$1,000 per day. That's what Tanda, an Australian workforce management company, budgets for AI coding tools - and their CEO Alex Ghiculescu says the number hasn't been hit yet.
In a recent talk, Ghiculescu makes an argument that sounds obvious but has radical implications: every software development methodology ever invented - agile, waterfall, Shape Up - exists primarily to rate-limit product ideas to a pace that engineering teams can handle. AI code generation removes that rate limit entirely. So what happens to the processes, roles, and organizational structures built around a constraint that no longer exists?
The Three Types of Work That Suddenly Become Possible
Ghiculescu identifies three categories of previously inaccessible work that AI coding opens up:
"Olympics problems" are ideas that resurface every few years but get rejected because of accumulated technical baggage. Every company has these - features that would take a senior engineer three months of untangling legacy code to build. With AI tools like Claude Code, the baggage becomes manageable because the AI can process and navigate complexity that would exhaust a human.
Architecture paralysis is the blank-page problem. Even experienced engineers sometimes cannot figure out where to put the first line of code for a new feature. AI eliminates the starting friction by generating an initial structure you can react to rather than create from nothing.
Gruntwork - the tedious, labor-intensive tasks traditionally assigned to junior developers - becomes trivially cheap. Database migrations, boilerplate endpoints, test scaffolding. The AI handles these without complaints or morale costs.
The Budget Philosophy
Ghiculescu's stance on AI spending is unusually aggressive for a mid-sized company. His explicit position: "The point of the budget is to decrease the friction around spend, not to constrain spend." He observed engineers making pull requests to optimize prompts that cost five cents to run, which he sees as deeply irrational when engineering time costs orders of magnitude more.
His implementation path is concrete: test Claude Code on your actual codebase to fix a real bug, run a company hackathon to reset expectations about what's possible in a day, put every willing employee on Claude Code's Team plan without making them ask, and set budgets generous enough that no one hesitates before using the tools.
What's Missing From This Picture
Ghiculescu opens with admirable honesty - "Nobody knows anything" and "Don't trust experts, particularly me" - but the piece has notable blind spots. There's no discussion of code quality at scale, architectural debt from AI-generated code, security review processes, or what happens when the AI confidently builds the wrong thing. The talk also sidesteps the workforce question: if AI handles gruntwork and starting friction, the traditional junior developer role changes fundamentally, and "become a lead as quickly as possible" isn't a career path everyone can follow.
The five-cents-per-prompt observation is genuinely useful though. Too many teams treat AI tool budgets like infrastructure costs to minimize rather than productivity multipliers to maximize. If your company has engineers manually writing boilerplate because nobody approved a Claude Code license, you're leaving the most obvious gains on the table.