STKD.DIGITAL · BUSINESS IN MOTION
STKD.digital
Menu
All software and integrations insights

Software and integrations · 5 MIN READ

Why CRM Integrations Break, and What to Ask Before You Buy One

A working demo is not a reliable integration. Ask how it handles duplicates, stale data, outages and the person responsible for exceptions.

THE SHORT ANSWER

CRM integrations break when records have no clear owner, the same event runs twice, updates arrive out of order or failures have no recovery path. Before buying, ask the supplier to demonstrate those cases with your workflow. A successful data transfer is only the start; the business needs a correct, traceable result and an owner for exceptions.

The demo usually shows the easiest case

For an established business, reliability means more than transferring information quickly. It means the right record changes, only the intended action occurs and someone can explain what happened. A system that silently creates duplicate tasks or overwrites approved information can increase work while appearing automated.

Guide map: The demo usually shows the easiest case; Decide who owns each piece of data; Plan for repeated and out-of-order events

Decide who owns each piece of data

Before connecting applications, write a field-level ownership map. Which system owns the customer name, delivery address, sales stage, invoice status and consent record? Who can edit each field? What should happen when the values disagree?

Avoid an automatic last update wins rule unless the business actually accepts its consequences. A late import should not casually replace a correction made by an employee. Define conflict handling explicitly and send uncertain cases to review.

Illustrative example: the CRM owns relationship details while a billing platform owns whether an invoice is paid. A project tool may display both, but it should not become a competing authority just because it contains copied fields. This is a suggested design pattern, not a client implementation claim.

Plan for repeated and out-of-order events

Stripe's webhook documentation states that events can arrive more than once and are not guaranteed to arrive in their original order. The precise behavior varies by provider, so your supplier must check each connected system rather than assuming all notifications behave identically. Stripe webhook documentation.

A public LinkedIn discussion by Ayush Chugh focuses on repeated writes and idempotency. In plain language, the desired property is that retrying the same intended operation does not create a second business effect. Use that as a buying question, not merely a technical term to put in the proposal. Read the discussion.

There are separate problems to test: receiving the same event twice and making the same outgoing request twice. Stripe documents idempotency keys for safely retrying applicable API requests. Deduplicating an incoming notification alone does not prove every downstream action is protected. Stripe idempotent requests.

Design visible failure, not silent failure

Ask what happens when a destination is unavailable or rejects the record. The integration should distinguish a temporary problem that can be retried from an invalid record requiring a decision. Repeatedly retrying bad data is not resilience; it is repeated failure.

Require a visible exception queue with the failed step, affected record, last attempt and next owner. Use logs that support diagnosis without unnecessarily copying sensitive information. Decide how alerts escalate if nobody responds and how normal processing resumes after a repair.

A reconciliation check should compare the states that matter to the business. For example, do approved orders have the expected project record? Do completed jobs have the expected billing handoff? A green automation run does not establish that the overall business outcome is correct. Reconciliation asks the question the individual task log cannot answer.

Keep AI within a defined boundary

An AI component can help interpret an unstructured request, propose a field mapping or summarize an exception. It should not silently invent missing customer data or decide that two uncertain records are the same person. Separate suggested interpretations from verified facts.

For consequential actions such as issuing a refund, changing payment details or sending a binding commitment, define authorization and human approval requirements explicitly. Give the integration only the permissions needed for the workflow. A broad administrator credential is not a substitute for designing the access model.

Start with a narrow workflow and representative test records. Run normal cases and failure cases before allowing real side effects. Where practical, compare proposed actions with human decisions first, then gradually enable only the actions the business has approved.

The acceptance-test checklist

Ask the supplier to demonstrate these cases rather than answering yes on a questionnaire. The results belong in the handover documentation so the next maintainer can repeat them.

  • The same input arrives twice without creating duplicate business effects.
  • An older event arrives after a newer one without corrupting the current state.
  • A destination outage produces bounded retries and a visible exception.
  • A required field is missing and the record is routed for correction.
  • Conflicting edits follow the agreed field-ownership rules.
  • Access is revoked, an owner is unavailable and recovery is tested.
  • A reconciliation check detects a missing downstream outcome.

What you should own after delivery

Keep the system map, field definitions, test cases, account ownership, maintenance responsibilities and recovery instructions. Your team does not need to become an integration engineering department, but it should understand the boundaries and know who acts when something goes wrong. The best integration is not the one with the most arrows in the diagram. It is the one that removes a real handoff while making exceptions easier to see and resolve.

Sources and further reading

Find my biggest bottleneck