What Was the Original AI Sales Coaching Tool?
About a year into building Innovative Automations, the team put together a first-generation coaching tool to help outside sales reps get feedback on their outbound calls. The calls were already being recorded and transcribed through their dialing platform, so they built a custom GPT loaded with their sales framework. Reps or managers could feed in a call transcript and receive structured feedback: what was missed, what could be improved, recommended next steps, and follow-up items.
The impact was immediate. The team had previously paid an outside sales coach to do this work on an ongoing basis. That budget was redirected to hiring additional team members, meaning AI expanded headcount rather than reducing it.
Why Did the Single-Call Review Model Fall Short?
The original tool had a structural limitation: it was a point-in-time review. A rep had to remember to download a transcript, choose a call worth reviewing, and manually submit it. In practice, this meant the tool got used reactively, typically when a call went poorly or something stood out as a problem, rather than consistently.
During sales meetings and huddles, managers would remind the team to use it, but busy schedules got in the way. There was no aggregate view of performance across all calls, no trend data, and no automated delivery. The team wanted something that ran on its own and gave a fuller picture.
How Does the Rebuilt AI Sales Coach Work?
The rebuilt system connects directly to the dialing platform through an API (the platform has since added MCP functionality, though the team developed against the API). From there, it pulls two categories of data: performance metrics and call transcripts.
What Metrics Does It Pull and Analyze?
The system looks at dials made, conversations reached, how many of those conversations were with decision-makers, and how many converted into booked appointments. Critically, it compares today against the prior week and the week before that, so managers can spot trends rather than just snapshots.
Each conversion step points to a different problem. If the dial-to-conversation rate is low, it likely indicates a list quality issue. If conversations are happening but decision-maker contact is low, that points to a scripting or gatekeeper problem. If decision-maker conversations are not converting to appointments, that reflects how well the rep is communicating value.
How Does It Analyze Call Transcripts in Aggregate?
Rather than reviewing one call at a time, the system pulls all transcripts from a defined period (a day, a week, or two weeks) and analyzes them together against the sales framework. The output includes specific coaching insights, a breakdown of what is impeding performance, and manager action flags with recommended steps the manager can take to support their reps.
Because the analysis is based on transcript data rather than manager opinion or rep self-reporting, it removes bias from the review process. The team is clear that AI is not perfect, estimating accuracy has improved from around 30 to 40 percent a year ago to 80 percent or more today, but they treat the output as a starting point for discussion rather than a final verdict.
How Is the Report Delivered Each Morning?
The automation runs on a schedule at 7:00 a.m. each day. It reviews the previous day and the prior week, generates a PDF report, and delivers it to both individual email inboxes and the team's Microsoft Teams channel. When the morning sales huddle starts, the report is already there.
This removes the dependency on any individual to remember to run the analysis or pull the data. Even if a manager is out or a huddle is skipped, the report exists in the archive and is available to review at any point.
What Other Teams Could Use This Approach?
The framework is not limited to sales. Any team with recorded and transcribed calls, such as a customer service group, can apply the same structure. The requirements are a system that stores transcripts and exposes an API, a defined standard to evaluate against, and an AI model configured to produce the output format you want. From there, scheduling the automation handles the rest.
The team estimates the full build took a couple of days. If you are interested in applying something similar to your own operations, you can reach out to the Innovative Automations team at ideas@innovativeautomations.ai.