Treat AI like an intern: what NVIDIA AI Amplified confirmed about how we build FieldTaskora

By Mark Kuch, CTO, FieldTaskora

Last Wednesday I spent the afternoon at AI Amplified, hosted by NVIDIA and MMT at Evoqé Events on Queen Street in Melbourne. Much of the program, from humanoid robots to autonomous mine trucks, is a long way from field service software, but the same questions kept coming up. Who is accountable when AI produces something? How do you know what it did? What does it cost? And will the people on the job actually use it?

Those are the questions we've built FieldTaskora around. Our software uses AI to help field teams draft quote emails, estimates, method statements (RAMS), inspection checklists and reports. So it was a good afternoon: engineers from NVIDIA, Atlassian and Orica arrived, in their own words, at the principles we already work by.

Accountability doesn't move to the machine

The agentic AI panel, hosted by Raul Kumar of Nexus42, opened with a show of hands. About 60% of the room said they were working with AI agents. Fewer knew what their agent had done yesterday. When he asked who would trust an agent to contact their largest customer with no human in the loop, one hand went up.

Pushyami Gali, a senior engineering manager at Atlassian who leads engineering on Jira, gave the line I've been repeating since: "treat your AI coding agents as your intern." Even when an agent writes the code, she said, the developer who merges it and the team that owns the feature are accountable for what ships. Handing over the work doesn't hand over the responsibility.

That's exactly how we use AI at FieldTaskora. The AI does the intern's job: it gets a solid first draft of a quote, a RAMS or an inspection report in front of someone who knows the work. A person reviews it, and it goes out under that person's name. A RAMS or an inspection report gets relied on, for safety or in a dispute, so it needs an owner, and the owner is a person.

The moderator quoted a line from John McGiffin, CEO of Natural Velocity, that agents "must only know what they must know" and "only do what they must do". John uses a RACI model (who is responsible, accountable, consulted and informed) to decide who signs off on what an agent may do. He'd keep agents out of situations where empathy is involved or where a decision could affect a person's life. Many safety decisions on a job site fall into that category. That's why, in FieldTaskora, AI sits on the drafting side, not the deciding side.

If you can't trace it, you can't trust it

Haritha Thilakarathne, a senior developer relations manager at NVIDIA, made a point every engineer should hear. Software fails loudly, but an agent can be slightly wrong at step two and you won't find out until step nine. His advice was not to wait: "get your guardrails right, get your tracing and logging", early in development, and build your evaluations before you build the thing.

Rad Kanapathipillai, head of engineering at DX1, put the questions any regulated business should ask of its AI. Can you trace and monitor everything the agent is doing? Who is making the decision? What data is it exposing?

For a field record, being traceable is the whole point. If a report says a defect was found, you should be able to see where that came from, who checked it and when. We treat that trail as part of the record, not an add-on. Pushyami summed up Atlassian's security stance as "close by default whether it's agent or humans". We think the same way: people see what their role needs and nothing more.

Right model, visible cost

Haritha said the most common mistake he sees is assuming you need the most powerful frontier model for every problem. Using one for small, routine tasks is, in his words, like "killing a fly using a sledge." He described an enterprise experiment where requests were routed between a frontier model and smaller ones. It cut cost by 73% while keeping 93% accuracy. That's his figure from one example, but the principle stands: match the model to the job.

Rory Vining, AI field application manager at Orica, came at it from the user side. Orica's Citizen Developer program teaches staff what a token is and what a big query actually costs the business. Our view is that field businesses run on tight margins, and they shouldn't have to understand tokens to know what AI is costing them.

Get your boots dirty

The physical AI panel was mostly robots and mining, but Rory's answer to what makes a project succeed applies to any software used in the field: "you really have to have the end user buy in." If you don't work with the people who'll use it, you'll put a lot of time and money into something that, once it's on site, "it's just not going to cut it."

This is where I think FieldTaskora has the right instincts. Our users are on roofs, in plant rooms and on building sites, not at desks, which is why we have a dedicated field app as well as the office dashboard.

Matthew Adams, founder of Tekt Industries & Augment Robotics, who builds safety products, put the trust question bluntly: "People are not a percentage. It has to be 100% safe." AI outputs are probabilities. His approach is to wrap them in rules and constraints and to "constrain your promise", so the product only claims what it can reliably deliver. That's the bar we hold our own AI features to. Johan Barthelemy, a senior AI scientist at NVIDIA's AI Technology Center, said about 90% of the effort in building physical AI goes into the data: "garbage in, garbage out." So our AI drafts start from what the team actually captured on the job, not a blank prompt.

The gap is in the doing

John McGiffin said Australia sits ninth last for enterprise AI adoption, even though we're near the top of the world for using ChatGPT. Paul Mason, CEO of Robotics Australia Group, said we rank fourth for published robotics research but 75th at commercialising it. One panellist gave the example of Vixia AI, a two-person company about a year old that has already signed contracts in Singapore and Japan.

The gap isn't access to AI. It's putting it to work on real jobs, with real accountability, and that's exactly where FieldTaskora sits.

How we build, and what I'd tell anyone starting

AI drafts, people sign off. A named person stays accountable for every record that matters.
Build the trail first. Logging, review and evidence go in at step one, not step nine.
Right model for the job. Keep the cost of AI visible and predictable.
Design with the people on site. If it doesn't work in the ute, on the roof or at the switchboard, it doesn't work.
Start with the problem. As Pushyami advised, check whether plain automation would do, write down what the AI should and shouldn't do, and think backwards about what could go wrong.

Thanks to MMT, NVIDIA, Nexus42 and the speakers for a useful afternoon. If you run field teams and want to see what AI that drafts the paperwork, with your people still in charge, looks like in practice, get in touch.

Event partners

Organiser

Hosts