Existing business systems contain the records and rules that keep an organisation working. Adding AI should begin by understanding those systems, rather than treating them as a barrier to remove.
A useful integration gives the application a clear, limited way to read information or request an action. It keeps responsibility for authoritative records and business controls visible.
Inventory the process before the interface
Follow one task from start to finish. Identify where information enters, which system owns each record and where people reconcile discrepancies. Document scheduled jobs, spreadsheets and manual workarounds as well as formal APIs.
A field with the same name in two systems may have different meanings. “Customer” might refer to a billing account in one system and an individual contact in another. An integration needs an explicit mapping, not a model’s best guess at every step.
Ask which failures the current process already handles. A human may be catching an exception that the new workflow otherwise leaves invisible.
Prefer bounded operations
Expose tools that match the work: retrieve an approved record, prepare a proposed update or submit a validated request. Avoid giving a model a broad administrative interface when the task only needs a narrow capability.
Separate reading from writing. A document-search assistant may need no write access at all. A workflow that updates records needs an application-side permission check, validated fields and confirmation of the external result.
Keep credentials outside prompts and store them through an appropriate secrets mechanism. Identify whose authority each request uses and how access can be revoked.
Translate formats in software
An adapter can handle provider-specific fields, identifiers and errors while the workflow uses an internal representation. Validate incoming data at that boundary.
Models may help interpret unstructured material, but ordinary code should handle exact date conversions, required fields and calculations where the rules are known. Ambiguous interpretations should remain visible for review.
Version the integration contract. If a connected API changes, a failing contract check is more useful than a model silently adapting to an unexpected response.
Plan for partial completion
A multi-system workflow can fail after one step succeeds. Record each completed action and its external identifier. Do not tell the user that the whole task succeeded because the final language-model response sounds complete.
If an external write times out, reconcile the target’s state before retrying. Use idempotency keys or stable references where supported. Decide which changes can be reversed and which need a separate authorised correction.
A catch-all retry loop is not a recovery strategy. Give unresolved work an owner and a visible status.
Modernise in manageable boundaries
Some systems need replacement. Others can remain useful behind a tested integration layer. Select a boundary small enough to validate without requiring an immediate rebuild of every dependency.
Migration taxonomies are planning aids rather than laws. For example, AWS describes seven migration strategies; the right decision still depends on the application and its operating requirements.
Measure whether the change improves the actual workflow. Consider exception handling, user effort, support and reconciliation, not just the speed of an isolated model call.
Test against the real constraints
Use representative records and permission combinations in a controlled environment. Include missing fields, duplicate events, stale data, unavailable services and a user losing access midway through a task.
Check that diagnostic data is protected. Copying live records into prompts, traces or developer tools creates additional data flows that need review.
Start with the smallest integration that serves the task and establish how it will be maintained. Our Secure System Connections capability covers the interface, identity and validation work that connects AI to an existing operation.
Written by Raj Dalal, Advisor at Ryzolv.
Back to Insights