The AI Pilot Worked. That’s When the Bill Starts.

We spend a lot of time on the pilots that fail. In 2026 the more expensive problem is the ones that succeed.

One company spent half a billion dollars on AI in a single month. Not over a year — one month. The cause wasn’t a bad model or a failed rollout. It was the opposite: the tool worked, people used it, and nobody had put a usage limit on the licenses. Axios reported it in late May 2026 — a client that burned through its entire annual AI budget by April after employees leaned hard into AI coding tools.

That’s the story that should be keeping mid-market operators up at night, because it inverts everything we’ve been telling you. We’ve written about why 95% of AI pilots show zero P&L impact, and why scope, baseline, and ownership separate the winners from the rest. All of that still holds. But there’s a second failure mode that doesn’t look like failure at all. It looks like adoption.

Success is the cost event

Here’s the uncomfortable mechanic. A pilot that fails is cheap — it stalls, people stop using it, the meter stops. A pilot that works is the one that runs up the bill, because the thing you were trying to achieve — adoption — is the same thing that drives cost.

Modern AI doesn’t bill like software. It bills like electricity. Token-based pricing, autonomous agents running background workflows around the clock, large context windows re-reading the same documents, employees using a frontier model to check the weather. Every one of those scales with usage. The more successful your rollout, the faster the number climbs — and unlike a SaaS seat, there’s no ceiling unless you build one.

The macro numbers say this isn’t an edge case. Average enterprise AI spend is projected to jump roughly 65% in 2026 — from about $7M to $11.6M — while Deloitte finds only around 10% of organizations using agentic AI can point to significant ROI today. Spend is compounding; returns are arriving for one company in ten. Uber’s COO called AI costs “harder to justify” after the company exhausted its 2026 budget for one coding tool by April. Microsoft — Microsoft — canceled most of its own licenses for a competing AI coding product, partly over cost. If the companies closest to this technology are hitting the wall, the firm running three overlapping subscriptions has no margin for surprise.

Alberta just made the math worse

Two local forces are pushing more operators straight into this trap. First, the Microsoft bill went up on July 1: E3 moved from $36 to $39, E5 from $57 to $60 per user per month. For larger orgs the sting is bigger than the list price — Enterprise Agreement discounts are being stripped out, pushing the effective increase closer to 20% — and Copilot Chat is now bundled into most plans whether your team touches it or not. Your AI line item grew before you added a single tool.

Second, Alberta is booming — a 2.7% GDP forecast, strongest in Canada, on the back of an oil-price surge. Confident, growing companies invest in new tooling, and that’s healthy right up until growth-mode adoption meets uncapped, usage-based billing. It’s exactly the environment where cost discipline gets skipped because everyone’s focused on moving fast.

The fix is a meter, not a budget

The reflex when the bill spikes is to slash access — cancel licenses, lock down agents, tell everyone to stop. That’s the same overcorrection as banning shadow IT: it kills the value you were trying to capture. The companies handling this well aren’t spending less on AI. They know what a unit of it costs. Four moves, none of which require a FinOps platform.

1. Meter the unit, not the total. A total AI bill tells you nothing — of course it’s growing, you’re using more. The number that matters is cost per unit of work: per ticket resolved, per document processed, per report generated. If cost-per-unit falls as volume rises, you’re winning. If it’s flat or climbing, you have a problem no amount of enthusiasm will fix.
2. Put a hard ceiling on every license and every agent. The $500M month happened because a spend limit that takes ten minutes to set was never set. Cap monthly spend by role. Cap what an autonomous agent can consume before it pauses and asks. Treat an uncapped AI license the way you’d treat a company credit card with no limit — because that’s what it is.
3. Attribute cost to the workflow it serves. “AI spend” as a single line item is undebuggable. Tie each dollar to the workflow it runs — intake, triage, drafting, coding — so you can see unit economics per use case. Some workflows are paying for themselves five times over; others are the weather-checking kind. You can’t tell them apart until the cost is attributed, and once it is, the kill-or-scale decision makes itself.
4. Give one person the meter to watch. Every version of this problem traces back to nobody owning the number. Assign one operator to review cost-per-unit and spend-by-workflow on a set cadence — monthly at first, quarterly once it’s stable. Not a committee, not the vendor. The same principle that stops agent sprawl stops cost sprawl: an inventory nobody reviews grows back in ninety days.

The quiet part

A failing AI pilot announces itself — it stalls, the demo energy fades, someone cancels it. A succeeding one makes no noise at all. It just works, quietly, while the meter runs in the background. That’s what makes runaway AI cost the more dangerous of the two: the failure you can see; the one dressed as success, you find on the invoice.

You don’t need to spend less on AI. You need to know what a unit of it costs, cap what it can consume, and put one person in charge of watching the number. Do that before you scale, and the pilot that works stays a win instead of the most expensive thing you shipped this year.