The problem
A Victorian services business — roughly 25 staff, around A$5M turnover — ran most of its sales and support through the phone. Reps were good on the call. The trouble was everything that happened after it.
A rep would finish a call, mean to log the follow-up, then take the next call. The note never got written, or got written as a two-word placeholder. Upsell signals — a customer mentioning a second site, a contract renewal date, a competitor they were unhappy with — lived in the rep's head until they didn't. Accounts that sounded frustrated on a Tuesday weren't flagged to anyone before they churned.
The owner estimated the team took between 400 and 600 calls a month. Each one carried perhaps five minutes of note-writing that should have happened and often didn't. The cost wasn't only the lost admin time; it was the opportunities that were spoken aloud and then forgotten. Nobody had a reliable picture of what was actually said across all those calls.
What we built
We ran a six-week Pilot, scoped after a short readiness review of their phone system and CRM. The build does three things after each call:
- Transcribes the recording into text, speaker-separated.
- Writes a structured summary into the CRM — a short synopsis, the key points, and a list of concrete action items — attached to the right contact record automatically.
- Flags likely sales opportunities and at-risk accounts into a review queue, with the specific quote from the call that triggered the flag.
The flag is a suggestion, not an action. A rep or the sales lead sees each flagged item alongside the evidence and decides whether to pursue it. The system never emails a customer, never changes a deal stage, and never closes an account on its own.
How it works
When a call ends, the recording is passed to a transcription model, then to a language model that produces the summary, the action items, and any opportunity or risk flags against a fixed set of categories the team defined with us (renewal mentioned, second location, pricing objection, competitor named, frustration signal). The output is written back to the CRM through its API and lands on the contact record within a few minutes of hang-up.
Flagged items go to a daily review queue rather than straight into a pipeline. Each flag shows the verbatim snippet that prompted it, so the reviewer can judge it in seconds and dismiss the noise.
Because calls contain personal information, privacy shaped the design from the start. Recording happens only with consent captured at the top of the call, in line with the team's obligations under the Privacy Act 2026. Transcripts and summaries are stored in their own tenancy with access limited to the sales team, retention is time-boxed, and a customer can ask to have their recording and derived notes removed. We treated the consent step and secure handling as prerequisites, not add-ons.
The results
Over the Pilot, at an estimated 400–600 calls a month:
| Before | With the miner | |
|---|---|---|
| Note-writing time per call | ~5 min (when done) | near zero |
| Admin time saved per month | — | ~30–40 hours |
| Opportunities surfaced vs manual logging | baseline | +~30% |
| Cost per call analysed | — | ~A$0.15 (model + transcription) |
| CRM notes on completed calls | patchy | consistent |
Roughly 35 hours a month of note-writing came off the team's plate — worth on the order of A$2,000 a month in recovered time at their loaded rates, and more valuable as attention returned to the customer instead of the keyboard. Opportunity flags surfaced around 30% more actionable items than the team had been logging by hand. A modest but real share of those flagged opportunities converted once a human followed up — enough that the owner extended the tool past the Pilot rather than because we promised a headline multiple.
How the team owns it
The sales lead owns the flag categories and can adjust them as the business changes — that's a settings change, not a redevelopment. The daily review queue is part of the team's morning routine, and dismissing a bad flag is one click, which keeps the reviewers honest about quality. We documented the consent script, the retention settings, and the removal process so the team can answer a customer's privacy question without calling us.
The design keeps the judgement where it belongs. The tool listens, writes the note, and points; a person decides what to do about it. If you're weighing something similar, our AI automation work and the readiness audit are the usual starting points — or just get in touch.