- Datacom's CEO says AI token spend, about NZ$500,000 a month, is easy to count, while proving what agents deliver is the harder task.
- Reviewers who cannot say no turn oversight into a formality, so authority passes to the machine without anyone deciding it should.
- Leaders should track override rates, measure cost per successful task, and write down what agents may not decide before they go live.
A company can see exactly what its AI coding tools cost. It is much harder to see whether the work they produce can be trusted. That gap is the real management problem, according to Greg Davidson, CEO of Datacom. He spoke at the AWS Cloud and AI Day in Auckland, as reported by Reseller News on October 9, 2026.
Davidson put numbers on the easy side. Datacom has around 500 core developers and close to 1,000 people who code as part of their roles. He said adoption of AI tools across that group is about 50 per cent. Spend on tokens, the units AI providers charge for, is about NZ$500,000 a month.
Cheaper tokens, bigger bills
Davidson described two trends across enterprises. The price of each token dropped by 67 per cent. Total usage grew by more than 1,000 per cent. He did not name a source or time period in the remarks reported.
The arithmetic is simple. A buyer who saw both changes would pay about a third of the old price per token. They would use more than ten times the volume. The total bill would still be more than three times larger. Falling unit prices do not shrink a budget when the work handed to the machine keeps growing.
Davidson's own explanation is that the tasks are now more expensive to process, with better outcomes. This is consistent with his argument for counting outcomes rather than tokens.
How Datacom tries to make the work provable
Davidson described a method rather than a tool. Before an agent starts, a person defines acceptance criteria and tests. The agent must then show it is finished by passing those tests. It must also leave a record of what it touched.
He called evaluations "the unit testing in the AI world". A unit test is a small automated check that a piece of code does what it should. His point is that an agent you cannot evaluate "is just a demo".
The same logic applies to cost. Davidson wants to measure cost per successfully completed task, with quality evaluated. He said outside benchmarking shows OpenAI's GPT-6 Astra matching its closest rival on intelligence, while each task costs 57 per cent less. Reseller News did not report who ran the benchmark.
Readers should weigh one more fact. Davidson said Datacom is both a customer and a partner of OpenAI on AWS, and calls itself "customer zero". That does not make the figure wrong. It does mean the claim comes from a party with a commercial relationship.
The idea that matters: review can become a rubber stamp
The most useful part of the talk is not about money. The common view is that machines predict and humans judge. Davidson said that view has a flaw. A person who receives an agent's work to check must be equipped to judge it properly.
He listed seven things a reviewer needs. The first is time to think. The second is a view of what the model looked at and how sure it was of its answer. Third comes knowledge of the system's known failure modes. Fourth is authority to override, with enough career safety to use it. Fifth is independence to reach their own judgement. Sixth is a named path to challenge or dissent. Seventh is proof, in the form of override rates reported up to the board.
The test he offered is blunt: "If your reviewers never override, it's not a real review." In his words, that is a transfer of authority. This is the human side of the story. The engineer, analyst or claims officer who signs off on an agent's work carries the risk. If they lack time, context or standing to say no, the organisation has handed the decision to a machine without deciding to.
WebPulse's view is that this mirrors the audit trail in finance. A signature only means something if the signer could have refused. An override rate gives leaders a number that shows whether that is true.
Questions to put to your team
Davidson's advice maps to a short list of questions for any organisation paying for AI agents.
Which one workflow has a written goal and a one-month baseline? He says there is no correlation between usage and benefit. What are the three most frequent AI workloads, and what does each cost per successful task? Who owns the evaluations, and can an agent prove it met them? Is model choice made per workload, so you can move as capabilities and costs shift?
Then ask the hardest one. What are agents not allowed to decide? And what is the override rate on the decisions they do make? Davidson's advice is to settle those terms in writing before an agent is switched on.
The token bill will keep arriving every month. The better question is what it bought, and whether anyone was free to say it bought too little.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Reseller News.
