In voice agent evaluations, conversation quality is always front and center, while what remains after the call is an afterthought. Operationally the second matters more: an agent that speaks beautifully but leaves no trace is an investment you can neither measure nor hand over.
This post covers how to set up post-call CRM writing.
Four outputs
Every call should produce four things:
1. A summary. What was discussed, what was asked for, what was done. It must be short — a summary longer than a paragraph goes unread. A representative opening the record should understand the situation in thirty seconds.
2. An outcome tag. How the call ended. Not free text, but a selection from a fixed list: appointment created, information given, complaint received, purchase intent, unreachable, handed off. The fixed list is what makes reporting possible.
3. Field updates. Things learned in the call that have a home on the record: a corrected phone number, an updated address, the product of interest, a budget range, a preferred channel.
4. A task. If work came out of the call, it should be a task: call back, send a quote, forward a document. With an owner and a date.
When those four are not complete, the call is only half finished. The one most often skipped is the fourth: the agent writes a nice summary but leaves the work as a note nobody owns.
Designing outcome tags
The tag list should be neither too long nor too short. A long list gets tagged inconsistently; a short one produces no distinctions.
An approach that works has two layers:
- Layer one: outcome. Resolved, unresolved, handed off, unreachable.
- Layer two: topic. Order, return, appointment, pricing, complaint, technical support.
Combined, reporting becomes meaningful: "what percentage of return-related calls went unresolved?" becomes answerable. With single-layer tagging those questions stay unanswered.
Review the tag list at the end of the first month. Tags never used should go, and if the share of "other" is high, new tags are needed.
Three rules for writing summaries
Write the decision, not the conversation. Not "the customer asked about their parcel, the agent said it was in transit, the customer said thanks," but "shipping delay reported, delay record opened, customer to be contacted Wednesday."
State uncertainty. If the agent was not sure of something, the summary should reflect that. "Customer probably wants a corporate invoice" is better than false certainty.
Preserve the customer's own words. Especially in complaints. A summarized complaint loses its tone; carrying the critical sentence verbatim gives the representative context.
Which fields should update automatically
Care is needed here. Automatic updates are convenient, but wrong updates corrupt data quality.
Can update automatically: contact preference, product or service of interest, call note, last contact date, channel preference.
Should update on approval: phone number, address, billing details, tax number. On these the agent proposes the change and a representative approves. A single digit misheard on a call corrupts the record.
Should not update automatically: derived fields like customer value, segment, credit limit or risk score. Those should not move on the basis of a single call.
Merging the written and voice channels
CRM writing must be channel-independent. If the same customer wrote on WhatsApp, called the phone line, and called over WhatsApp, all three should accumulate under the same record.
Without that merge two problems appear: the representative cannot see the history, and reporting fractures by channel — the same customer ends up with three separate stories in three places.
The key to merging is number matching. On a WhatsApp call that is easy because the number already matches the record; on a traditional call, a matching rule based on caller ID has to be defined.
Measurement
To measure the quality of CRM writing:
- Untagged call rate. Calls with no outcome tag. The target should be near zero.
- Share of "other" tags. If high, the tag list is inadequate.
- Task creation rate. What share of handed-off or unresolved calls became tasks?
- Task closure time. Are the created tasks actually being done?
The fourth is the most neglected. If the agent creates tasks but tasks never close, the problem is not the agent but the operation — and in that case every task the agent creates is just debt added to a queue.
You can review the CRM and data infrastructure side on the data and CRM page.