Back to the reading room
Business insights

AI systems integration: permissions, approvals and safe retries

Connect AI to existing business systems with clear permissions, human approvals and safe retry handling, plus a practical checklist for testing failure paths.

Stanic & Co.

Perspectives for business leaders

7 min read

Integrating artificial intelligence into your existing business systems is rarely about replacing your current software stack. Instead, it is about building bridges between the tools you already trust—your CRM, document management systems, and finance platforms—and the new capabilities AI offers. For many Australian businesses, the goal is to reduce the manual "swivel-chair" work where staff move data between disconnected applications. By connecting AI to these systems via supported APIs, you can automate repetitive tasks while maintaining the governance and reliability required for professional operations.

Mapping your workflow

Before writing a single line of code or configuring an integration, you must map the workflow. This involves documenting the current process from start to finish, identifying every system involved, and noting where data enters and exits. A common mistake is attempting to automate an inefficient process; instead, use this phase to simplify the steps.

Identify the specific triggers that start the process, such as a new email, a form submission, or a change in a CRM record. Determine the exact data points required for the AI to perform its task and where the final output needs to reside. By visualising the flow, you can identify potential bottlenecks and ensure that the integration serves a clear, measurable business purpose.

Connecting via supported APIs

Most modern business platforms provide Application Programming Interfaces (APIs) that allow external systems to communicate with them. Using these official channels is critical for stability. Avoid "screen scraping" or unofficial workarounds, as these are fragile and often break when the vendor updates their software.

When connecting your AI layer, look for platforms that offer well-documented REST or GraphQL APIs. These allow your AI to query data, update records, and trigger actions in a structured, predictable way. If a vendor provides an OpenAPI or JSON schema, use it; this allows your integration tools to "discover" the available functions, making the connection process significantly more robust.

Read-only pilots and least privilege

Start your integration journey with a read-only pilot. By restricting the AI to "read" access, you can observe how it interprets your data and interacts with your systems without the risk of it accidentally modifying or deleting records. This phase is essential for building confidence in the AI’s logic and ensuring it handles your specific data formats correctly.

Once you move to write-enabled workflows, strictly adhere to the principle of least privilege. Never use a master administrator account for your AI integration. Instead, create a dedicated service account or use OAuth scopes that limit the AI to only the specific objects and actions it requires. If the AI only needs to update lead statuses in your CRM, it should not have permission to export your entire customer database or modify system settings.

Human approval for writes

For critical business actions, such as sending invoices, updating financial records, or deleting customer data, implement a human-in-the-loop requirement. Even if the AI is highly accurate, a human approval step acts as a final safety net. This can be configured as a simple notification—such as an email or a message in your team chat—that requires a "click to approve" action before the AI executes the write command.

This approach balances the efficiency of automation with the accountability of human oversight. It is particularly useful during the early stages of an integration, allowing your team to verify the AI’s output before it impacts your live data. As your trust in the system grows, you may choose to automate lower-risk tasks entirely while keeping the human-in-the-loop for high-stakes operations.

Idempotency and safe retries

In a distributed system, network errors are inevitable. An API request might fail, or a response might be delayed, leading your system to retry the action. Without idempotency, these retries could result in duplicate records, double-charged invoices, or corrupted data. Idempotency ensures that performing the same operation multiple times has the same effect as performing it once.

An idempotency key works only when the receiving API explicitly supports and honours it. Stripe's documentation describes its own implementation; it is an example, not evidence that every CRM behaves the same way. If a destination does not support idempotency, design and test a deduplication strategy using a stable business-event identifier, a unique constraint where available and a durable record of completed actions. Concurrency and uncertain responses still need careful handling.

Audit logs and maintenance

Record the events necessary for accountability and troubleshooting, while minimising sensitive information in logs. Audit logs should capture who (or what) initiated the request, what data was accessed or changed, and when the action occurred. These logs are invaluable for troubleshooting errors, investigating anomalies, and meeting compliance requirements.

Integration maintenance is the unglamorous work that keeps your systems running long-term. APIs change, vendors update their platforms, and business requirements evolve. Establish a routine for monitoring your integrations, reviewing error logs, and testing your connections after major software updates. Treat your integration layer as a core piece of business infrastructure, not a "set and forget" project.

Illustrative example: automated lead qualification

Consider a hypothetical Australian consultancy firm that receives dozens of enquiries via a web form each week. Currently, a staff member manually copies these details into their CRM and checks if the lead is a good fit.

To automate this, the firm connects their web form to an AI service via an API.

  1. The AI reads the new form submission (read-only).
  2. It compares the lead details against the firm’s qualification criteria.
  3. The AI drafts a summary and a recommended action in a separate review queue, without changing the CRM.
  4. A staff member reviews the draft and clicks "Approve" to create the lead record.

The CRM update uses its documented idempotency feature if supported, or a tested deduplication mechanism otherwise. This reduces duplicate risk; it does not remove the need to reconcile uncertain responses and concurrent events. The entire process is logged, and the AI operates under a restricted service account that can only write to the "Leads" module.

Checklist for successful integration

  • Map the full workflow and identify the system of record.
  • Use official, supported APIs rather than unofficial workarounds.
  • Start with a read-only pilot to validate AI logic.
  • Apply the principle of least privilege to all service accounts.
  • Implement human approval for all write-based actions.
  • Use idempotency keys to prevent duplicate data during retries.
  • Maintain comprehensive audit logs for all automated actions.
  • Schedule regular maintenance to monitor API health and system updates.

The boundary that matters: proposals are not permissions

The language model should propose an action; a deterministic application layer should decide whether that action is allowed. Do not execute an arbitrary URL, database query or API operation because it appears in a generated response. Validate the permitted action, target record, required fields and authorised user before any write.

For a hypothetical customer-enquiry workflow, allow the model to suggest a category and draft a response. Do not let the contents of the enquiry choose the destination account, override permissions or request a data export. A malicious email can contain instructions aimed at the model. Treat customer messages, retrieved documents and model output as untrusted data, even when they are formatted professionally.

An approval screen should show the exact target and change, not merely an attractive summary. If the proposed action changes after approval, require a new approval. Protect approval links with authentication and an expiry; a forwarded link should not grant authority to modify a business record. The NIST AI Risk Management Framework provides a voluntary risk-management structure, not a substitute for these application controls.

Test the failure path before switching on writes

Use a sandbox or non-production records to simulate five events: the same event delivered twice; two workers handling it simultaneously; a timeout after the destination accepted the change; a permission revoked mid-process; and a destination schema changed by the vendor. Write the expected result for each before testing.

The difficult retry case is an uncertain response. A timeout does not prove the write failed. Query the destination using the original transaction or event identifier where supported before deciding to retry. Never create a new business-event identifier merely because the network request is repeated. If you cannot determine what happened, send the case to reconciliation rather than blindly writing again.

Some actions cannot be rolled back. An email already sent or an external payment instruction requires a different recovery procedure from an editable CRM note. Record the owner and recovery steps for each irreversible action. Set a kill switch that stops new writes without losing the queue of pending cases.

Finally, check retention for prompts, outputs and integration logs. The OAIC's privacy guidance addresses the handling of personal information in commercial AI products; the integration design must consider the actual product and contract. A narrowly scoped integration is easier to operate and assess than one given blanket access to all systems. Choose the first workflow using our AI pilot selection guide.

Sources

General information only; verify sources before making business decisions.

Turn an idea into a next step.

A conversation about what matters to your business.

Let's talk