Budgets and AI Spend

Every AI call through the gateway is priced as it happens, so you can see what your organization spends and cap it before the invoice arrives. Budgets set the limits. AI Spend shows where the money went.

Both live under LLM Governance in the portal.

Quick Start: What do you want to do?

I want to...	Go to
Stop one person running up a large bill	Giving everyone their own allowance
Put a ceiling on total company spend	Setting a shared budget
Stop a runaway agent	Capping an AI agent
Work out which limit applies to someone	Checking what applies to a person
See who is spending what	Reading AI Spend
Try limits without blocking anyone	Blocking versus tracking

---

The three kinds of budget

When you create a budget the first question is what you are limiting.

Kind	What it does	Typical use
Each employee's spend	Everyone covered gets their own allowance of the same size	A monthly allowance per employee
A shared budget	One budget pool shared by everyone it covers	A ceiling on total spend by a team or the company
An AI agent's spend	Caps what an agent persona spends, whoever runs it	Stopping a runaway automation

Shared and agent budgets stack. A single call must fit inside every budget that covers it — the caller's own limit, any shared budget they fall under, and the agent budget if they are using a persona.

Person limits are the exception: exactly one applies to any given person. See Checking what applies to a person.

Creating a budget

Go to LLM Governance → Budgets and select New budget. If you have no budgets yet, three starting points are offered instead — a monthly limit for each person, a company-wide cap, or a limit for an AI agent — and you can change every detail before saving.

1. Choose what you want to limit — each person's spend, a shared budget, or an AI agent's spend.
2. Choose who it covers — everyone, or specific teams. Teams include every unit beneath them, and people join automatically as your identity provider adds them.
3. Set the spend limit in dollars, per month or per day.
4. Choose what happens when it's reached.
5. Optionally set different amounts for specific people.
6. Name it, and save.

A preview on the right shows the budget in plain language as you build it. Changes reach the gateways within about a minute.

Giving everyone their own allowance

Choose Each employee's spend, covering Everyone. Every employee gets the full amount for themselves, so one heavy user cannot consume everyone else's headroom. A team with its own limit replaces this one for that team's members.

Setting a shared budget

Choose A shared budget. One amount covers the whole group — the first people to spend it use it up. This is the right shape for "we will not spend more than $10,000 a month on AI, whoever is spending it".

Capping an AI agent

Choose An AI agent's spend and pick the personas. The budget counts every call those personas make, whoever runs them, and applies on top of the caller's own limit.

An Agent Persona carries its own spend limits alongside its other policies, so if you are already deciding what an agent may do, capping what it may spend belongs in the same conversation.

Blocking versus tracking

Option	What happens
Block new AI requests	Requests are refused once the budget is used up, with a message naming the budget and when it resets
Keep working, just track it	Nothing is blocked. Overspend simply shows on the Budgets page.

Tracking still counts spend. A budget set to track accumulates its totals exactly as a blocking one would, so when you switch it to block you start from accurate figures rather than zero. Running a new budget in track mode for a cycle is the safest way to find out whether the number you picked is realistic.

Setting different amounts for specific people

On a per-employee budget you can list individuals by email with their own amount. A listed employee gets that amount instead of anything else — it replaces even a team limit. Use it for the few people whose work genuinely costs more, rather than raising the limit for everybody.

Checking what applies to a person

When several budgets could cover someone, exactly one sets their personal limit. Check a person on the Budgets page shows which one applies, and why the others did not.

The gateway resolves this with a fixed precedence, evaluated top to bottom until one rule settles it:

Named exception → Team limit over company-wide → Most specific team → Tie-breaker setting

Precedence	Rule	How it resolves
1	Named exception	An amount set for that person by email wins outright, overriding every team and company-wide limit
2	Team beats company-wide	With no named exception, a limit on one of the person's teams takes priority over a limit covering everyone
3	Most specific team wins	Among team limits, a sub-unit's own limit replaces the one it would otherwise inherit from its parent
4	Tie-breaker setting	If two team limits are equally specific, the tenant setting Someone in more than one team gets decides — set to either the higher or the lower of the two

Shared budgets and agent budgets sit outside this precedence. They apply in addition, and a call must fit inside all of them regardless of which personal limit wins.

What happens when a budget is reached?

The request is refused with 403, formatted as an error from the AI provider the agent was calling. The message names the budget and says when it resets, so the person can tell what happened without going to an administrator.

The response is marked as not worth retrying.

note

If the gateway cannot reach the service that tracks spend, calls are allowed through rather than refused. Budgets are a cost control, not a security boundary.

Reading AI Spend

LLM Governance → AI Spend shows what has been spent, for this month by default, and you can switch to the last 7, 30 or 90 days.

Group the figures by Teams, Users, Agent Personas, Models or Providers, and select any row to drill into it. Grouping by provider is usually where a surprise shows up, and each LLM Registry connection also shows the budgets covering it and how much of each is left. The four cards at the top show total spend with a comparison against the previous period, average per day, the number of AI calls, and how many were blocked at a budget limit.

Two things to keep in mind when reading the numbers:

• Someone in two teams counts toward both, so team totals can add up to more than the organization total.
• The list shows the top 200 by spend.

How accurate are the figures?

Spend is calculated as each call happens, using your contracted rates where you have set them and public list prices otherwise. It is designed to be useful for allocating cost and spotting runaway usage.

caution

Treat AI Spend as a close estimate, not a bill. The size of each request is estimated rather than measured exactly, and nothing reconciles these figures against the invoices your AI providers send. Use it to manage usage and settle budgets; use the provider's invoice for accounting.

Contract rates you set apply to new calls only — past spend is not recalculated.

If your agents use more than one gateway

Each gateway counts spend on its own. A tenant served by several gateway pools therefore allows up to the full budget amount in each one — a $1,000 monthly budget across three pools can permit up to $3,000.

This is shown as a note under the enforcement options when you create a budget. If it matters to you, size the budget accordingly or keep the agents that share a budget on one pool.

Real-world examples

Keep a lid on experimentation without policing it

Setting	Value
What	Each employee's spend
Who	Everyone
Limit	$100 per month
When reached	Block new AI requests

Everyone can explore freely, and nobody can quietly spend thousands.

Protect against a runaway automation

Setting	Value
What	An AI agent's spend
Who	The nightly report agent persona
Limit	$50 per day
When reached	Block new AI requests

A retry loop stops within a day's budget instead of running all weekend.

Tips

• Start in track mode for one cycle. It is the only reliable way to find out whether your number is right, and it costs nothing to be wrong.
• Use day budgets for agents and month budgets for people. An agent fails fast and loudly; a person's usage is lumpy across a month.
• Set exceptions rather than raising everyone's limit. A handful of heavy users should not define the default.
• Check a person before you field the complaint. It answers "why was I blocked" in one screen.
• Revisit after a model change. Moving to a more expensive model can quietly double spend against an unchanged budget.