How our categorisation engine cut manual work by 67%
We are busy implementing NineAnts for our first clients. Every implementation starts in the same place: the client's historical data.
Why categorisation comes first
Before our AI agents start forecasting, we feed them historical transactions from the client's bank accounts. Sometimes those come from Excel. Sometimes from old bank statement files or other systems.
Every transaction needs a category. The categories are what the agents learn from: patterns per customer group, per supplier, per tax calendar, per entity. A forecast built on badly categorised actuals inherits every one of those errors.
A recent implementation
Last week, during one implementation, we put our transaction categorisation engine to the test.
The client's treasury team had already categorised around 180,000 historical transactions themselves. Partly with an advanced Excel model full of keyword formulas, partly by hand. Years of daily work, and every one of those decisions shows how the team thinks about their cash flows.
That history gave us a benchmark. Our agents read the files, learned the team's logic and reasoning, and built a new rule set. Then we graded that rule set against all 181,031 decisions the team had already made.
The results
89% of transactions auto-categorised, up from 58% with the existing Excel formulas
99.3% agreement with the team's own historical choices
Manual categorisation work down to roughly a third
The rules are readable, so the team can open any rule and see why a transaction gets its category. They still check and manage the rules, but the repetitive daily work is largely gone.
What we take from this
Two things stand out for me.
First, the training material was there all along. Most treasury teams have years of categorised transactions sitting in their files. That history is enough to automate a large share of the work, in the team's own logic.
Second, this is a first version. Over time I think we can further optimise the coverage, but the first results are promising.
We are building NineAnts at lightning speed together with our clients. An agentic AI platform for treasury teams, designed to be the most user-friendly one out there. Curious how it could help your treasury team? Reach out.
