Notes

Field notes

Four short notes on working with AI, by Rodrigo Gutiérrez.

From Working with AI

Your first automation should be boring.

By Rodrigo Gutiérrez

The work that lasts rarely begins with a grand “AI transformation.” It starts with one small, frequent piece of work that people are genuinely tired of doing by hand.

Choose a task that happens often, follows a recognisable pattern and has a clear end point: preparing a weekly update, sorting incoming enquiries or drafting a first reply. The first version does not need to do everything. It needs to remove one useful slice of the work.

Pick the repeat

Look for work that happens weekly or more, not a one-off ambition.

Keep review in place

Let AI prepare the first pass; keep a human decision before anything important leaves the team.

Earn the next step

Run it for a few cycles, learn where it needs help, then expand only once the value is real.

Small wins are not small if they change a weekly habit. Once a workflow is reliable and embedded in the day, it becomes a foundation rather than another forgotten pilot.

Opportunity discovery

Map the process before you touch AI.

By Rodrigo Gutiérrez

An AI agent can be remarkably capable, but it does not know your clients, your exceptions or what good work looks like in your practice. You have to teach it the process first.

Before comparing tools or pricing pages, pick one task you completed this week that felt repetitive. Treat it as a process, not a tool-shopping exercise.

Pick one task

Choose something repeated often enough to matter, not your whole job.

Time it honestly

Estimate or measure the time it takes, so you can see whether change is worth it.

Write the walkthrough

Record the source, checks, decisions, software, first touch point, last touch point and output.

A client-meeting brief makes the point: calendar details, the recent email thread and CRM notes need to become one useful page. Mapping those handoffs is already most of the instruction an AI will need.

AI delivery

A wish is not an instruction.

By Rodrigo Gutiérrez

“Check my calendar and email, then give me a meeting summary” sounds clear to a person. To AI, it leaves too much to guess, and those guesses change from run to run.

Turn the process you mapped into four parts. This is the difference between a one-off prompt and a repeatable piece of work.

Role

Who is the AI acting as, and for what specific job?

Steps

Write the process in the same order you would teach a new colleague.

Decision rules

Replace vague directions with real conditions, dates, quantities and exceptions.

Output

Specify the result: a one-page brief, table, draft or another consistent format.

If the workflow needs calendar, email or CRM information, check the connectors before committing to a tool. Then test it using a question you already know the answer to. It is much better to find the gap in a safe test than in front of a client.

Governance field note

Australian hosting is only half the answer.

By Rodrigo Gutiérrez

Where a file is stored and where an AI model processes it can be different. When sensitive information is involved, a useful vendor conversation traces the whole path, not just the hosting claim on a homepage.

TRACE is a plain-language prompt for asking better questions of any AI vendor.

The processor. Which provider and subprocessors handle the information?

Region. Where is the file kept, and where does the AI read or process it?

AI training. Is information used to train models, and how long are interactions retained?

Consistency. Do the privacy, security, terms and subprocessor pages tell the same story?

Everyone named. Is there a real company and accountable team behind the tool?

Use this as a first screening conversation, then pair it with the privacy, legal and insurance advice appropriate to your organisation and the information you handle.

Tell us what you need built.

Send a short note on the problem and who it affects, and we will reply by email.

Email us