How we reduced after-call note-writing time by over 89%

Agents were spending time after each call writing notes. If a manager wanted to understand what had happened, they often needed to listen to the entire recording.

We built an AI summarisation tool that automatically summarises each customer interaction and stores the summary for later reference. It reduced after-call note-writing time by over 89%.

My contribution drew on systems engineering: connecting services, tracking processing, handling failures, and making the result available where people needed it. We used GCP, n8n, and language models to build the workflow.

Start with the work people need to do

For an agent, the task continued after the customer hung up. They needed to record the conversation so someone else could understand the request, the actions taken, and anything left unresolved.

That context mattered to managers too. Listening to a recording could answer their questions, but required working through the call to find the relevant details.

Summarisation gave both groups a quicker starting point. Agents could review generated notes, and managers could read a summary before deciding whether they needed to listen to the original call.

The reduction in note-writing time was a concrete outcome. Storing the summary also made the conversation easier to review when another colleague needed it.

Keep the customer journey connected

A customer’s latest interaction rarely tells the whole story. They may have contacted us before, received an explanation, or been waiting for an action. Someone picking up the next conversation needs that history to avoid making the customer explain everything again.

Summarising and storing interactions helps colleagues follow that journey. They can review what the customer asked, what had already happened, and what still needs attention across conversations.

A useful summary needs to preserve those distinctions. A short paragraph that misses an unresolved issue can leave the next colleague with an incomplete picture, even if it reads well.

Build the workflow around the model

The workflow needs to collect the right conversation material, coordinate processing, generate a summary, and store the result. Voice introduces a transcription step before summarisation. Each dependency creates another point where processing can stop or return something unexpected.

My systems background makes me pay attention to those boundaries. I want to know what a component received, what it produced, and which step is responsible for continuing the work.

A generated summary is useful only when it reaches the right destination and remains associated with the correct interaction.

Make unfinished work visible

Receiving an event, generating a summary, and storing it are separate stages. Success at one stage does not confirm that the whole workflow completed.

In the pipeline work, we used lifecycle state, audit logging, retries, and scheduled reconciliation to track processing and identify stale events. State records where processing reached. Audit logs help reconstruct what happened. Retries can recover from temporary failures. Reconciliation helps find work that remained unfinished.

The person supporting the workflow needs a way to investigate missing results without manually tracing every interaction through several services.

Check the summary as well as the execution

There are two checks around generated output: whether it has the expected structure, and whether it accurately represents the conversation. Structured output and validation help make missing or unexpected data easier to detect before the workflow continues.

Accuracy needs separate attention. A response can contain all the expected fields while omitting a customer’s concern or making an action sound complete when it is still pending.

The original conversation remains the source for checking those details. Summaries reduce the material colleagues need to work through and give them a starting point for deeper review.

I judge the result against the original task: can an agent finish their notes sooner, can a manager understand the interaction, and can the next colleague follow what happened before? The reduction of over 89% in after-call note-writing time gives us one measurable answer.