Start with the tool you already have.

If the job is to assign a lead, create a reminder, send information to the right queue or require a field before the next stage, check the CRM’s existing configuration first. Clear rules are often easier to understand and maintain than an AI decision.

Before comparing tools, write the desired process in plain language and identify what is actually missing. Poorly defined responsibilities will not be solved by a new platform.

If two systems hold the information but people keep copying it between them, investigate an integration. Check how records are matched, what happens to duplicates and who is alerted when an update fails.

An integration still needs ownership. Agree who maintains it and how the team can tell whether it ran successfully.

AI becomes more relevant when the job involves reading varied documents, interpreting free-text information or drafting something for a person to review.

Test representative inputs and awkward cases. Agree which outputs can move on automatically and which require a person. The review process should be part of the scope.

A custom system may make sense when the team needs a review interface, a specialist workflow or functionality its existing tools cannot provide. Account for the handover, maintenance and user training alongside the build.

If you are in the middle of a CRM rollout, understand what it will deliver before commissioning overlapping work. An AI audit can help identify the remaining opportunities and dependencies.

A sales manager wants an alert when a quote has been untouched for seven days: start by checking the CRM’s existing rules. A finance team is copying approved records between two tools: investigate a supported integration and duplicate handling. A caseworker receives the same information in many different document layouts: test AI extraction against representative files, with an explicit review step.

The label on the technology does not determine the right approach. Choose the simplest option that meets the requirement, has a clear owner and can be checked when something goes wrong.

Ask the person who administers your CRM whether the required data can be read and updated through a supported integration. Check the relevant plan, permissions, record identifiers and limits. A connection that works in a demonstration may still need a way to recover from duplicates, missing fields or a temporary outage.

For example, when an enquiry becomes a CRM record, decide what uniquely identifies it, whether a repeat submission updates or creates a record, and where failed submissions are reviewed. Test those cases before adding AI to qualify or summarise the enquiry.

Can the existing tool meet the requirement with configuration? If not, what specifically needs to be built? Ask to see the proposed inputs, outputs and human review step, alongside the accounts and permissions the solution needs.

Agree who will maintain the connection after launch, what happens if a supplier changes its service, and how the team can continue working during an outage. Choose a pilot with a measurable outcome before replacing a business-wide workflow.