Practical AI in Business Operations: the series

Most of the AI conversation in Australian business is about the big swing: the customer-facing assistant, the strategy deck, the platform decision that takes a year to unwind if it goes wrong. This series is about the other end of the scale. Timesheets that don’t get filled in. Leave requests that sit in someone’s inbox. Contracts with the wrong version number in the filename. The daily “what happened yesterday” question that someone answers by hand every morning.

None of it is exciting. All of it costs money, and it costs the same money every week whether anyone measures it or not. The return on fixing it is usually better than the return on the flashy project, mostly because the problem is small enough to define properly and the failure modes are cheap.

Every post in this series is based on something we run. Some of these automations we built for ourselves at PicNet and have been living with for long enough to know where they annoy us. Others we operate for clients. Where something didn’t work, or where we changed our approach after the first version, I’ll say so. I’d rather write about a system with known limits than a demo that behaves well on stage.

The approach behind all of it has three parts, roughly in order of preference.

Deterministic automation comes first. If a rule can be written down, write the rule. A scheduled job that checks who hasn’t submitted a timesheet and sends a message is not AI, it is a cron job and a query, and it will still be working in five years without anyone paying per token. A surprising amount of what people currently want AI for is this.

AI goes where judgement or language is involved. Reading an unstructured email and working out that it’s a leave request for three days in October. Summarising twenty tickets into something a manager can read in a minute. Deciding whether a document looks like the signed version. These are jobs where the rules can’t be fully enumerated, and where being right most of the time is genuinely useful because a human still sees the output.

A human sits in the loop wherever money or people are affected. Approving payments, changing someone’s pay, sending anything to a client, making a decision about a person’s employment. The automation can prepare the work, gather the evidence and present the recommendation. It doesn’t press the button.

The series is written for IT managers and executives who have been asked what their organisation is doing about AI and want an answer that survives contact with an operations team. If you’re looking for a first project with a defensible business case and a small blast radius, this is where I’d start.

PicNet builds production AI systems for Australian organisations. Talk to us about what a first project could look like.

Posts in this series

• Reconciling timesheets, leave and payroll automatically

Tagged: #business-automation #back-office #workflow-automation #ai-adoption #operations