An engineer burned $1,000 in a single day using Anthropic’s API. As Gergely Orosz relayed in The Pragmatic Engineer, the developer’s reaction when asked how they afforded it: “My company pays for it, so it’s practically free for me.” 2 Source 2 Gergely Orosz, The Pragmatic Engineer. Developer burns $1,000 via Fable API in a single day. X/Twitter, June 12, 2026.
Uncapped, unaccountable per-developer AI spend is why finance now owns the next conversation. Not because the tools don’t work, but because nobody budgeted for them.
Engineering Teams Are Running Audits. Most Don’t Call Them That.
Yes, enterprise engineering teams are auditing their AI tool subscriptions in 2026, and not because AI failed to deliver value. The spend accumulated outside any budget approval process: individual developers expensed tools, teams bought seat licenses, and the aggregate only became visible when pricing shocks forced finance to look. The informal audit is already underway.
Gergely Orosz, whose Pragmatic Engineer newsletter reaches 350,000 subscribers tracking engineering practices at hundreds of companies, put it plainly in June 2026: “Breaks are in. No company will allow unlimited budgets.” 1 Source 1 Gergely Orosz, The Pragmatic Engineer. X/Twitter, June 12, 2026. https://twitter.com/GergelyOrosz/status/2065535488425947349 — 109K impressions. That’s not a prediction. It’s a field report from inside the orgs running the audits right now.
The trigger: Fable pricing made invisible spend visible
Anthropic’s Fable launch forced engineers to look at what they’d been accumulating across every AI subscription, not just this one. The Public AI Foundation summarized the shift on June 10, 2026: “The Free Ride Is Over: Why AI Coding Tools Are Getting More Expensive to Use.” That framing landed because it was accurate.
The benchmark Uber accidentally set
Uber put a $1,500/month cap on per-developer AI spend in June 2026. 3 Source 3 First reported by @DgtlDreamscapes, X/Twitter, June 9, 2026: “Uber Put a $1,500 Wall Around AI — And Accidentally Priced the Whole Boom.” The headline that followed caught on because it named something engineering leaders were already sensing. If your per-developer AI spend exceeds $1,500/month without a utilization report behind it, you’re at the threshold where finance steps in.
The Utilization Problem Is Worse Than You Think
Those aggregate numbers on the finance dashboard don’t show who’s actually using what. That’s what makes this expensive.
AI tools sit at the bottom of the enterprise stack on actual usage, and developers routinely run two of them at once. Four data points make the case:
- AI tools carry the lowest monthly active seat utilization of any category — 34% on average. 4 Source 4 Zylo 2025 SaaS Management Benchmark. zylo.com/resources
- 44% of AI software licenses go unused within 90 days — double the abandonment rate of traditional software. 5 Source 5 Productiv 2025. productiv.com/research
- The combined per-worker AI bill runs $85 to $140 a month across the whole subscription stack, not a single tool. 6 Source 6 BetterCloud 2025. bettercloud.com/research
- 41% of developers already run more than one AI coding assistant. 10 Source 10 JetBrains Developer Ecosystem Survey 2025. 41% of developers report using more than one AI coding assistant. https://www.jetbrains.com/lp/devecosystem-2025/
Redundancy isn’t the edge case. It’s the default.
That 34% average means most engineering orgs are paying full price for a tool that’s idle six days out of seven. The CFO’s question isn’t whether AI is useful. It’s how many tools any single developer actually uses in a given week, and what the overlap costs.
44% of AI SaaS licenses go unused within 90 days — double the abandonment rate of traditional software.
How Do You Know Which AI Tools to Cut?
The threshold most engineering managers are using is monthly active utilization: tools used by 60% or more of the team at least weekly stay; tools used by fewer than 30% get cut. Everything in between gets a 90-day window to prove measurable output or exit.
Three tiers. The data makes the framework simple.
Tier 1: Keep (above 60% monthly active utilization)
A tool earns this tier by meeting all of these:
- Real adoption. 60% or more of the licensed team uses it at least weekly.
- Workflow integration. It’s wired into a primary workflow — the code editor, the build pipeline, or the code-review process.
- Habit, not just access. The team has built working habits around it, not merely turned it on.
- A measurable output. Some metric moved: test coverage, review time, deployment frequency.
- Switching cost. Removing it would mean rebuilding workflows, not just cancelling a subscription.
Tools that clear these bars are load-bearing. They stay.
Tier 2: Evaluate (30–60% utilization, 90-day prove-or-cut)
Most tools end up here. The team knows the tool exists but uses it inconsistently. ROI is theoretical rather than evidenced. The action: assign an owner, define what “earning its keep” looks like as a specific metric, set a 90-day review date. The key question for any Tier 2 tool: has anyone on this team changed how they work because of it?
If no one can answer yes, that’s the answer.
Tier 3: Cut (below 30% utilization)
Downloaded, trialed, largely ignored — still billing. A tool lands here if it meets any one of these:
- Below 30% monthly active seat utilization.
- Functional overlap with a Tier 1 tool you already own.
- No owner who can articulate what workflow it improves.
- Vendor data practices that create unacceptable risk.
That last point matters. Orosz flagged it directly in June 2026: “If you use Claude as a business you should want to have an off-ramp.” 7 Source 7 Gergely Orosz, The Pragmatic Engineer. X/Twitter, June 11, 2026. https://twitter.com/GergelyOrosz/status/2064940216868061320 — 71K impressions. Some tools belong in Tier 3 for trust reasons, not utilization.
Which Categories Are Getting Cut First
Pattern recognition from engineering managers running actual audits in mid-2026.
Duplicate coding assistants. Teams that accumulated both GitHub Copilot and Cursor are consolidating. Not because one is objectively better, but because the monthly active seat report makes one choice obvious. Most teams landing here are keeping Cursor (the tool they built IDE workflow habits around) and pausing Copilot on inactive seats.
Generic AI writing assistants. When the same output is available through five tools the team already pays for (LLM API access, email client AI, document editor AI), the standalone writing subscription is the easiest cut to defend.
Standalone AI testing tools. Unless they hook into CI/CD and produce measurable coverage improvements, they sit at Tier 3. The ones staying are the ones where developers immediately notice their absence, because coverage drops.
Bundled AI features on existing SaaS. Premium tiers from existing vendors with AI add-ons and low adoption rates relative to cost are getting scrutinized. If the AI feature in the upgrade tier isn’t used, the upgrade doesn’t justify its price. This is the same extraction pattern the SaaS pricing squeeze documents: vendors bundling AI at renewal to justify price increases on locked-in customers.
What’s staying: tools with workflow integration so deep that removing them creates immediate friction; tools where the team’s output is measurably different because of them; and increasingly, self-hosted or open-source alternatives, where the cost model is predictable and the data stays in-house.
The Build-and-Route Alternative
The audit conversation defaults to keep or cut. There’s a third path engineering orgs are actively pursuing.
Model routing is emerging as a distinct cost-control category: software that routes each query to the most cost-effective model capable of handling it, rather than defaulting every call to the most expensive option. Orosz identified this as a significant demand opportunity in June 2026 (The Pragmatic Engineer, June 11, 2026). The math is direct: not every AI task requires frontier model capability, but flat-rate subscriptions charge frontier model prices for all of them.
For non-US organizations, there’s an additional dimension. Orosz noted: “every non-US company sees how they can be cut off from US vendors with a snap of a finger” (The Pragmatic Engineer, June 13, 2026). 8 Source 8 Gergely Orosz, The Pragmatic Engineer. X/Twitter, June 13, 2026 — 166K impressions. Open-source models with local inference aren’t just cheaper. For teams with data sovereignty requirements, they’re becoming the only architecturally defensible option.
56% of companies now prefer building internal AI tools over buying external ones (Retool, 2025 State of AI). 9 Source 9 Retool 2025 State of AI. https://retool.com/reports/state-of-ai-2025 That’s a mature capital allocation shift: tools that proved their workflow value are worth keeping or building deeper; tools that didn’t are worth replacing with infrastructure the company actually controls.
This is where the GTM stack audit framework applies directly. The same discipline that surfaces seven-figure hidden costs in the go-to-market stack is running on engineering AI spend now.
What to Ask Your Engineering Manager
The CFO or CTO who delegates this gets results faster than the one who runs it personally. Four deliverables to request.
A complete tool inventory. Every AI subscription: who approved it, monthly cost, seat count. Not a spot-check. A complete list.
Three months of usage data. Most enterprise AI tools expose admin dashboards with monthly active user counts. For tools that don’t, login frequency or API call volume works as a proxy. Three months classifies each tool into Tier 1, 2, or 3 without debate.
A 90-day Tier 2 commitment for each tool in that tier. Each needs an owner and a measurable usage target. Without both, the 90-day window becomes a delay, not a decision.
Renewal criteria documented before vendor conversations start. Vendors see cancellation signals in their own dashboards and will proactively offer discounts and extended trials. If the internal decision criteria aren’t written down before the vendor calls, the vendor’s sales team fills in the definition.
The audit isn’t a sign of AI fatigue. It’s a sign of AI maturity. Companies running the most effective audits aren’t asking whether to use AI. They’re asking which tools have earned their place in the workflow, and which ones were just easy to expense.