An AI agent for CRM upkeep

The CRM problem is not the CRM

Every company of this size has bought a CRM, and most are disappointed by it. The software is rarely the reason. The reason is that keeping it current is administrative work assigned to the people with the least patience for administrative work, at the end of a day when they have already done their actual job.

So records get filled in from memory on Thursday for a call on Tuesday, or not at all. Then someone builds a report on that data and everyone quietly agrees not to trust it.

What the agent does

• Creates the record from the conversation — from the email thread, the call transcript where one exists, the appointment that was booked.
• Fills the structured fields: who, what was discussed, what was agreed, what the next step is and when.
• Links related items, so the enquiry, the call and the quote sit on one record instead of three.
• Flags what it could not determine rather than leaving a field silently blank — an empty field and an unknown value look identical later, and they are not the same thing.

The hard limit

An agent cannot recover what was never captured. A conversation in a car park, a decision made in a corridor, a customer’s reservation that someone picked up on but never wrote down — none of that reaches the CRM, and no amount of capability changes it.

That matters for expectations. This use case makes existing information structured and complete. It does not make undocumented knowledge appear, and a project sold on the second promise fails on contact with the first month.

Recording is a legal question first

Call transcripts make this use case considerably stronger, and recording calls requires consent from both sides in the EU and in many other jurisdictions. That is a decision to settle before anything is built, not a detail for the launch checklist. Where recording is not appropriate, the agent works from written material and delivers less — which is still more than the current state.

What stays with people

The factual layer is the agent’s: who, when, what was said, what was agreed. The assessment layer is not. Whether the deal is likely, whether the customer sounded uneasy, whether it is worth another call — those are judgments from someone who was present, and a confident-sounding automated guess in that field is worse than an empty one.

What it connects to

The CRM with write access, the email system, and the phone system where transcripts are in play. Calendar too, since appointments are part of the record.

How to tell whether it worked

• Field completeness on records from the last 30 days.
• Time between a conversation and its record existing.
• Contacts with no record at all — the number most people have not looked up.
• Time the team spends on CRM admin per week.

Record the first three before starting. They are the argument for doing this at all, and they are the only way to show afterwards that anything changed.