Shelf Life | Vol. 54 — The Recoupling: Your Operating Model Coupled Up When AI Was Free.

You built your org chart, your talent strategy, and your budget around AI that cost almost nothing. The subsidy era is over, the meter is running, and it might be time to stop being closed off.

Minimize image

Edit image

Delete image

Top Shelf Insights

💸 The AI subsidy era is ending. OpenAI, Anthropic, Google, and Perplexity are raising prices, restricting usage tiers, and moving high-value features into higher-cost SKUs. The unit economics that funded 2024's AI rollout no longer apply.

💸 AI was introduced to many organizations with promotional economics. The real enterprise cost is now showing up through tokens, usage credits, subscriptions, rate limits, tool calls, agents, storage, context, and compute.

🎯 Token cost is strategy now. Cost per token decides which use cases work at scale. The AI copilot that made sense at $0.01 per query does not make sense at $0.10.

🔀 Staying “closed off” to one AI platform may simplify security, adoption, and governance. It may also create vendor dependence, price exposure, and capability blind spots. Multi-vendor is the new norm. Enterprises are moving to multi-model architectures with cheap models for simple work and premium models for the hard stuff.

⚡ Token optimization is becoming an enterprise capability. It belongs in the operating model, talent model, procurement strategy, governance framework, and workflow design. Token maxing is a new operational discipline. Individually, employees are learning to prompt for efficiency. Enterprise-wide, teams are optimizing token usage the way marketing teams optimized ad spend a decade ago.

🏗️ Operating models built in 2024 assumed unit economics that no longer hold. Talent assessments, org charts, budget lines, and success metrics all get rebuilt.

🏝️ Welcome to the Villa: How We All Coupled Up

For two years, enterprise AI worked like the first week of Love Island. Everything was paid for. The drinks flowed, the pilots were subsidized, and the only job was to couple up with a platform and declare it your type on paper.

And couple up we did. Companies built entire operating models on the assumption that the marginal cost of intelligence was effectively zero. Talent assessments were redesigned around "AI fluency." Headcount plans assumed every knowledge worker came with an unlimited copilot. Workflows were re-architected so that agents would handle the volume humans used to. Some organizations went further and made AI usage itself a performance metric, which is how we ended up with token maxing as a corporate philosophy. Jensen Huang told the world he would be alarmed if a $500,000 engineer consumed less than $250,000 in tokens. Consumption became the KPI. Activity became the proxy for productivity.

Here is the problem with building your operating model in the villa: someone else was paying for the villa. Vendors subsidized pricing to win land grabs. Flat subscriptions absorbed usage that cost far more to serve than it earned. That was a customer acquisition strategy so successful that drug dealers have been using it for years.

📱 I Got a Text: The Invoice Arrives

The text has landed, and it is not a date card.

GitHub Copilot's shift to usage-based billing this June saw individual developers report monthly costs jumping from roughly €67 to over €900 for the same workload. Anthropic moved heavy enterprise users to consumption pricing. Sam Altman said the complaint "my company spent its entire 2026 budget in Q1" went from unheard of to constant. Uber, where 84% of developers had become agentic AI users by March, famously torched its annual AI budget in weeks.

The cruelest part is that per-token prices are actually falling, fast. But total spend equals unit price times volume, and agentic workflows detonated the volume side of that equation. A single prompt used to be a single completion. Now it is a planning step, five tool calls, a retry loop, a verification pass, and a subagent (bombshell) or two.

CIOs now need to look into managing tokens with the rigor of energy or capital allocation. When consultants start publishing "tokenomics" frameworks, the free era is officially over. Someone always pays for the compute.

💔 Closed Off vs. Playing the Field: The Token Optimization Recoupling

In the villa, being "closed off" means staying loyal to one partner no matter who walks in. In enterprise AI, being closed off means routing every workload to one frontier model because it topped the leaderboard the week you signed the contract.

For eighteen months, closed off was the default. It was also mugging you off. Classification, extraction, summarization, and routine logic make up the majority of enterprise AI volume, and none of it needs frontier reasoning. Paying frontier prices for those tasks is the difference between $2.31 and $18.40 per million tokens.

So the recoupling ceremony has begun, at two levels:

✅ Individually, the savviest operators are learning token hygiene the way they once learned search skills. Tighter prompts, curated context instead of dumping entire document libraries into every request, choosing reasoning effort deliberately, and knowing which model fits which task. Model choice plus effort setting is the new personal productivity lever.

🛒 Organizationally, leading enterprises are standing up AI FinOps as a real function: intelligent model routing that sends the easy 80% of requests to cheap or open-weight models, prompt caching, context compression, and, most importantly, a shift from measuring tokens consumed to measuring cost per resolved outcome. Processing a billion tokens is an activity. Resolving 40% of Tier 1 support tickets is an outcome. Reward the second one.

The bombshells walking into Casa Amor are the cheap, capable mid-tier models. Staying closed off to your frontier couple for every workload is no longer loyalty. It is negligence doused with a minor case of Stockholm syndrome.

Minimize image

Edit image

Delete image

Made to Measure (aka Let Me Pull You for a Chat)

This is the uncomfortable fire pit conversation. If your org design, talent assessment, and workflow architecture were built on free AI, they were built on a subsidy, and subsidies are not strategies.

Three moves for leadership teams right now:

🔧 Re-underwrite every AI-dependent workflow at real prices. Take your top ten agentic workflows and model them at current consumption rates, then at rates 30% higher, because usage caps, rate limits, and repricing are all live risks. If a workflow only made sense at subsidized prices, it needs a redesign or a retirement. Gartner projects more than 40% of agentic AI projects will be canceled by end of 2027 over cost and unclear value. Decide which side of that line your portfolio sits on before your CFO decides for you.

🧑‍💼 Rewrite the talent thesis from "uses AI a lot" to "uses AI well." If your performance framework rewarded token consumption, you incentivized exactly the behavior now wrecking your budget. The scarce skill in 2026 is judgment: knowing when a task needs frontier reasoning, when a cheap model clears the bar, and when a human is still the most economical option. That skill belongs in your competency models, your hiring rubrics, and your promotion criteria.

🏛️ Give tokens an owner. Cloud spend got FinOps a decade ago. AI spend needs the same: consumption visibility by use case, model mix governance, unit economics per outcome, and a leadership triad (CFO, CTO, business owner) that reviews it monthly. Ungoverned AI spend is the new shadow IT, except it compounds daily.

Minimize image

Edit image

Delete image

On the House

The phrase "token maxxing" now has its own Wikipedia page, complete with a criticism section noting that workers will maximize whatever metric management rewards. Goodhart's Law got a glow-up and a spray tan. If your dashboard celebrates token consumption, congratulations, you have built the villa's most expensive popularity contest.

Here is my take. The AI honeymoon phase made everyone look more mature than they were. When access felt cheap, the operating model conversation was easier. Give people tools. Create policies. Train teams. Count use cases. Call it momentum. Now comes the real test.

Love Island contestants know this well. It is easy to declare loyalty when the cocktails are free, the lighting is flattering, and no one has asked what happens after the villa. AI operating models are entering that same moment. The question is no longer whether people are using AI. The question is whether the organization has designed a model that can handle what AI actually costs, how fast that cost can scale, and where AI consumption creates measurable business value.

That means matching the model to the task. Pricing workflows honestly. Teaching employees that more prompts do not automatically mean better work. Giving leaders visibility into token consumption before the budget starts looking like a Casa Amor recoupling gone wrong.

The AI operating model cannot be built only for how people use AI today. It has to be built for how the business will use, scale, govern, fund, and optimize AI tomorrow.

The subsidy era gave everyone a free summer in the villa, but have I got a challenge for you...

Minimize image

Edit image

Delete image

The Last Look

If AI is no longer being subsidized by the market, should companies stay closed off to one strategic platform, or is token maxing across multiple models the smarter way to play the long game?

Tomorrow night, in the AI villa: the token bills arrive, the public has been voting, and any operating model built on “free forever” is officially vulnerable.

Jackie Swanson advises retailers and brands on AI readiness, agentic commerce, and transformation at Gartner. She lives in NY with her partner and three children, which is great prep for complex client engagements, or the other way around.

📅 Book a 1:1 | 📬 Follow Shelf Life on LinkedIn | 📸 Instagram: @ShelfLifebyJKS