← All articles
Where AI actually helps in a CRM (and where it doesn't)
A grounded look at AI in CRMs in 2026: the tasks it genuinely speeds up, the ones it quietly makes worse, and how to adopt it without eroding trust in your data.
Every CRM now has an 'AI' button. Some of it is genuinely useful; a lot of it is a demo that doesn't survive contact with your messy data. Here's how I split the two when advising clients.
Where AI earns its place
- Drafting, not deciding. First-draft replies, follow-up emails and meeting summaries save real time — because a human still approves them.
- Enrichment and tidying. Normalising job titles, guessing a company from a domain, deduping — bounded tasks with a checkable answer.
- Summarising a timeline. 'What's the history with this account?' across dozens of activities is exactly what a language model is good at.
Where it quietly makes things worse
- Silent field-writing. AI that edits records without a human in the loop erodes trust in your data the first time it's wrong — and you won't notice for months.
- Scoring you can't explain. A black-box lead score sales can't interrogate gets ignored. Prefer transparent, rule-based scoring.
- Automation without a rehearsal. Letting AI trigger customer-facing sends unsupervised is how you get the 3am blast.
The adoption rule
Put AI where a human reviews the output, and keep deterministic rules where the system acts on its own. That single boundary is the difference between AI that compounds and AI that costs you.
This is the kind of practical rollout I help teams with — more on that here. Or see automation workflows worth stealing.