Building custom AI plugins for CRMs, e-commerce and SaaS platforms

The AI feature people actually use is the one that appears inside the tool they already have open. Plugins, whether for HubSpot, Salesforce, Shopify, Zendesk, Slack or your own platform, are how AI stops being a separate destination.

Three patterns

1. Enrich: when a record is created, fill in what's missing (company data, classification, summary).
2. Draft: propose the next email, note, or reply for a human to send.
3. Act: update fields, create tasks, move stages, with approval where the action is hard to undo.

The engineering that keeps it safe

• Scoped credentials: the plugin gets exactly the permissions it needs.
• Idempotency: the same event processed twice must not create two records.
• Approval gates: anything that touches money, deal stage or customer communication waits for a click.
• Audit log: who/what changed which field, when, based on what input.
• Dry-run mode: run in shadow for a week and compare with what humans did.

A worked example

For a distributor we built a CRM agent that enriches inbound leads, classifies intent, drafts a follow-up and proposes updates. Low-risk fields apply automatically; deal-stage changes wait for the rep. Reps saved ten-plus hours a week and missing CRM fields dropped by around 90%.

Start small, expand on evidence

Enrich first, then draft, then act. Each step earns the trust that justifies the next.