HEADCOUNT & CODE · A Building Tech Teams publication
01 · THE PATTERN
Costs. Productivity. Revenues.
Three articles this week. The pattern consistent across all of them is ROI. And it ties in well with this week’s podcast guest Adam Witanowski, who said “what you don’t measure you can’t change”, and that when you can measure the uplift one person brings, “they’ve paid for their salary multiple times over”.
AI conversations are becoming a cost and value question. The question is no longer whether companies are using it, but whether they can measure some form of ROI.
The three big articles this week cover Canva’s unexpected AI costs that changed their forecasts, Westpac using AI to save 150,000 hours of processing time, which it then reinvested in customers, and Atlassian increasing profitability following its restructuring earlier this year.
It’s a theme this week, and it’ll be an ongoing theme as companies move past the hype and want to see tangible results from AI in practice across their teams, their customers, and ideally their P&L.
And it’s the same question I have been asking while working with Claude. Some weeks the time I put in doesn’t actually lead to any significant return or value. I’m currently writing off that wasted time in the “learning” bucket, but at some point the outputs need to significantly outweigh the inputs, and hopefully at that point I sleep a little more, argue with Claude a little less, and see a greater return on investment.
02 · THE MOVES
1. Canva provided a blueprint for controlling model costs
Canva cut its annual revenue growth forecast from 30 per cent to around 20 per cent, and the reason is the one every company building on AI, or building with AI as a significant part of their operating systems, needs to be aware of. In the Canva example, customers used the AI suite more than the cost model assumed, and the company was leaning too heavily on expensive frontier models to serve them.
To combat this, Canva spent three months rebuilding its AI infrastructure around its own models and cut the cost of processing a single AI task by almost 90 per cent. Its in-house style-transfer model now runs 23 times cheaper than comparable frontier models. (Startup Daily)
This has been a conversation I’ve been having almost weekly. If you’re building, you'll need to build model-agnostic systems, maintain the flexibility to route each task to the right model, and know when running your own or open models beats using frontier models.
And only two days ago, Meta made this easier for companies not going to the lengths of building their own models. Meta released Muse Glimmer, a 30-billion-parameter open agentic model, compressed for single-consumer GPU deployment. (Meta AI Research) That’s a lot of words to understand. But essentially, it’ll allow users to run high-quality open models themselves rather than using Frontier (or Chinese) models where appropriate. The build-your-own path Canva spent three months engineering keeps getting cheaper and more accessible to run for those without the engineering capability. Frontier models for the hard tasks and open local models for the volume work is becoming a default.
Whilst the headline is negative, Canva is simply ahead of most companies on this curve, and that means their mistake play out in public. But they are aware of the problem and have been fixing it. Revenue still grew 25.2 per cent to US$921.9 million for the June quarter, so I read this as a short-term negative press on the cost side, with a long-term solution that provides Canva with the best platform to continue scaling.
For leaders: the question this story hands leaders is ownership. Who in your company owns AI unit economics? Someone needs to own cost per task the way someone owns cloud spend, with the authority to route/architect the work to the right model.
For tech professionals: companies rebuilding themselves with AI as a core function of their operating systems will be looking for people with this knowledge around AI architecture and unit economics. When you can tie your work directly to cost savings or ROI, your salary negotiations become significantly easier.
2. Westpac has five AI agents inside real lending decisions
Andrew McMullan, Westpac’s chief data, digital and AI officer, was clear about what matters: these are running in production systems, not prototypes. Five specialist agents now work within the bank’s mortgage and credit card processes, classifying payslips, extracting data, running calculations, and checking policy before handing the output to a banker for review. The scale is significant, more than 32,000 payslips and 1.5 million transactions every week, and more than 150,000 hours of processing time saved and reinvested in customers. (iTnews)
The achievement itself is noteworthy and worth a read, but what stood out to me was the time to production - significantly shorter than most enterprises. Westpac has shrunk the time to get a new agent into production from six months to about four to six weeks. When agent deployment moves to a monthly rollout, the operations team beneath it changes, customer delivery speeds up, and this is happening in one of the most regulated environments in the country.
For leaders: the metric to learn from is the deployment speed. Six months to six weeks is the difference between AI as a project / POC and AI as an operating rhythm. It shows whether you are truly investing in AI as a core function of your business or just building slices around the edges.
For tech professionals: “the banker reviews the agent’s output” is the point I’ve been making for weeks. AI + Human is the real value play for most organisations at the moment. The person with industry knowledge who owns the exceptions and the policy calls is vital to the workflow; use your knowledge, overlay AI, and you can be one of the most valuable parts of the workflow.
3. Atlassian cut 1,600 people in March. This week it posted its first operating profit in two years.
Issue #6 of this newsletter ran the Financial Times finding that the market punishes companies that blame AI for layoffs, marking their shares down almost 10 per cent against the Nasdaq in the month after the announcement. (TechCrunch) Atlassian just delivered the biggest bounce-back yet. Five months after cutting 1,600 people, 10 per cent of its workforce, in a restructure framed around an AI pivot, the company reported June quarter revenue of US$1.77 billion, up 28 per cent, and its first operating profit in more than two years. And the market added more than a third to the share price overnight. (Bloomberg)
That March cut is a part of the Index’s 3,610 EXPLICIT count, which is one of only five Australian events where the company itself named AI as a driver of layoffs. And now we’re reading the first public result of what the restructure delivered.
Personally, I expect Canva will have a similar story in the coming quarters; the pattern is a company investing heavily in AI, finding an issue, making changes, the market reacting negatively, those changes eventually leading to positive results, and the company being rewarded on earnings/share price.
Because these companies are ahead of the curve in AI investment, they are the ones learning in public about infrastructure, team size, and team makeup. The rest of us get to watch and learn from the sidelines.
For leaders: one luxury most companies have is being able to learn from those ahead of the curve. We’re not all the same size; we’re not all product companies investing in AI, but there are still lessons we can learn from both Canva and Atlassian, even if it’s on a smaller scale.
For tech professionals: the people who rebuilt Atlassian’s teams through the pivot now hold some of the most transferable experience in Australian tech. If your company is mid-restructuring, that experience is the asset you are building. Understanding how a leaner team works, and what a team truly investing in AI looks like, gives you experience that can set you up for future opportunities.
03 · FROM THE INDEX
One read a week from the AI Jobs Index, the live tracker I run at aijobsindex.com.au.
A third of the job ads the Index scans now mention AI. The diffusion index hit 32.99 per cent this week, 924 of 2,801 technology and data-adjacent Australian ads. It has climbed every data pull since the Index launched.
The number of job cuts is staying the same. Nothing new has been announced, and after the Financial Times’s market-penalty finding from last issue, staying quiet about AI restructures has never made more commercial sense.
What does this mean: the reshaping of roles is running through job descriptions.. AI language is spreading across job ads. The money is being allocated to the reshaping of teams and roles, and there is an expectation that AI will at least be a part of more and more job roles of the future.
04 · WHAT I’M BUILDING
A change this week: I wrote about it on LinkedIn. Tech Teams Tuesday is now Headcount & Code, and it arrives on Wednesday mornings - starting today!
The reason is not pretty. Four straight weeks of building the issue past midnight on a Monday was unsustainable. Tuesday is my writing day, so the newsletter now gets written on Tuesday rather than Monday nights. And the new name says what this is actually about: the two things every tech team is made of, the headcount and the code.
Building a newsletter has similarities to building tech or a business: we all start with great plans, but sometimes we need to acknowledge what’s not working and iterate.
05 · THIS WEEK ON THE PODCAST
Adam Witanowski is an AI Architect and one of the few people in the country who has actually built AI in production at enterprise scale. He built an AI software delivery capability for a team of more than 300 engineers in a heavily regulated environment, 18 systems deep, including a knowledge graph across 3,600 repositories that shows the agents what their changes break downstream.
What we get into:
Back pressure. His term for the testing harnesses, CI signals and organisational context that tell an agent its code is not fit for your environment. Agents will generate code happily all day; back pressure is what makes it good code for your context.
Whole processes, not thin slices. Why bolting AI onto one sliver of a role returns five or ten per cent, and why the real gains come from re-engineering the process end to end.
The measurement problem. Companies have become good at measuring what a token costs but have stayed poor at measuring what the token adds. His line of the episode: what you don’t measure you can’t change.
My Favourite - How to interview an AI engineer now that the coding test is dead.
Why it matters: this issue is about AI landing on the P&L. Adam talks AI’s impact is not a technology problem, but a measurement problem.
Listen: YouTube · Apple · Spotify
06 · THE ASK
One question this week - is there an AI workflow inside your company where you can actually prove the ROI?
Cheers,
James
Headcount & Code is the weekly read from Building Tech Teams. Written by James MacDonald, MD of NTP Talent. The AI Jobs Index is a Building Tech Teams publication.