Gartner expects 40 percent of enterprise applications to embed task specific agents by the end of this year, up from under 5 percent in 2025. Your CRM is almost certainly in that number, and the agent probably arrived without a decision being made.
It appeared in a release note, defaulted to enabled, and began summarising calls, drafting follow ups, updating fields and scoring leads. Some of that is genuinely useful. The problem is that nobody chose which parts, and a CRM is the one system where the difference between useful and harmful is hard to see until a forecast is wrong.
Writing to the record is the line that matters
Draw a distinction the vendor's settings page probably does not. There is an enormous difference between an agent that suggests and one that writes.
A suggestion is reviewed by a person who accepts or ignores it. The record stays a record of what happened, and a wrong suggestion costs a moment of attention. A write changes the data other people rely on without anyone confirming it was right.
The trouble is that CRM data is used far beyond the person looking at it. Pipeline reports, forecasts, commission, territory decisions and board packs are all downstream of those fields. An agent that helpfully updates a deal stage based on the tone of an email has changed a number the finance team will present next month.
So the first configuration pass is straightforward: find everything the agent can write and decide, field by field, whether it should. Most teams doing this discover the agent can write to more than they assumed.
We treat this as the first configuration decision on any vendor agent rollout, before anything else is discussed. Reading is recoverable and writing is not, and the default is almost always set to write.
Summaries are not the same as notes
Call and email summarisation is the feature most people are happy about, and it is worth being precise about what it produces.
A rep's note says what they judged important. An automated summary says what the model extracted. These are usually similar and occasionally very different, particularly around commitments, objections and anything said carefully. A summary that records "customer is interested in expanding" when the customer said "we might look at that next year if budget allows" is not wrong exactly, and it will mislead whoever reads it in six weeks.
The practical fix is to keep them separate rather than blended. An automated summary stored as an automated summary, clearly labelled, is useful context. The same text stored in the notes field where a human note would go is a quiet loss of information about who said what.
Where the transcript is available, keep it. When a deal goes wrong the question is always what was actually said, and a summary cannot answer that.
Read More - A Search Plugin Scored 9.8: How to Read Severity Without Panicking
Lead scoring deserves particular scrutiny
Automated scoring is old, and agent based scoring differs in one respect that matters: it can consider unstructured signals like email tone and meeting content, which makes it more useful and much harder to explain.
Before you let a score influence routing or prioritisation, establish whether anyone can say why a lead scored as it did. If the answer is no, you have introduced something that will eventually deprioritise a good lead and nobody will be able to reconstruct why.
Also check what it was trained or prompted on. A model that learned from your historical conversions will reproduce your historical biases, including the ones you would rather not repeat, such as systematically underrating an industry a former rep happened to be bad at selling into.
Run it alongside your existing process for a quarter before letting it decide anything. Compare where they disagree, and look at those cases specifically. That is the cheapest way to find out whether it is adding judgement or noise.
Adoption problems get worse, not better
Most CRMs have a data quality problem originating in a familiar place: reps update records reluctantly, often just before a review. We wrote about that in why CRM adoption fails, and an agent interacts with it in a way that is easy to miss.
If reps see the agent filling fields in, the incentive to maintain records themselves weakens. Why write a note when something writes one for you. The data then becomes progressively more automated and less observed, and the moment it is wrong, nobody notices because nobody was looking.
An agent can genuinely fix adoption by removing the tedious part of the job. It can also hollow it out. Which one you get depends on whether the rep still has to confirm what went in, and that is a configuration choice rather than a property of the software.
Where we see this bite hardest is teams who were already sceptical of the CRM. An agent adding entries nobody made confirms the existing suspicion that the system does not reflect reality, and adoption goes backwards.
A configuration review worth an hour
List every agent capability now enabled in your CRM. Vendors do not always make this easy, which is itself informative.
For each, decide whether it suggests or writes, and whether writing is appropriate for that field. Anything feeding a forecast, a commission calculation or a report should require human confirmation, because the cost of quiet wrongness there is measured in decisions rather than in tidiness.
Then check the audit trail. When a field changed, can you see whether a person or an agent changed it, and on what basis. If your CRM does not distinguish those, you have lost the ability to investigate your own data, which matters the first time a number is questioned.
None of it works on records nobody has cleaned, which is the constraint we went through in why dirty pipeline data breaks automation.
Finally, decide who owns this. Agent capabilities will keep arriving in release notes, and defaults will keep being set by the vendor. Somebody in sales operations needs to review each one rather than discovering it later. If your CRM is custom or heavily configured, our CRM development team treats this as part of the system rather than a feature toggle.
Frequently Asked Questions
Should we turn the agent off entirely?
Rarely the right answer. Summarisation and drafting genuinely save time, and switching everything off forfeits that to avoid a risk you can configure away. Turn off automatic writes to fields that feed reporting, keep the assistive features, and review the rest deliberately.
How do we know what it has already changed?
Check the field history or audit log, filtered by the agent's identity if your CRM records one. If it does not distinguish agent edits from human ones, that is the first thing to raise with the vendor, because without it you cannot audit your own pipeline.
Can we trust automated call summaries?
As context, generally yes. As the record of what was agreed, no. Summaries compress and interpret, and interpretation of a hedged commitment is exactly where they drift. Keep the transcript where you can, and label summaries as summaries rather than storing them as notes.
Will this fix our data quality problem?
It can, and it can also disguise it. Fields get populated, so the CRM looks healthier while fewer people are actually checking anything. The improvement is real only if someone still confirms what went in. Otherwise you have replaced missing data with unverified data, which is harder to spot.
What about compliance and recorded conversations?
Summarisation usually implies recording or transcription, which carries consent obligations that vary by jurisdiction and are often stricter than people assume. Confirm what is being retained, for how long, and whether your existing notices cover it, before the feature has been running for six months.
Our vendor enabled this without asking. Can we push back?
You can ask for granular controls and an audit trail, and you should, because you will not be the only customer asking. Vendors respond to that. In the meantime configure what you can and document what you cannot, so the gap is visible rather than assumed away.
Does this change how we should think about custom CRM work?
It raises the value of owning the configuration. A packaged CRM will keep making these decisions for you at the vendor's pace and in the vendor's interest. Where the pipeline data is central to how a business runs, controlling what writes to it is worth more than it was a year ago.
Talk to our team if an agent is writing to your pipeline and nobody has decided what it should touch.