A language model charges for each call it answers, and the calls fall in two places: while a policy is being written, and while an agent runs.
Five steps; three of them run before the model reads the request.
Model cost follows the shape of the product. At design time a person states intent in plain language, and the cloud model of your choice reads that intent and drafts the policy: tokens in, tokens out, a price per call. At run time the Lima Rules Engine executes the published rule.
Where a run needs help with language, it calls the small language model that runs inside your estate. That model serves the published plan, sits on your own hardware and carries zero token cost per call, so the cost of a thousand runs matches the cost of one.
The token manager sits in front of both. Before a call it reads the budget for the agent, the budget for the team and the cap you set for the month. It then picks the route the link allows: the cloud model at design time, the model in your estate at run time.
After the answer it records the reading: tokens in, tokens out, the time taken, the model that served it and the decision it belonged to. That reading joins the decision record, so a cost per decision is a figure your team reads on any agent.
Budgets are read before the call, so the month stays inside the figure your finance lead approved.
Showcase estate, with one built agent. The amber row is above the warning line the team set at 80%. The safety stock and lead time drift rows are designed roles from the agent blueprint, yet to be built, with zero spend; they show what the screen would hold once several roles are built. blueprint
A budget is a figure per month, set by the person who owns the agent or the team. Juno reads it before every call, so a month of work stays inside the number a finance lead approved. The warning line sits where your team puts it; 80% is the usual starting point.
Above the warning line the owner hears the same day, in the manager report and in the alert route the team uses. The figure shown is the spend so far, the budget and the days left in the month.
At the cap, work that carries a person continues and scheduled work waits. A planner writing a policy keeps the session; rule discovery, which runs on its own schedule, waits for the next month or for a raised budget.
Raising a budget is a change like any other: a named owner makes it, and the change joins the version history with the person and the minute on it. KPMG, June 2026, found that 36% of organisations set usage or token budgets at all, so this is still the minority practice.
Design time carries the token cost; run time runs on the model inside your estate.
| Kind of call | When it happens | Which model answers | What it costs |
|---|---|---|---|
| Reading a planner’s intent | Design time, while a policy is written | The cloud model of your choice | Tokens in and out |
| Drafting the planning table | Design time, in the same session | The cloud model of your choice | Tokens in and out |
| Rule discovery | A scheduled task over past decisions | The cloud model of your choice | Tokens in and out |
| Putting a basis into words | Run time, on each recommendation | The model inside your estate | Zero token cost |
| Reading free text in a step | Run time, inside a published plan | The model inside your estate | Zero token cost |
| Explaining a drift finding | Run time, when a reading leaves its range | The model inside your estate | Zero token cost |
The cap holds the scheduled work; the work a person is doing carries on.
Run time stays steady through all of this, because the model inside your estate answers those calls and carries zero token cost. An agent keeps deciding, keeps drafting and keeps writing back while a design-time budget is being discussed.
Gartner, 25 June 2025, expects over 40% of agentic AI projects to be cancelled by the end of 2027, and names cost among the three reasons. A budget per agent, read before every call, is the control that answers it.
Showcase estate.
On the Friday from 09:00 L. Tan stated the PO cancellation policy to the Planner Agent, with D. Okafor named as the approver of every action it drafts. The session made 62 calls to the cloud model, read 41,000 tokens and wrote 9,200, and cost $420 against the purchasing team’s September budget. Simulating and refining it through the middle of that day added $260.
From 13:30 that afternoon the cancellation agent ran on the published rule, once each working morning: the cancellation agent ran 40 times over the eight weeks, each run reading the rule store and calling the model inside the estate for wording, at zero token cost. It is the one agent built in the showcase estate, so those 40 runs are the whole of run time. Rule discovery, which reads six weeks of decisions on a schedule, added $604. J. Harrison read the eight weeks as one figure: $1,284 of model cost, against 40 runs of the cancellation agent and 212 drafted actions.
How a model link is made, what the boundary rule allows across it, and how a live handshake confirms the link before it serves.
The difference is whether the cost is known before the call or after the month.
Consumption pricing is the common pattern for agents, and the unit differs by vendor: a credit, an action, a conversation, a resolution. The spend depends on how an agent was designed, which makes a forecast hard before go-live and a surprise likely after it.
KPMG, June 2026, found that 26% of organisations have full real-time visibility of AI operating cost, and PwC, May 2025, found cost tied with security as the top barrier to agents. Juno answers both with a budget read before the call and a cost figure per agent and per decision.
Run-time calls go to the small language model that runs inside your estate, on your own hardware. The rules engine does the deciding, so the model is asked for wording alone.
Yes. Juno reads the agent budget, the team budget and the monthly cap before each call, and the tightest of the three applies.
The metered cost of the model calls that belonged to that decision. Design-time calls are recorded against the policy and shared across the decisions made under it.
Run time carries on, because those calls use the model inside your estate. A cap holds design-time work such as rule discovery until an owner raises the budget.
Each one builds on the one before it.
Reads your rows in place.
Connect and move ↗One record per real thing.
Master and govern ↗Apps for every decision.
Operate and decide ↗Asks and acts, at your level.
Ask and act ↗Bring two systems and one decision your planners make every week. In the first week we register the data, confirm the mapping with your team, answer your questions in ChatLima and run the first agent at the level you choose.
