Your CRM almost certainly gained a set of AI features in the last year or two, most of them enabled without anyone asking. Some are genuinely useful. Some produce confident output that nobody checks, which is worse than no feature at all.
Here is what the common ones actually do, and whether to leave them on.
Worth using
Call and meeting summaries
Transcribes a recorded conversation and writes the key points, objections and next steps to the record. This is the standout. It solves the real problem – CRM notes are thin because writing them happens after a call when the next one is starting.
The catch: participants must be told they are being recorded and agree. That is a legal requirement, not a courtesy. And the summary needs a glance before it is relied on, because what a salesperson judged important is not always what the model picked.
Drafting follow-up emails
Generates a first draft from the conversation history. Useful because it removes the blank page, which is the actual cause of slow follow-up.
The catch: it drafts, a person sends. Unchecked AI email in your name will eventually promise something you cannot deliver.
Data enrichment
Fills in company size, sector and similar from public sources so your records are complete enough to segment.
The catch: enrichment is often wrong in boring ways – outdated turnover, wrong sector code. Fine for broad segmentation, not for anything where accuracy matters.
Duplicate detection
Spots that two records are the same company spelled differently. Unglamorous and genuinely valuable, because duplicates quietly destroy trust in reporting.
Treat carefully
Lead scoring
Ranks leads by how likely they are to convert, based on past patterns.
The problem is that it learns from your history, and your history reflects who your salespeople chose to pursue. If the team historically favoured a certain type of enquiry, the model learns that preference and presents it as a prediction. You get a confident loop that is hard to argue with.
It also needs a decent volume of closed deals to be meaningful. Below a few hundred, it is a guess with a number attached.
Use it as a sorting aid when there are more leads than a person can assess. Do not use it to discard leads automatically, and compare its scores against your ideal customer profile rather than treating it as the authority.
Forecasting
Predicts what will close this quarter from deal data.
Output quality depends entirely on input quality. If deal stages are inconsistent and close dates are optimistic, the AI forecast reproduces that optimism with more decimal places. Fix your pipeline stages first; the forecast is a symptom, not the cure.
Chat and prospecting agents
Tools that engage website visitors or run outreach automatically. These can work, but the risk is reputational rather than technical – an agent that misunderstands and answers confidently is talking to a prospect in your name. If you use one, constrain what it is allowed to say and give it an easy route to a human.
The questions to ask before switching anything on
Where does the data go? Some features process data on the provider’s infrastructure or with a third-party model. Check what your client contracts and privacy notice allow, and what the provider does with what you submit.
Does it make decisions about people? Anything that scores or filters individuals – including candidates – may create obligations under UK GDPR around automated decision-making. Keep a human deciding.
Does your privacy notice cover it? If you record and transcribe calls, say so.
Who checks the output? If the answer is nobody, that feature is generating unverified content into the system your business runs on.
The underlying point
AI in a CRM is only as good as the CRM. Features that summarise and tidy work well because they operate on something that exists. Features that predict depend on data most businesses have not yet got right.
If your pipeline stages are vague and half your deals have no owner, no feature will fix that — it will present the mess back to you more confidently. Our wider look at where AI automation actually saves time follows the same rule.
Frequently asked questions
Should we pay extra for an AI tier?
Only if you can name the specific feature you want and the hours it saves. The transcription and summary features are usually the ones that justify it.
Can we use AI on client data?
Depends on your contracts and the provider’s terms. Some clients explicitly restrict where their data may be processed. Check before enabling, not after.
Will AI replace the need to update the CRM?
Partly, for notes. It will not decide what stage a deal is at or whether it is still real — those are judgements.
How do we stop it making things up?
Use it where the source material exists — summarising a call, drafting from a thread. Features that infer or predict are where errors appear, and those need human review.
Get the foundations right first
We set up CRMs around a defined sales process, then add automation only where it removes work — see our lead generation and sales service, or get in touch.
