What an AI consultant actually does, and what to ask before you hire one
"AI consultant" can mean almost anything, from someone who's handy with ChatGPT to someone who rebuilds how your business runs. A good one works out where AI and automation will actually save you time and money, picks the right tools, plugs them into what you already run, sets the rules so it's used safely, and gets your team using it. The great ones build it too. Here are the five parts of the job, and the question to ask about each one before you hire anyone. Including us.
The five parts of the job
1. Finding the work worth doing
Ask them: "What will you measure before and after?"
Most businesses start in the wrong place. Someone's seen a slick demo, someone's had a play with ChatGPT, and suddenly the question is "what can we do with AI?" The better question is "what's eating our hours?" A good consultant starts on your floor, looking for work that's repeated, done by hand and big enough to measure. In most businesses that's quoting, intake, scheduling, the handover between the field and the office, and the paperwork that turns a finished job into an invoice. Plenty of what gets called an AI problem turns out to be a workflow problem, and AI only goes in where a step needs judgement a simple rule can't handle. Before anything is built, a good consultant measures the hours, the errors and the days it takes to get paid. If nobody measures the before, anyone can claim the after.
2. Putting AI to work, safely
Ask them: "What happens when the AI gets it wrong?"
Generative AI drafts, summarises, sorts and pulls information out of documents. Agentic AI goes a step further and takes action: it reads an email, looks something up and updates a record. Both are useful. Both get it wrong sometimes, usually with total confidence. So every AI step should sit inside a process, with a person signing off wherever a mistake would cost money or trust, and a log of what the AI did. An agent that can touch your systems gets one job, an off switch and an undo button. A good consultant will name the approval step and the person responsible for it. If the only answer is that the AI is accurate, keep asking.
3. Picking the tools and plugging them in
Ask them: "Who builds it?"
Copilot, ChatGPT, Claude, Gemini and a growing list of models hosted here in Australia all do similar things, at different prices, with different rules about where your data goes. The choice is rarely about which one is smartest. It's about what it needs to connect to, what it costs each month at your volume, and whether your team will actually use it. Then it has to be plugged into your practice software, CRM, accounts or rostering system. That plumbing is usually more work than the AI itself. The firm giving the advice should be the one that builds it. If the build is handed to "a partner" or "your IT provider", ask who is accountable when they disagree.
4. Governance and risk
Ask them: "Where does our data go?"
If you use AI on customer, patient or resident information, you need a short policy your team can actually follow, data handling that meets the Privacy Act, and a record of what the system did. From 10 December 2026, privacy policies also have to say when a computer program uses personal information to make decisions that significantly affect people. It sounds heavy, but for most businesses it comes down to a one-page policy and a log, as long as the system is built to keep the log. Ask for a list of every service your data touches, where each one is hosted and whether it is used for training. If the answer is vague, the homework hasn't been done.
5. Bringing your people along
Ask them: "Who owns it, and who runs it once you've gone?"
A system nobody uses gives you nothing back. The people doing the work need to know what the automation does, where it hands back to them, and how to change it when the business changes. Training should happen on the system built for them, using their own work, so they can run it from the first day. At the end, you should own the code and the data, and your team should be able to run the system day to day. Check the contract for the word "licence".
Three things to watch for
- The same product in every answer
That usually means a reseller, which is fine as long as you know.
- A report with nobody to build it
A list of opportunities leaves the problem where it was.
- A plan that starts by replacing everything
Replacing your core system on day one is the riskiest way to begin. Fix the workflows costing the most time first, then decide about the core system with real numbers in front of you.
Where to start
Not with a tool. Walk the floor for thirty minutes with someone who can build, and show them what's done by hand, who does it and how often. Start with the dullest, most repeated job in the building. It's never glamorous, but that's where the hours are hiding and where the proof turns up first. Halcyon does this on site from its bases in Coffs Harbour and Newcastle, and by video anywhere.
Frequently asked questions
Do I need an AI consultant or a software developer?
If the problem is a repeated manual job, you need someone who can map the process and build the fix. A lot of AI consultants stop at the recommendation, and a lot of developers build what they are told without asking whether it is the right thing. Look for someone who does both. At Halcyon, the consultant who maps the process also builds the fix, so the advice and the build come from the same team. See AI consulting.
How long does an AI or automation project take?
Reviewing the work takes one to two weeks. A first working automation is usually live four to six weeks after the build starts. A bigger system with several parts goes live in stages over three to six months, and you're using each stage before the next one starts.
Is my data safe with AI tools?
It depends on the tool, how it's set up and what you put into it. Business and enterprise versions can be set not to train on your data and to keep it in Australia; free consumer versions generally can't. Identifiable customer, patient or resident information should never go into a free consumer tool.