Perspective · July 26, 2026

Are Jira Tickets About to Have Dollar Signs Instead of T-Shirt Sizes?

The cost of a feature used to be a guess wrapped in a t-shirt size. Now there’s a receipt, and somebody has to read it.

By the superpenguin team·6 minute read

What happens when the subsidized Claude Code and Codex subscriptions go away and we are all left to pay per token? Will companies still let their engineers reach for the most expensive frontier models whenever they feel like it? What happens when an engineer accidentally (or not) spends $10K in a day on a research side quest?

Software development effort used to be measured in engineering-hours. We are all hurtling toward a future where it is measured in dollars. In weekly syncs across the world, PMs ask their engineering teams how many tickets they can pull into the next sprint. For a long time, the answer was a function of bandwidth (“how much can the team take on?”). It is quickly becoming a function of budget (“how much do we want to spend?”).

01The engines run on cash, but the dollars aren’t always donuts

The engines burning those dollars, the agents, are finicky. Model choice, prompt, provided context, harness, and a dozen other factors can swing the cost of task completion by orders of magnitude. It is often unclear whether a task designated to an agent will cost cents or hundreds of dollars, which goes a long way toward explaining why so many finance teams have been left staring at their monthly reports in horror.

They are in good company. Uber rolled Claude Code out to roughly 5,000 engineers in December 2025, encouraged staff to use AI “as much as possible” (there were internal usage leaderboards), and had exhausted its entire annual AI budget by April (TechCrunch). Uber’s CTO burned $1,200 himself in a single two-hour demo session (The Information). The fix was a $1,500 monthly cap per employee per tool, which stops the bleeding, but doesn’t go far towards creating a more sophisticated system for balancing productivity and cost. Microsoft went further and pulled Claude Code from the division that builds Windows and M365, effective June 30, conveniently the last day of its fiscal year, after power users hit $500 to $2,000 per month (Forbes). Thousands of engineers were marched over to Copilot CLI, reportedly away from the tool they preferred.

And this is the early, subsidized innings. Ramp, which now watches AI spend across 70,000+ businesses, reports token spend among its customers grew 21 timesin a year. Its CEO put it plainly on CNBC: the AI companies “have functionally set up a tab. You can spend as much as you want” (Ramp). The median firm on Ramp’s AI Index spends a modest $11.38 per employee per month on tokens; the top 1% spends $7,450 (Ramp AI Index). That gap between the median and the top arguably represents the adoption curve everyone below is about to climb.

These companies had no shortage of smart people or money. What they were short on was a polished feedback loop. The bill arrived after the decisions were already made.

02Why your finance systems aren’t prepared for this

The reason existing finance systems fail in this new paradigm is that the place where the dollar gets spent has absconded from its more centralized origins. Capital used to be allocated in advance, by hiring engineers. Think of headcount as a yearly subscription to software development: it grants the company a finite pool of engineering resources; PMs contend for allocation based on priority and scope, and the hours get divided among them. Finance’s entire toolkit (annual planning, quarterly reforecasts, monthly closes) is built around that cadence. Sure, sometimes headcount gets reforecast, reqs get frozen, budgets get trimmed mid-year. But each of those adjustments is a large, deliberate decision made by a handful of leaders, with an approval chain and a paper trail.

Agents have moved the spending decision to the vanguard. Every engineer becomes a budgeter, an allocator, a spender, one prompt at a time. Finance sees only the aftermath, weeks later, on an invoice, often devoid of the initiatives, projects, features, or customers the spend was in service of.

The first generation of fixes proves how wide the gap is. Ramp’s token spend product recently surfaced the term “prompt caching” to a Controller at AngelList who had never heard of it; he routed it to engineering, and the fix saved $10,000 per month the same day (Ramp). The person accountable for the money and the people spending it sit far enough apart that basic vocabulary doesn’t survive the journey across the office space between.

03 The AI buck stops with whom?

So who should be responsible for this new agentic spend? How does it get budgeted? Does each agile team appoint a tiny finance manager who owns the squad’s P&L and financially optimizes each sprint? It’s an amusing picture, but probably not the solution anyone would prescribe.

We don’t claim to know the org chart of 2027, but a few patterns seem durable enough to bet on.

Firstly, spend needs policy, because per-case judgment doesn’t scale. When you book a flight, you don’t ask the CFO. Economy is the default and business class needs a justification (if you’re at Amazon, a really good one), and the booking tool acts as the automatic enforcer. AI spend needs the same treatment: every firm deciding for itself what counts as economy, what counts as business class, and what machinery guides the needle threading them.

Secondly, accountability has to sit alongside the people making the decisions, at the team level. In the cloud era the answer was FinOps software and a central platform team, and it worked because ten infra engineers controlled the org’s resources, and therefore its costs. It probably won’t work in this era where agentic cost generators toil away discreetly at the periphery. Even the FinOps establishment senses this: 98% of FinOps teams now manage AI spend, up from 31% two years ago, and the Linux Foundation just spun up an entire Tokenomics Foundation because no standards exist for any of it (State of FinOps 2026, Linux Foundation).

Thirdly, the feedback loop has to run at the speed of development, because firms running at the speed of accounting will be out-competed in an AI-driven market.

And finally, nobody is hiring a finance manager per squad.

If policy has to live where decisions happen, it’s worth asking where the decision is actually made. The prompt is the point of sale, but the purchase was approved earlier, when the ticket was written and prioritized. The natural home for the dollar is the ticket. Imagine that tickets carry dollar estimates the way they carry points today, except the estimate says $180 and the retro asks why the actual came in at $212. Sprint planning starts to look like allocation: the PM holds $8K for the sprint and a ranked backlog with price tags, and picks tickets the way a fund manager picks positions. The EM owns the budget the way they own cloud spend today. Engineers keep the judgment calls, deciding when the expensive model earns its keep and when few token do trick (we have opinions on that already), except now the price is visible at the moment of decision. Finance sets the guardrails, gets the telemetry, and stops being surprised.

The prompt is the point of sale. The purchase was approved when the ticket was prioritized.

04 Our bet, loosely held

That’s our fun hypothesis, and we hold it loosely. Maybe intelligence gets so cheap that the bottleneck moves again, to human decision-making and the meetings where we argue about what to build. Maybe companies buy subscriptions generous enough to feel unlimited, and the optimization problem centralizes into a handful of negotiations between the CFO and the frontier labs. DM us and we’ll gladly hear your best argument! What is definitely true is this: AI spend compounds faster than any budget process built on quarters, and the loop between spending money and deciding to spend it has to become shorter than the billing cycle. The companies that close that loop will out-build everyone else, dollar for dollar. The rest will find out what they spent the way Uber did: afterward.

Figures as of mid-2026. Sources linked inline.