Token Spend Is the New Cloud Bill

An executive asked us last month: "How do I control AI spend across the organisation without limiting my employees' productivity?" Then, in the same breath: "And should I be setting spending limits?" Two questions, and they pull in opposite directions. That tension is the whole subject of this post.

I have seen this movie before. In the early cloud years, nobody watched the AWS bill because it was small. Engineers spun up instances the way they opened browser tabs. Then one month the bill arrived with an extra digit, finance panicked, and companies swung to the opposite extreme: approval forms for every server, spend frozen, and the engineers quietly went back to doing things the slow way. The overcorrection cost more than the overspend ever did.

Token spend is following the identical arc. For a year it was a rounding error, a few hundred dollars a month across a handful of enthusiasts. Now agents run for hours, chew through millions of tokens per task, and the invoice has become a line item the CFO asks about. The temptation is to ration. The right answer is to manage, and those are very different things.

Should I Be Implementing Spending Limits?

Yes, but at the team level, and as alerts before hard stops. The mistake I keep seeing in mid-market companies is the tight per-seat cap: every employee gets $30 a month of AI usage and the meter cuts them off. It feels prudent. It is quietly disastrous.

Consider what happens when a capable employee hits a per-seat cap mid-task. They do not stop working. They finish the job manually, at their fully loaded hourly rate, which for a $90K salaried analyst is $60 an hour or more once you count overheads. A task the agent would have completed for $3 in tokens now costs $180 in human time. The cap saved you $3 and cost you $177, and it taught the employee that using AI gets them punished with friction. Adoption dies quietly, one capped session at a time. Manual work is the most expensive compute you own.

The structure that works instead has three layers:

• Team-level budgets. Give each department a monthly token budget sized to its workload, owned by the team lead, with genuine headroom. The lead reallocates within it; individuals never see a meter.
• Alerting at thresholds, not ceilings. Fire a notification at 60 and 85 percent of budget. A human looks at the trajectory and decides. Nobody gets cut off mid-task; someone gets a conversation about what changed.
• Hard limits only at the blast-radius level. One organisation-wide circuit breaker, set well above normal usage, that exists purely to catch a leaked API key or a catastrophic loop. It should never fire in a normal month.

The platforms are converging on this pattern too. Anthropic's new Claude for Government Desktop shipped its public beta this month with department-level admin and spend and model limits built in, which tells you where enterprise tooling thinks the control point belongs: the department, not the individual.

How Do We Prevent Runaway Sessions?

Separate the spend problem from the runaway problem, because they need different tools. Budgets manage intent. Runaway guards manage malfunction: the agent stuck in a loop, retrying the same failed step for four hours while the meter spins.

Every agent deployment we set up gets three mechanical guards:

1. 1Turn limits. Cap the number of agent turns or tool calls per session. A task that genuinely needs 200 turns is rare; a session at turn 200 is far more likely to be circling. Set the ceiling at twice your longest legitimate task and treat hitting it as a defect to investigate, not a cost to absorb.
2. 2Wall-clock limits. A hard time budget per unattended session. If a reconciliation job normally takes 20 minutes, a 90-minute session should pause itself and ask for a human.
3. 3Auto-pause on repetition. The cheapest loop detector is a check for the agent repeating the same action with the same result. Three identical failures should suspend the session and flag it, not fund attempt thirty.

None of these limit productivity, because none of them fire on productive work. They fire on malfunction, which is exactly the distinction the blunt per-seat cap fails to make. This is also a design-quality question when you are choosing agent products in the first place; we covered what to demand from vendors in how to evaluate AI agents, and session guards belong on that checklist.

How Do I Start Controlling Token Spend?

Start by measuring the right unit, because the wrong unit will mislead every decision that follows. The unit is not the token. It is the completed task.

Cost per token tells you what a model charges. Cost per completed task tells you what your work actually costs, and the two have diverged sharply. This month's pricing reset made the gap visible: OpenAI's GPT-5.6 launch priced its tiers at $5/$30 per million tokens for Sol, $2.50/$15 for Terra, and $1/$6 for Luna, with Terra matching GPT-5.5 performance at half the price. Anthropic's Claude Fable 5 sits at $10/$50, twice Sol on both sides.

Read those numbers per token and Fable 5 looks expensive. Read them per task and the picture scrambles: on the Artificial Analysis Coding Agent Index, index-reported results have Sol scoring 2.8 points above Fable 5 while using under half the output tokens, landing at a third of the cost per solved task on that index. A model with a higher sticker price can still win on cost per task if it finishes in fewer attempts with less output; a cheap model that needs three retries is not cheap. That spread is your management opportunity, and per-token dashboards cannot see it. It is also why routing different work to different models matters, which we unpack in our multi-model strategy piece.

The practical starting sequence:

1. 1Tag every agent workload with a task type: proposal draft, invoice reconciliation, report build, code change.
2. 2Log tokens and dollars per session, and whether the task completed without human rework.
3. 3Divide monthly spend per task type by completed tasks. That number, cost per completed task, is your unit economics.
4. 4Only then tune: move task types down a model tier where quality holds, and up a tier where retries are eating the savings.

One caching footnote while you are instrumenting: the pricing mechanics moved under everyone's feet this month too. OpenAI's new scheme charges cache writes at 1.25x uncached input with explicit cache breakpoints, so a workload built on heavy cached context needs its assumptions re-run rather than carried forward.

How Do I Attribute Value to a Token?

Price the counterfactual. A token has no intrinsic value; the task it completes does, and the honest yardstick is what that task would have cost in human hours.

The arithmetic is not subtle. Take a monthly board report that consumed six hours of a finance manager's time, $400 or so fully loaded. An agent now assembles it for $5 in tokens and 30 minutes of review, call it $45 all-in. The token spend is not a cost to minimise; it is the cheapest labour you have ever purchased. Whether the run cost $2 or $8 is a rounding error next to the $350 it displaced.

Do this at the task-type level, not for every session. For your ten biggest agent workloads, write down three numbers: what the task costs in tokens, what it used to cost in human hours, and how often it completes without rework. Those thirty numbers will settle every budget argument, because they turn "AI spend is up 40 percent" into "we bought $60,000 of work for $4,000". One reframe helps finance enormously: tokens are cost of goods for knowledge work, not overhead. Nobody celebrates a factory that halved its raw-material purchases by halving production.

How Do I Capture the Full Scope of ROI?

Time saved is the floor, and it is the only number most companies count. The full return has four layers, and the largest one never appears in a time study.

• Time saved. Hours displaced on existing tasks. Easy to measure, so it gets measured. This is the floor.
• Cycle time. The proposal that went out in four hours instead of four days changes win rates as well as labour cost. Speed compounds commercially in ways a timesheet never shows.
• Error rates. Reconciliation that catches what a tired human at 4pm misses. Count reworks avoided and disputes that never happened; they were always a cost, just an unbudgeted one.
• Work that previously did not happen at all. This is the big one. The 40 dormant leads that never got follow-up because nobody had capacity. The competitor analysis that was never commissioned. When the marginal cost of a task drops from $300 of staff time to $4 of tokens, work that was silently rationed out of existence starts happening. There is no baseline to compare against because the old volume was zero, which is precisely why it never makes the ROI slide and precisely why it should.

An honest ROI account also changes what "expensive" means at the top end. For long-horizon, high-value work, a complex migration, a due-diligence pass, a multi-day analysis, the frontier tier at $10/$50 is frequently the correct spend even when a cheaper model could limp through. If the task is worth $20,000 and the premium model finishes it reliably for $60 in tokens where the budget one needs two supervised retries, the extra $40 is the best money in the building. Spending more per task is often the optimisation, not the failure. Cheap models for cheap tasks, frontier models for consequential ones, and cost per completed task as the referee.

Where to Start if the Bill Is Already Climbing

Manage the spend the way you eventually learned to manage cloud: team budgets with alerts, mechanical guards on runaway sessions, and cost per completed task as the number everyone steers by. Rationing feels responsible and quietly reroutes your payroll back onto the most expensive compute you own.

If your AI bill is growing and nobody can say what it is buying, that is a solvable measurement problem, and it usually takes days to fix, not months. Our Deep Dive maps your actual workloads to unit economics: what each task type costs, what it displaces, which model tier it belongs on, and where the guards go. If you want help turning the token bill from a source of anxiety into a managed cost of goods, that is the place to start.