
ARGENT FORBUDGET AND TOKEN SPEND
Master Control System · by Nyedis
Your goal: control what AI costs
If your goal is to control what AI costs, start from how it is billed: every request an agent sends to a model such as OpenAI's GPT, Anthropic's Claude or Google's Gemini is metered and billed, by the token. An agent can send call after call, all day, and the total shows up on the invoice, after the money is spent.
- Gartner predicts AI inference costs per agentic workflow will rise more than five times through 2028, as advanced reasoning models typically use five to 30 times more tokens than basic chatbots.¹
- 98% of State of FinOps 2026 respondents now manage AI spend, up from 31% two years earlier, and the report names allocating AI costs to business units as harder than traditional infrastructure.²
- Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls.³
A dashboard tells you what was spent. A budget has to act before the money is gone. Argent gets you there in order: see what your agents spend, give it an owner, then cap it before it is spent.
Six questions a budget owner asks
Controlling spend starts with straight answers, in one console, for the AI calls your agents make through Argent.
| Question | Where Argent answers it |
|---|---|
| What are we spending, right now? | Live traffic streams each request as it happens, with a cost view beside it. Operations shows spend for the last three hours, top spend and budgets about to run out. |
| Who spent it? | The Spend page breaks dollars down by company, organization, department, owner, agent, model and provider, with top spenders and a watch list. |
| On which models, and how many tokens? | Each agent's Models tab lists every model it calls, with requests, tokens in, tokens out, cache tokens, cost and cost per day. |
| Who owns this money? | Every registered agent has a named owner, and the Spend page counts the dollars spent by agents with no live owner. See Identity Tethering. |
| What work did it pay for? | With ADE, the Argent Development Environment, a coding agent's spend is booked to the tracker story it works on and rolled up through its epic to its project. See Agent Work Tracking. |
| Is something off? | An unusual-spend check flags an agent whose spend spikes against its own 28-day median, and a loop detector spots an agent repeating the same paid step. |
The path: from a live view to a hard cap
Each step adds control, and nothing stops until you decide.
- Route agents through Argent. Agents call models with an Argent key. The company's AI provider key stays in its vault and Argent attaches it to each call on the way out, so by default an agent never holds it. See Keyless Agents.
- Find the AI nobody registered. Proxy-log import and managed machines show which AI services people reach, and unregistered clients land in the Agent inventory tagged Discovered. See Shadow AI Discovery and the AI visibility goal.
- Read live spend. Metering runs in every mode, so you see dollars and tokens by agent, owner and model from the first request.
- Put an owner on every agent. Each agent is tied to a named, active employee. When an owner leaves, their agents are refused on the next request and can be reassigned. See Owner Offboarding.
- Fund budgets while nothing stops. An admin sets budgets down the org chart. The budget enforcement switch ships off: until it is on, budgets count spend and refuse nothing.
- Watch. Turn budget enforcement on with the cost area in watch mode, and Argent records each call it would have refused while the call goes through. See Watch Before You Enforce.
- Enforce. Switch to enforce and every call must fit every budget above its agent, or it is refused. See AI Spend Caps.
How every call is counted
A cap is only as good as the meter behind it. Argent counts from the model's own reply and prices from a price book you control.
- From the reply. Cost comes from the usage figures in each model's own reply, across many reply formats, including streamed answers.
- Cache tokens priced their own way. Cache tokens are counted the way each provider bills them and priced at their own cache rates, so cached prompts are counted at the rate the provider charges for them.
- Your price book. Prices live in Argent's Pricing page, in US dollars per million tokens by provider and model. A price feed can be synced on a schedule, and its changes wait in Pending until a person approves them. Every price edit is a signed admin event.
- No free pass. A capped agent calling a model with no price is refused, so an unpriced model cannot get around a budget.
- On the record. Each signed record carries the cost of its request, so spend is provable from the record itself.
Budgets that act before the money is spent
Budgets follow the org chart: company, organization, department, person and agent, in dollars, for a day, a week or a month.
- Fund. You can only hand out what you hold: the budgets under a level never add up to more than it has. A person sees their own funding chain in the owner portal and splits their budget across their own agents.
- Price the call. Each request is priced before it is sent, on the high side: the prompt plus all the output it asks for (or a set default when it names no limit), at the model's catalog price.
- Hold it at every level. That amount is held against the agent, its owner, the department, the organization and the company, so many calls at once cannot slip past the cap.
- Allow or refuse. If every level can cover it, the call goes through. If any level cannot, it is refused: nothing reaches the model, and the reason names the level that ran dry.
- Settle. When the answer comes back, the hold is trued up to what the call actually cost.
More money is a person's decision. An agent never raises its own limit. The owner chooses Request more and says why; an approver who holds the money approves for now or for next period, or writes a reason to deny. Nobody approves their own request. Connect ServiceNow or Jira Service Management and Argent opens a ticket for every new request and pushes the decision onto it.
Brakes on top. An operator can set an emergency cap, a dollar limit over a rolling hour, day or week on any agent, person, department, organization or the company, and it stops in watch mode too. With ADE, an admin can cap a story, its epic or its project: at the cap, with Pause chosen, the next prompt is held until the budget owner chooses Raise cap or One-time overage.
Spend less for the same work
Capping spend stops the bill. These controls cut the waste before the cap has to.
| Control | What it does |
|---|---|
| Cheaper model near the limit | A budget can switch an agent to a cheaper approved model when it reaches a level you set (80% by default), so the work keeps going for less. At 100% the budget still stops it, and spend is metered at the price of the model that ran. See Model Control. |
| Only approved models | Only AI models you have approved can answer; a request to any other model is refused. See the model control and routing goal. |
| Model findings on each agent | The agent page points out a cheaper model, a price to watch and cache wins beside the agent that causes them. |
| Loops refused | The loop detector spots the same tool call with the same result, two steps back and forth, or the same error again and again. Set to enforce, it refuses the repeating request while the budget still has money. See Agent Behavior Watch. |
| Quarantine | An admin can quarantine a runaway agent from its page: every request from it is refused until it is released, and nothing is deleted. |
| Carry-over | Unused money can be added to the next period, up to one period's allotment, and an owner or admin can let an agent borrow from a shared pool instead of raising its cap. |
Proof finance and audit can check
A spend report says what was spent. A signed record proves it, and proves who decided.
- Every dollar is signed. Each signed record carries the cost of its request.
- Every decision is signed. Every refusal, budget change and approved raise is a signed record, and so is every call Argent would have refused in watch mode.
- Every price is signed. Every price edit and price-feed sync is a signed admin event.
- Periods close cleanly. At period close Argent writes settled figures, carries unspent money forward where carry-over is on, and reconciles its counters against the signed record.
- Branch admins see their own money. An organization or department admin sees only their own unit's budgets and spend.
Package it for an auditor with One-Click Audit Evidence, replay a costly run prompt by prompt with Session Replay, or take it to the board with Give the Board What It Wants.
Where to go next
Budget control works best with the goals around it.
- Capabilities on this path: AI Spend Caps, Model Control, Agent Work Tracking, Agent Behavior Watch, Keyless Agents, Identity Tethering, Owner Offboarding and Watch Before You Enforce. All capabilities.
- Related goals: see all AI in your company, model control and routing, audit and accountability, agent identity and ownership, data loss prevention, threat defense, compliance evidence and sovereign AI. All goals.
- Use cases: Securing AI coding agents, Rolling out copilots and chat safely, Governing customer-facing AI assistants, Controlling agent tool access and AI in regulated and air-gapped environments.
- Industries: Financial Services, Insurance, Healthcare, Public Sector, Retail and all industries.
- Where it runs: inside your own environment: air-gapped on-premises, private cloud or your public cloud VPC. Compare deployment options.
- Connects to: 20 AI providers and 60 AI clients and agents, plus MCP and agent-to-agent. Browse our integrations.
- See it yourself: take the walkthrough or download the data sheets.