Do you need custom AI, or a better process?

“We need AI in our CRM” describes a technology choice. It does not yet describe the problem. The useful conversation starts a little earlier: what is someone trying to do, what stops them, and what would a good result look like?

A custom AI integration can be valuable when a task needs language understanding and information from several systems. It can also become an elaborate way of avoiding an unresolved decision about who should answer an enquiry.

First, check whether the process is agreed

Ask two people to explain how the task should work. If their answers differ, settle that before automating it.

For instance, one colleague may think a sales enquiry belongs to the person who opened the email. Another may assign it by area. Neither a workflow tool nor an AI assistant can resolve that policy for you. They can only apply a rule, or guess at one.

Write down the owner, the input, the next action and the point where the task is finished. You may find that a clearer handoff is enough.

Use ordinary automation for ordinary rules

When the rule is predictable, a conventional integration is often the easier option to test and maintain. A submitted form can create a record. An accepted quote can trigger an internal task. A missing mandatory field can stop the process and ask for a correction.

AI is more useful where the input varies: summarising a long conversation, finding relevant passages in documents or preparing a draft from rough notes. Even then, a simple rule may be the right way to handle what happens after the draft is approved.

Our enquiry-to-job article shows where those two approaches can sit in the same workflow.

Try the existing tool before commissioning one

If an existing assistant can handle the task safely with a few examples, test that first. You may need training or a reusable instruction rather than software development.

Some tools can connect to business systems already. Anthropic’s connector guidance describes connecting Claude to tools and data. Check the specific connector’s supported actions and permissions; the existence of a connector does not establish that it supports your whole process.

A custom integration becomes more plausible when the standard option cannot access the right records, enforce your review steps or fit where staff do the work. Be precise about the gap. “We must draft from the approved project record and save for the account owner’s review” is a buildable requirement.

Include the person who checks the output

Review is part of the workflow, not a final disclaimer. Who checks the draft? Can they see the source information? What happens if they reject it? Does anything get sent or changed before they approve it?

If an assistant can take actions, give it only the access needed for the task. Keep a record of the actions and a way to recover from a mistake. The NCSC’s AI security guidance explains risks including misleading outputs and prompt injection. Those risks matter more when text from outside the business can influence a tool with permissions to change records.

Ask what happens after launch

A working demonstration is only part of the decision. Documents change, connected systems change and users find cases the initial test missed. Someone needs to look after the integration and its running costs.

Ask for the ongoing charges, the support arrangement, how changes are tested and how you can export your data. Managed services can cover that work, but the scope should be explicit.

If you cannot yet name the task or judge a good result, start with a small AI trial. If you can, bring the task and a few representative examples. They give us a much better starting point than a request to put AI everywhere.