AI for finance

Before you put an agent on it

An agent sitting on unreconciled data does not give you fewer wrong answers. It gives you confident wrong answers, faster.

Most of what is being sold as finance AI assumes the boring work has already been done: that the chart of accounts matches how the business is run, that master data is maintained in one place, that a definition means the same thing in two reports. Where that is true, these tools are genuinely useful. Where it is not, they industrialise the confusion.

So we start where we always start. Fix the data and settle the process, then automate the parts that are repetitive and rule-based, and only then hand judgement to something that learns. That order is not caution for its own sake — it is the difference between a pilot that scales and a pilot that quietly gets switched off.

Where it actually works today

Invoice exceptions and coding suggestions that improve as corrections are made
Supplier and internal queries answered from your own documents rather than from a mailbox
Reconciliation investigation — proposing the likely match, leaving the decision with you
First-draft variance commentary, written from the same source as the figures
Collections prioritisation based on behaviour rather than on age alone

What we will not do

We will not put an agent in front of a process nobody can describe, and we will not automate a judgement that has no owner.

How to start

You do not need a full blown strategy from the start. You need one process, a baseline, and somebody willing to say afterwards whether it actually helped.

1. Find the friction, not the technology.

Two or three sessions with the people doing the work, listing where the hours actually go and where the same judgement gets made over and over. We score what comes out on three things: how often it happens, how much of it is judgement rather than rule, and what it costs when it is wrong. The best first candidates are high in volume, low in consequence, and already have somebody who can check the answer.

2. Check the ground before you build on it.

For each candidate we look at what it would have to read, whether the definitions are agreed, whether the master data behind it is maintained in one place, whether there is one source of truth or four. This kills the wrong candidates in days rather than in months, and it is the same work as the data health check, so it costs nothing extra if you have already done that.

3. Run one, small, and measure it against a person.

One use case, four to six weeks, with the human baseline measured before anything is switched on. A named owner in your team. And a written rule for what the tool may decide by itself, what it may propose, and what it must never touch. Without a baseline you will only ever argue about whether it feels faster.

4. Decide honestly, then keep it or stop.

If it did not beat the baseline we say so and you stop, which is a cheap outcome after six weeks. If it did, the work is to write down the controls, train the team, put it into the close calendar like any other step, and hand it over. The same rule applies as for everything else we do, it has to run without us.

Supporting the Finance Shift

The right expertise at the right time.