolved through canonical connection identity. They are not pasted into individual workflows, prompts, or generated artifacts.
Action logging
Execution carries state and traces: what triggered the work, which connection it used, and what came back — including an explicit failure when something did not run.
Approval and revocation
Consequential actions can be made to require a person to approve them. Access can be changed or revoked at the connection, and ARIA loses that reach without unpicking the work already completed.
Start with ARIA
Ask ARIA to run this integration.
Describe the outcome you need across this system. ARIA works out the scopes, data, and actions the job requires, and operates inside the access you grant — which you can change or revoke.
- ARIA acts only through the systems and permissions you connect.
- Connections use scoped credentials you can change or revoke.
- Actions are recorded, and consequential ones can require approval.
Start here
Turn the integration into a working business outcome.
Start with ARIA to describe the outcome, then continue into the product path that fits the workflow. Connector availability and required scopes should be validated for the specific workspace before production use.