Paying 100+ salespeople the right commission, on time, every month.

The problem

The client pays commission to more than 100 salespeople under an annual policy. Each salesperson's contribution to sales and revenue is measured against a personal target, and the rules change every year.

Measuring that was a manual job. Data came from 52 sources, sent by individuals and reporting teams via email, Teams, and SharePoint, and stitched together in Excel and Access on one person's laptop. One late file held up the whole run. One wrong report meant salespeople were paid the wrong commission. As the policy grew more complex, with rules depending on other rules, parts of it became nearly impossible to build in a spreadsheet.

What we built

We rebuilt the process in three phases, each one useful on its own.

Automate what was there

First we rebuilt the existing manual process in code, so the new output could be checked against the old one. We put everything under version control with a proper change process, so a new policy year became a reviewed change rather than a new set of spreadsheets. Once the data was in, processing went from days to hours.

Take the data from source

We replaced all 52 manual spreadsheets with an automated data pipeline. The platform collects the data itself, directly from the client's data warehouse, ERP systems, and scheduled file transfers. The data arrives on time and comes from the system that holds it, not from someone's copy. This is the data foundation we argue every AI project needs, and it is the same pattern as our payroll work: many sources, one shared rule set.

The pipeline runs as containerised jobs on serverless AWS infrastructure, so volume is no longer limited by the machine it runs on. Every change is tested and deployed automatically. The processed data feeds back into the business's reporting at a level of detail the old process could never carry.

Ask the data

With trusted data in place, we built an assistant on AWS Bedrock. We designed it around this business and its commission rules, including how those rules depend on each other, so it can explain how a number was reached rather than just repeat it. Salespeople and sales managers ask questions in plain language to check their numbers, spot anomalies and find opportunities. It is in the final stage of testing, with a limited release to power users.

Where it got to

The platform has been live for six years and a Chisl team still runs and extends it. The monthly run now takes 3 days instead of 12. Disputes and queries from the sales force have dropped because every number can be traced back to its source and to the rule that produced it. When sales priorities shift, the business changes the policy and sees the effect in the next run. The next step is agents that act on what the assistant finds, not just report it.