Why AI Costs Spiral When Economic Governance Isn't Architected

Value Is the Story — the Bill Is the Consequence

The deciding factor in enterprise AI adoption is no longer what the technology can do. It is whether anyone can prove what the technology returned — and in most organizations, no one can.

The bill is what gets the headlines, and the numbers are genuinely startling. Worldwide AI spending is forecast at roughly $2.59 trillion in 2026, up about 47% year over year. But a large bill is not, by itself, a problem. Enterprises spend enormous sums on things they can justify every day. A bill becomes a crisis only when the other side of the ledger is empty — when the spend is real and the value is a shrug.

That is the actual condition of enterprise AI today. Fewer than one-third of corporate decision-makers in a Gartner survey could identify specific financial outcomes attributable to their AI investments. MIT found that 95% of organizations deploying generative AI saw zero measurable P&L impact. The newest wave repeats the pattern at higher stakes: companies anticipate an average 171% return on agentic AI, yet only 39% can attribute any earnings impact to AI at all — and Gartner projects that more than 40% of agentic AI projects will be canceled by 2027, citing unclear ROI, escalating costs, and inadequate controls.

Read those numbers carefully and the story is not overspending. It is unprovable value. Some of those deployments are almost certainly paying for themselves — but their owners cannot demonstrate it, which in a budget review is the same as not happening. The money is responding accordingly: Forrester found enterprises postponing 25% of planned AI spend to 2027 as financial scrutiny increases. The CFO is now in the room, asking the ordinary questions asked of any line item — what did we get, where did it go, who approved it. What is remarkable is not that the questions are being asked. It is how few organizations can answer them.

None of this means the technology failed. The capability is real, and the organizations that instrumented their deployments are measurably ahead. What failed, in most places, is the economics around the capability: value that was never defined, spend that cannot be traced to outcomes, and budgets discovered on invoices rather than decided in advance. The market is not pulling back from AI. It is pulling back from AI whose value cannot be proven — and closing that gap is the subject of this article.

Two Skipped Steps, One Predictable Outcome

The runaway-bill stories of the past year were not freak accidents. They were the predictable result of two skipped steps — and the second one matters more than the first.

The stories themselves have become industry folklore. An unnamed enterprise reportedly ran up a $500 million bill on a single AI provider in one month after failing to set usage limits for its employees. Uber reportedly burned through its entire $3.4 billion 2026 AI budget in four months before capping employees at $1,500 a month; Microsoft terminated internal AI coding licenses after per-engineer bills reached $500 to $2,000 monthly. One engineering team discovered a multi-agent research tool had slipped into a recursive loop that ran for eleven days — two agents talking to each other continuously — before anyone noticed the $47,000 bill. “We are 3x over our entire 2026 token budget and it's only April,” one industry group's director recalled hearing from company after company this spring.

The first skipped step is the obvious one: nothing was enforcing limits. In nearly every case, controls existed — the $500 million incident happened on a platform that had spending controls available; the company simply had not configured them. Off by default, opt-in by design, fail open in practice. The invoice was the first alarm to fire, and it fired weeks after the money was gone.

But missing limits explain only how the spending ran away — not why no one could say whether any of it was worth it. That is the second skipped step, and the deeper one: no one defined the value before committing the spend. Consider the arithmetic those stories imply. An enterprise paying $1,500 a month per employee for AI tooling is spending roughly $18,000 a year per head — on top of salary, on top of existing software. That is not inherently unreasonable. It is unreasonable unanswered: What was that $18,000 expected to return? In saved hours, shipped features, reduced headcount growth, faster delivery? How would the return be measured, and when would the investment be evaluated against it? In the rush to adopt, those questions were clearly either never asked or answered poorly.

This is the pattern hiding inside the ROI crisis. When enterprise leaders say they cannot calculate the return on their AI investment, they are often describing the symptom of a harder failure: they never defined what the AI was supposed to do. The measurement choices confirm it — half of companies grade their AI on data quality improvements and 48% on employee productivity; far fewer tie AI directly to P&L or margin impact. Those are activity metrics, not value metrics. They measure that the tool was used, not that the business gained.

Basic economic planning — the discipline applied to every factory, every acquisition, every hire — was skipped for AI, because AI was treated as a race rather than an investment. Spend first, define value never. The bill arrived on schedule. The business case never did.

And here is the uncomfortable part: even the organizations that did set targets watched their budgets fail anyway — because the economics underneath AI spending break the assumptions budgeting was built on. That is where we turn next.

The Economics Underneath Break Every Budgeting Assumption

Even the organizations that did set targets watched their budgets fail — because token economics break the assumptions enterprise budgeting was built on.

Enterprise software has been budgetable for decades because its cost model is fundamentally static. You buy seats, you provision servers, you sign a license — and the cost is decided at procurement, in a contract, by a human. Usage might vary, but the spend is anchored to things that change slowly: headcount, infrastructure, terms. A budget is a reasonable prediction because the things that drive cost are things you control at planning time.

AI spend is anchored to none of them. The fundamental unit is the token — not a compute hour, not a seat — and tokens are consumed by activity, not by people. When the activity is generated by autonomous agents, the relationship between headcount and cost dissolves entirely. Per-developer AI consumption rose roughly 18.6x in nine months — the same employees, the same tools, an order of magnitude more spend — because agentic workflows multiplied the activity each person set in motion. Falling per-token prices have not slowed the curve, because volume is growing faster than prices decline. A budget anchored to headcount is measuring the wrong variable.

The mechanics compound quietly. An autonomous agent working through a task re-reads its accumulated context at every step — the same history, paid for again on every call. A task that costs three times more than a single call at five steps exceeds a 30x multiplier at fifty steps, and 100x past two hundred steps — the length of a typical autonomous debugging session. The cost of an agentic workflow is not the sum of its steps; it is the compounding of them. No budgeting model built on linear consumption survives contact with that curve.

And when it goes wrong, nothing announces it. This is the property that separates AI overspend from every operational failure a business is built to catch. The eleven-day recursive loop that produced a $47,000 bill ran on infrastructure where, the whole time, every dashboard was green — because nothing was failing. The agents were healthy. The systems were responsive. The work was even plausible. Traditional monitoring watches for things that break, and runaway AI spend breaks nothing. It just runs. The only instrument that registers the failure is the invoice, and the invoice arrives after the money is gone, aggregated into a number that cannot be traced back to the agent, the task, or the decision that incurred it.

Put the three properties together — cost decoupled from headcount, consumption that compounds rather than accumulates, and failure that looks identical to success — and the conclusion is uncomfortable but unavoidable: the problem is not that budgets were set carelessly. It is that a budget, as enterprises have always practiced it — a number agreed in advance and checked against a monthly statement — is structurally incapable of governing this cost model. The spend moves at machine speed; the oversight moves at billing-cycle speed. Whatever governs AI economics has to operate where the spend actually happens: at runtime, before the tokens are consumed, not on a statement after they are.

That is not what most of the market is building. Most of the market is building better statements.

Most of the Market Starts at the Bottom

Picture the full chain of AI economics as a cascade, flowing from intent to invoice:

Every layer is supposed to be governed by the one above it. Objectives define what AI is for. Governance turns objectives into policy. The budget boundary makes policy a hard line. Runtime decisions happen inside that line. Tokens are consumed as a result. The bill records what happened. Read top-down, the bill is a consequence — the last and least interesting line of a chain of decisions.

Now look at where the market is building. A cost-management industry has formed around AI spend with remarkable speed — two years ago, 31% of FinOps teams were managing AI spend; today it is 98% — and nearly all of it starts at the bottom of the cascade and climbs.

The first camp reads the bill. Cost-observability platforms ingest provider invoices, allocate spend to teams and projects, and flag anomalies — genuinely useful, and structurally after the fact. Traditional FinOps tooling, built to analyze spend in retrospect, is fundamentally mismatched to the real-time economics of AI token consumption — a windshield-versus-rearview-mirror problem, as one industry analysis put it. A dashboard can tell you where the money went. It cannot stop the money from going.

The second camp climbs one layer, to token consumption and runtime calls. Gateways and routing layers cache semantically similar requests and steer routine work to cheaper models — real savings, at the level of the individual call. But a gateway optimizes each request as it passes; it holds no concept of what the spend is for, whether the workflow should exist at all, or what the business expected in return.

The third camp is the providers themselves, offering caps and admin controls — which share one disqualifying property: they are opt-in. Someone must remember to configure them, and the defaults favor consumption. The $500 million incident happened on a platform where the controls were available; the company simply had not configured them. A control that fails open is a suggestion.

To be clear, none of this work is wasted — visibility, attribution, and per-call efficiency are necessary layers, and the discipline forming around them is real. The Linux Foundation launched the Tokenomics Foundation this year to develop best practices for managing enterprise AI use at scale — the clearest possible signal that the industry knows the problem is unsolved. But climb the cascade from the bottom and you can ascend forever without reaching the top two layers, because they cannot be built from below. No amount of billing data produces a business objective. No gateway derives your economic policy. The layers that govern everything beneath them — objectives and the governance that encodes them — have to be architected first, from the top, or they do not exist at all.

Most organizations are discovering exactly that gap the hard way. And for many of them, the timing is about to get worse — because a second force is arriving that assumes those top layers exist, and penalizes their absence.

Regulation Is About to Compound the Problem

For organizations already unable to control or attribute their AI spend, the regulatory environment is arriving as a second bill — on a schedule, with penalties attached.

The direction is unmistakable even where the timelines shift. In Europe, the majority of the EU AI Act's rules come into force on August 2, 2026, when enforcement begins at the national and EU level. The recent Digital Omnibus agreement deferred obligations for high-risk systems to December 2027 and product-embedded systems to August 2028 — relief on timing, not on substance: the requirements themselves — risk classification, documentation, human oversight, recordkeeping — remain unchanged, with the delay explicitly premised on the expectation that implementation efforts should already be underway. The penalties were never softened: up to €35 million or 7% of worldwide turnover for prohibited practices, and up to €15 million or 3% for other infringements — applying to EU and non-EU companies alike.

In the United States, the problem compounds differently: not one rulebook but dozens. State legislators introduced over 1,100 AI-related bills in 2025 alone, producing roughly 100 laws and proposed rules — California's transparency and frontier-model laws effective this January, Texas's governance act the same day, Colorado's replacement statute arriving in 2027, with states defining “AI,” “high-risk,” and “consequential decisions” differently, forcing businesses to analyze the same system under multiple frameworks. A federal executive order is now pushing to preempt the patchwork — which may eventually simplify it, but today adds a layer of uncertainty on top: companies are advised to continue preparing for compliance under current frameworks while watching legislative developments that could reshape obligations in the near term. Preparing for rules that may change is itself a cost. Industry estimates suggest compliance adds approximately 17% overhead to AI system expenses.

Step back from the statutes and notice what they all ask for. Disclosure of where AI is used. Risk assessment of what it is used for. Documentation of how it is overseen. Records of what it did, retrievable when a regulator or an affected consumer asks. Jurisdictions differ on thresholds and timing, but the demand converges on one artifact: a governed account of your AI — its purpose, its boundaries, its conduct, its record.

Now recall the cascade. Every one of those demands lives in the top layers — defined objectives, encoded governance, enforced boundaries, attributable records. An organization that entered AI from the bottom, discovering its deployment through invoices, has none of them. It cannot say with confidence where AI runs, what it does, what data it touches, or what any of it was for — which means compliance is not a reporting exercise but a reconstruction project, performed under deadline, at consulting rates, for every jurisdiction that asks.

This is how regulation compounds the problem rather than merely adding to it. The unprovable-value crisis and the compliance crisis are the same deficiency wearing two costumes: no architected account of what your AI is doing and why. The CFO asked first. The regulators are asking next, and unlike the CFO, they schedule the audit, define the format, and fine the gaps. For organizations that skipped the top of the cascade, the cost of not knowing is about to be priced by someone else.

The Budget Boundary: Control Flows Down, Not Up

Control has to flow the other direction — and this is where the platform argument begins, because flowing the other direction is not a practice you adopt. It is an architecture you build on.

Loriqa starts at the top of the cascade because it was built there. Business objectives define what your agents are for. Economic governance translates those objectives into policy: what each agent, task, and workflow is permitted to spend, and what happens when a limit is approached. And between policy and execution sits the layer that gives this section its name — the Budget Boundary: the line where economic policy becomes physically enforced, in the execution path, on every action, before the money moves.

Enforcement at the boundary means the order of events is reversed. In bottom-up approaches, spend happens and control reacts — the alert fires after the threshold, the dashboard reports after the month, the cap halts a process that already burned its budget getting to the wall. At the Budget Boundary, the platform evaluates before execution: a process that is going to exceed its token budget is stopped ahead of the spend, not discovered behind it. What happens next is a configuration decision, not an improvisation — the request is elevated to a human for exception approval, or it is denied. Either way, the outcome is decided by policy someone set deliberately, in advance, at the top of the cascade.

And the boundary does not fail open. This is a design principle, not a feature: nothing in Loriqa fails open, and the budget is no exception. There is no unconfigured default that permits unlimited consumption, no gap where a forgotten setting quietly authorizes a $500 million month. The posture inverts the market's: spending beyond policy is not what happens when no one intervenes — it is what cannot happen unless someone approves it.

Enforcement is half the boundary. Attribution is the other half. Because every agent action passes through the governed execution path, every unit of spend is captured where it happens — attributed to the agent that incurred it, the task it served, the model it ran on, and the authorization behind it. Not reconstructed from an invoice; recorded at the moment of execution, in the same tamper-evident chain as everything else the platform proves. This is the artifact the previous two sections established as missing: the CFO's unanswerable question — what did we get, where did it go, who approved it — becomes a query. The regulator's demand for a governed account of your AI becomes a report the record already contains.

Attribution at that grain also surfaces what aggregate invoices are built to hide. Track consumption per agent, per task, and per model over time, and provider-side drift becomes visible: token burn rates quietly rising though the work hasn't changed, or service levels degrading while the price stays the same. Without independent measurement, a customer cannot distinguish their own workload growing from their provider silently increasing the burn rate and spend. With it, the relationship changes character — AI spend stops being a bill you receive and becomes a contract you can verify.

That is the Budget Boundary: policy enforced before the spend, nothing failing open, and every dollar carrying its own explanation. What it is not — and this matters — is a wall. A boundary that only says no would make AI safer and smaller. The next section is about the opposite motion: how the same architecture that controls the spend continuously finds ways to reduce it.

Spend Control Is Not a Rule — It Is an Ongoing Process

A budget you set once is a snapshot of what you knew on the day you set it. Models change monthly. Prices move. Workloads evolve. Agents that were the right choice in January are the expensive choice by June — and no static rule notices. Spend control that stops at enforcement leaves money on the table every day after the policy was written. The Budget Boundary keeps spending inside the lines; this section is about the platform continuously redrawing the lines in your favor.

Loriqa does this through two capabilities built on the same foundation as everything else: the record.

The Continuous Improvement Program (CIP) evaluates your deployed agents against the metrics that define value — cost, speed, accuracy — and recommends changes. Not generic advice; specific, evidence-backed proposals drawn from how your agents actually run: where spend concentrates, where runs take longer than the work requires, where budgets are over-provisioned against observed need. The evaluation never stops, because the conditions never stop changing. What arrives is a stream of governed recommendations — each one reviewed and approved by you before anything changes, because an optimization pipeline that modifies production on its own initiative would be a new risk wearing a savings costume.

QIA answers a question most organizations have never had the instrumentation to ask: is each task running on the right model? The frontier model that demos best is rarely the right tool for every job — routine classification does not need the reasoning engine that architectural analysis needs, and paying frontier prices for commodity work is one of the quietest, largest leaks in an AI budget. QIA evaluates your agents and tasks against multiple models and scores the results against your priorities — lowest cost, highest accuracy, fastest response, or the weighting that fits the work. Not the provider's recommendation. Not the default nobody revisited. A measured answer, per task, to a question whose answer changes as the market does.

And the highest-value recommendation of all: stop paying for intelligence where none is needed. When the record shows an agent performing the same pattern over and over — same steps, same tools, same shape of work — that pattern is a candidate to move off LLM execution entirely, into a deterministic workflow: same output, near-zero marginal cost, and fully repeatable. This is the optimization no bottom-up tool can make, because it requires seeing the work, not the bill. A dashboard sees that an agent cost $40,000 last quarter. The record sees that $28,000 of it was the same routine performed four thousand times — work that never needed a model at all.

Notice what the three have in common: none of them is a cap. Caps limit damage; these reduce cost — structurally, permanently, and compounding, because every retired pattern and every right-sized model lowers the run-rate for good. The market's tools tell you what AI spent. An architected platform keeps finding what AI no longer needs to spend. That difference compounds every month it operates — and it is the difference between managing a cost and optimizing an investment.

Governance Is the Enabler

There is a belief, quietly held in most organizations, that governance and capability trade against each other — that every control subtracts some power, and a fully governed AI deployment is a cautious, diminished one. The past seven sections argue the opposite, and it is worth saying plainly: the organizations getting the least from AI right now are the ungoverned ones.

Look at what ungoverned actually bought them. Budgets exhausted in a quarter. Spending frozen mid-year. Projects canceled not because the technology failed but because nobody could prove it worked. Access rationed, caps imposed in panic, experimentation shut down by the CFO because the last experiment cost half a billion dollars. Ungoverned AI does not run free — it runs briefly, then gets restrained by the crudest instrument available: less AI. The spiral in this article's title ends the same way every time, and the ending is retreat.

Now run the same story on an architected foundation. Every agent operates inside a budget that cannot fail open — so leadership never faces the invoice that triggers the freeze. Every dollar carries its attribution — so value is provable, and what is provable can be defended, renewed, and scaled. Targets are set at the top of the cascade before spend begins — so the question “was it worth it?” has an answer scheduled from day one. The platform itself hunts for savings continuously — so the run-rate falls as the deployment matures instead of bloating with it. And when regulators arrive with their convergent demand — show us the governed account of your AI — the account already exists, because it was never a report to assemble. It was the architecture all along.

That is what economic governance actually purchases: not restraint, but confidence — and confidence is the scarcest resource in enterprise AI today. The confidence to give agents real autonomy, because the boundary holds regardless of what they attempt. The confidence to expand the fleet, because the next hundred agents are as governed as the first. The confidence to say yes — to the new workflow, the deeper integration, the more consequential task — because yes no longer means hoping the bill and the behavior stay reasonable. It means knowing the boundary is enforced, the record is complete, and the value will be measurable when the board asks.

The companies pulling back from AI this year are not pulling back from the technology. They are pulling back from what they cannot see, cannot prove, and cannot stop. Solve those three — see everything, prove everything, stop anything — and the retreat inverts. Spend becomes investment. Caution becomes velocity. The budget boundary, it turns out, was never a wall around your AI. It is the foundation under it — and it is the reason you can finally build as high as the technology allows.

This article was researched and drafted in collaboration with Claude, Anthropic's AI assistant — because the best thinking happens in good company.