From 5 hours to under 2 minutes: a legal firm's lead report

Challenge

Somebody spent five hours a week building a spreadsheet. The firm processed hundreds of leads a week across more than a dozen executives, and the only way to see whether those leads were being followed up was to compile the report by hand: pull the records, reconcile them, chase the gaps, format the result.

Five hours is the visible cost. The invisible one is that a report built that way is always a week old, so nobody used it to intervene. It was a record of what had already gone wrong. Management reporting on call activity had the same shape, produced monthly from telephony data, and the same problem.

Approach

We started by separating the two things the manual process was doing at once: gathering data, and judging it. Gathering is mechanical and belongs in a workflow. Judging, deciding whether a follow-up record is complete, consistent and plausible, is where a person was actually adding value, and where the errors were.

So the design put n8n in charge of the gathering and put AI-powered validation in the judging seat, with the results going somewhere people already look rather than into a file someone has to open. Nothing about this required the firm to change how it worked, which matters: an automation that demands a new process usually gets one week of compliance.

Solution

Three systems, built in that order.

Automated lead follow-up tracking. An n8n pipeline pulls the week’s leads, applies AI-powered validation to the follow-up records, and produces the report that used to be assembled by hand. It covers more than 12 executives and the hundreds of leads they work each week.

Daily telephony dashboards. Nightly RingCentral data (call duration, outcome, timing) is consolidated through n8n and Google Apps Script into dashboards that refresh every day. That replaced a monthly report compiled by hand with a performance interface management can open whenever they want it.

A company-wide policy assistant. Staff were filing internal support tickets to ask questions that already had documented answers. We built a retrieval-augmented chatbot over the firm’s own policy documentation so those answers are self-service and grounded in the source material rather than generated from nothing.

Results

• The weekly lead report went from 5 hours to under 2 minutes, covering hundreds of leads and 12+ executives.
• Monthly call reporting became a daily-refreshed dashboard, so management sees call activity while it is still actionable.
• Policy questions became self-service through the retrieval-augmented assistant.

One honest note on the third item. The assistant was built to reduce internal support tickets and the firm’s read was that it did, but nobody instrumented the ticket volume before and after, so we have no figure for it and will not invent one. The first two results are measured. This one is reported as what it is.

What made it hard

The five hours were never really about the compilation. They were about reconciliation: records that disagreed, follow-ups logged in one place and not another, entries that were technically present and practically useless. Automating the assembly of bad records would have produced a fast, wrong report. The validation layer is why the number moved, and getting validation to agree with what an experienced person would have flagged took more iteration than the pipeline around it.

The telephony work had a plainer problem. Nightly data arrives on the telephony provider’s schedule and in the telephony provider’s shape, neither of which is negotiable, so the consolidation had to tolerate a late feed and a partial feed without producing a dashboard that quietly showed the wrong day.