There’s a lot of hype about AI and Salesforce right now. Most of it skips the boring truth: AI only helps if your CRM is set up well in the first place. Bolt a smart assistant onto messy data and you get confident, fast, wrong answers.
After years as a certified Salesforce admin — and now building AI automations full-time — here’s how I think about adding AI to an org that people actually depend on.
Start with the data, not the model
The single biggest predictor of whether AI works in Salesforce is whether your data is clean and well-structured. Before any AI:
- Fix the data model. Clear objects, sensible fields, no duplicate “source of truth” everywhere.
- Tighten the security model. AI that can read records needs the same guardrails your users have.
- Automate the basics with Flow first. If a process can be a deterministic Flow, it probably should be — it’s cheaper and more predictable than AI.
Then add AI where judgment is needed
AI earns its place on the tasks that aren’t deterministic — reading unstructured text, summarizing, drafting, classifying intent. A few high-value patterns:
- Summarize long records (cases, opportunities, email threads) so reps get context in seconds.
- Draft the first version of replies, follow-ups, and notes — a human still approves.
- Triage and route inbound messages based on what they actually say.
Build vs. buy
Salesforce’s own Agentforce is powerful, and for some teams it’s the right answer. But you can also build lighter, cheaper automations with Claude Code that talk to Salesforce through its APIs — and you own them outright. The right choice depends on your scale and budget, not on whatever’s being marketed hardest.
The takeaway
AI in Salesforce isn’t a feature you flip on. It’s a layer you add once the foundation is solid. Get the data and automation right first, then let AI handle the judgment calls — and you’ll get tools your team actually trusts.
Thinking about AI in your Salesforce org? Let’s talk.