AI and your CRM: what actually helps and what is a demo
CRM vendors now bundle AI everywhere. Four uses that help a small sales team, three that mostly do not, and how to tell before you pay.
The short answer
Every CRM vendor now bundles AI into every screen, and some of it is genuinely useful for a small sales team. The pattern is simple: AI helps where there is text to process and the data underneath is decent. It mostly does not help where it is asked to predict from thin data or to replace human judgement at volume. Before paying for any of it, test it on your own last month of records, because a CRM full of stale, duplicate and empty entries turns every AI feature into confident nonsense.
Four uses that help
| Use | Why it works | What to watch |
|---|---|---|
| Summarising calls, meetings and email threads into the record | Text in, structured summary out: the task AI is best at | Review the first weeks; check nothing sensitive is stored where it should not be |
| Drafting follow-ups with the deal context | The context makes drafts specific; a person edits and sends | Never auto-send; generic drafts at volume hurt reply rates |
| Enriching and deduplicating records | Matching names, companies and addresses across messy entries | Confirm merges on anything with money attached |
| Flagging deals gone quiet or next steps missed | Simple signals surfaced at the right time | It only works if activities are logged; fix logging first |
Three that are mostly demos for small teams
- Predictive lead scoring. It needs many closed deals with complete data to find a pattern. A few hundred deals with half the fields empty gives a score that looks precise and means nothing. Rules you define from your own experience usually outperform it.
- Revenue forecasting from the pipeline. If stages are updated late and close dates are guesses, the forecast is a confident restatement of the guesses.
- Generated outreach at volume. Personalisation from a template plus a few fields reads as exactly that. Reply rates fall, and the domain’s sender reputation can follow.
How to test before you pay
Take last month’s data. Run the summarisation on ten real calls or threads and check them against what actually happened. Let it draft follow-ups for five live deals and have the salesperson grade them. Run the deduplication in preview mode and inspect the proposed merges. Turn on the quiet-deal flag and see whether it catches what the team already knew. Half a day of testing tells you what a quarter of subscription would.
What this means for you
Clean the data first. Then turn on the features that process text, summaries, drafts, enrichment and quiet-deal flags, with a person reviewing anything that reaches a customer. Leave scoring and forecasting off until the data can support them. Test everything on your own records before paying, and remember that the CRM’s API lets you build the step you actually need when the bundled one is a demo.
Frequently asked questions
Should we switch CRM to get better AI features?
Rarely. The features are converging across vendors, and the migration costs more than the difference. Fix data quality in the CRM you have, turn on the features that process text, and connect an AI step to your own workflows where the vendor's version falls short.
Can AI write our sales emails?
It can draft them well when given the context: the contact's history, the last conversation, what you sell. Sent unreviewed at volume, the drafts read as generic and damage reply rates. Drafted for a person to edit and send, they save real time. The difference is the human in the loop.
Is lead scoring worth turning on?
Only if you have enough closed deals, won and lost, for a pattern to exist, and enough fields filled in for it to be found. Without that volume the score is noise dressed as insight. Rules you write yourself, based on what you know, usually beat it.