'Not going to be that cheap': why agents outprice offshore teams – HiPages

Jeremy Burton, CTO of HiPages Group, describes how deploying Salesforce Agentforce on a complex first use case taught him where AI agents actually deliver – and where to start.

When Salesforce announced its agentic platform Agentforce, HiPages CTO Jeremy Burton was one of the first in Australia to sign up and test it out. HiPages, a marketplace for tradies, was already running its sales and services operation on Salesforce. And it moved early enough on AI agents that Salesforce was still building features around them.

The first question Burton faced with the new platform – where do you start?

Burton wanted to fix something tradies felt directly. A new tradie signing up on a Friday afternoon couldn't win work until Monday or Tuesday. The hold-up was licence validation – HiPages checks every tradie's credentials against 34 Australian licensing authorities before they can claim jobs. The process sat in a queue, handled manually by the Trust and Quality team, with a turnaround of a couple of days.

"We saw a real opportunity to automate that," Burton says. "Take that from something that takes a person 10 or so minutes to do, and sits in a queue... and speed that up to almost real time."

The proof of concept was running within two to three months. Refinement took another three to six. More than 12 months in, Burton has a clear view of what Agentforce costs, what it saves, and where it delivers the most.

‍

The cost case

There were two challenges Burton faced with the project. There was no problem connecting the systems to Agentforce. The hard part was making the agent behave consistently.

"We needed a lot of determinism in this piece of work," Burton says. "LLMs are not deterministic, and it's hard to convince them to be." The solution was heavy negative prompting and months of iterative testing against historical data.

Salesforce has since shipped a partial answer to this problem. Hybrid Reasoning, part of the Agentforce 360 release in October 2025, combines LLM flexibility with a graph-based deterministic layer that locks critical workflow steps to defined logic. Burton hadn't had the chance to test it at the time of this interview. "It would have made the development time a lot quicker," he says, "because that was always where the challenge came."

The second challenge was the cost. When Salesforce first released Agentforce pricing, Burton looked at the figures and they didn't stack up. "It would have actually been more expensive to handle the workload there," he says. Salesforce revised the model – moving from per-conversation to a credits system – and the shift made the economics work.

After more than 12 months of development, the numbers are in. The licence validation process that once took a team member in the Philippines 10 minutes now costs HiPages 15 cents per execution in token costs.

"Regardless of how well priced your people are in a low-cost location," Burton says, "they're not going to be that cheap."

‍

Both sides of the ledger

The 15-cent figure covers token costs only. It excludes the Salesforce platform subscription, Agentforce credits allocation, implementation work with Salesforce partner j4rvis, and the ongoing human QA that remains part of the workflow. Anyone modelling this for their own organisation needs the full cost stack, not just the compute number.

But Burton is equally clear about what a human cost comparison misses. "When you're just looking at 15 minutes for a person versus 15 cents for an automated process, which one is cheaper – you're not taking into account all the other variables," he says.

Those variables include onboarding, sick days, surge capacity and bench time. An agent scales up and down at no incremental cost beyond processing. It works at full pace from day one.

There is also a management dimension. Burton frames agent oversight not as a burden but as a different kind of management work – closer to reviewing employee output than to monitoring software.

"You're not going to watch over their shoulder every single thing they do," he says. "But what you are going to do is look at the work they do, form a judgment on, are they doing a good job, and if not, what's the corrective action?"

Which raises a question worth sitting with: is managing a fleet of agents actually less burdensome than managing the people they replace? The evidence from HiPages suggests it might be – but nobody has been doing this long enough to know for certain.

‍

Where to start

Running an agentic workflow requires a different kind of oversight than running conventional software. When an agent breaks, it may keep producing output – just the wrong output. The people best placed to catch that are not engineers. They are the people who know what correct looks like.

"It's shifting to a world where it's people who understand the context, people who understand the data, who are actually in a much better place to QA it," Burton says.

That shift has staffing implications. HiPages is designing what Burton calls an Agent Manager role – someone whose job is to manage the agentic workforce the way a people manager manages employees: setting expectations, reviewing output, and taking corrective action when performance drifts. "It's not engineering heavy," he says. "It's process thinking."

On ROI, Burton's advice to other CIOs is grounded in what he learned through the HiPages deployment. "Where I'd start is in customer service," he says. "I think that's where the most ROI is."

Part of the reasoning is economic. Part of it is risk. A wrong answer to a routine customer query is recoverable. A billing error is not.

Burton's first project was a complex, multi-authority data validation problem – harder than it needed to be as a first deployment, but one that produced real cost numbers and a clear lesson about sequencing.

Jeremy Burton is Chief Technology Officer of HiPages Group. The full interview is available on YouTube.