---
title: How do you forecast AI agent spend for next year's budget?
url: https://opencontroller.lyzr.com/blog/forecast-ai-agent-spend
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# How do you forecast AI agent spend for next year's budget?

**Forecast AI agent spend with a driver-based model: number of agents × runs per agent × tokens per run × blended price per token, plus seat licences as a separate line. Then run three or four scenarios — price drops, a heavier model mix and faster adoption — because each driver can move by 2× in a year. Base the drivers on measured data from your gateway or provider logs, not on vendor list prices alone.**

This post belongs to the gateways and cost cluster; the pillar is [LLM gateway vs agent gateway vs control plane](/blog/llm-gateway-vs-agent-gateway-vs-control-plane).

## Why does last year's spend not work as a forecast?

Agent spend does not grow in a straight line. Three things break trend-based forecasts:

- **New agents launch in steps.** One successful pilot can add ten production agents in a quarter.
- **Model choice changes cost per run.** Switching to a larger or reasoning model can multiply cost per task; see [how reasoning tokens change agent cost](/blog/reasoning-tokens-and-ai-agent-cost).
- **Per-token prices fall.** Providers have repeatedly cut prices for comparable capability, which pulls the other way.

A driver model separates these so finance can see which assumption moved.

## What are the drivers?

| Driver | Where to get it | Typical owner |
|---|---|---|
| Agents in production (start and end of year) | Agent inventory plus the launch roadmap | Platform team |
| Runs per agent per month | Gateway or platform logs | Agent owners |
| Tokens per run (input and output separately) | Gateway logs | Platform team |
| Blended price per million tokens | Provider pricing × current model mix | FinOps |
| Seat licences | Contracts for AI assistants and copilots | Procurement |
| Platform and tooling | Contracts for agent platforms and governance | Procurement |

If you cannot fill "tokens per run" today, that is the first gap to close. You need per-agent cost attribution; see [how to attribute LLM costs to AI agents](/blog/how-to-attribute-llm-costs-to-ai-agents).

## What does a worked example look like?

All numbers below are illustrative.

- Agents: 12 in production today, 30 planned by year-end. Average over the year: 21.
- Runs per agent: 20,000 per month.
- Tokens per run: 40,000.
- Blended price: $3 per million tokens.

Monthly cost per agent = 20,000 × 40,000 = 800 million tokens × $3 per million = **$2,400**.
Annual usage cost = 21 agents × $2,400 × 12 months = **$604,800**.

| Scenario | Change from base | Annual usage cost |
|---|---|---|
| Base | — | $604,800 |
| Price drop | Blended price falls 30% | $423,360 |
| Heavier model mix | 20% of traffic moves to a model at 5× the price (blended $5.40) | $1,088,640 |
| Adoption doubles | Runs per agent double | $1,209,600 |

Seat licences sit on their own line. For example, 2,000 seats at $30 per user per month is $720,000 a year — and it does not scale with usage. Keep the lines apart so a usage spike is not hidden by flat seat costs; see [combining seat licence and token costs](/blog/combine-ai-seat-licence-and-token-costs).

## How do you build the forecast step by step?

1. **Export three months of per-agent usage.** Runs, input tokens and output tokens per agent.
2. **Group agents into cohorts.** Live, launching this half, launching next half.
3. **Set drivers per cohort.** New agents often start at low volume and ramp over two to three months.
4. **Price the current model mix.** Use the provider price pages and your real input-to-output ratio.
5. **Add scenarios.** At least: price drop, heavier model mix, adoption doubles.
6. **Add seat and platform lines.** From contracts, not usage.
7. **Agree review triggers.** Re-forecast when an agent changes model, when a cohort launches, or when monthly spend is 20% off plan.

## How do you keep actuals close to the forecast?

- Set [budgets and alerts per agent](/blog/set-budgets-and-alerts-for-ai-agent-spend) that match the forecast lines.
- Report actuals by agent and version monthly, using the same drivers.
- Treat a model switch as a budget event that needs a re-baseline.
- Use caching and routing to keep tokens per run down where quality allows.

## How Opencontroller does this

Opencontroller's LLM Gateway (Lyzr Gateway) attributes cost per agent and per version, which gives you measured runs and tokens per run instead of estimates. Auto-discovery of agents, models and API keys across AWS, Azure and Google Cloud gives the agent count. Routing, fallback, retry and caching are configured in the gateway; in the October 2026 benchmark, caching lowered model cost by 23.5% on mixed synthetic traffic under the stated test conditions. [Book a demo](https://www.lyzr.ai/book-demo/) to see per-agent cost data.

## Frequently asked questions

**Should we forecast per agent or per team?**
Model per agent cohort, then roll up to teams for budget ownership. Per-team-only forecasts hide which agent drives a change.

**How often should we re-forecast AI spend?**
Quarterly at minimum, and whenever an agent changes model or monthly spend is about 20% off plan.

**Do falling token prices mean we can budget less?**
Not automatically. Price cuts are often outweighed by more agents, more runs and larger models. Model them as one scenario, not the base.

## Key facts

- AI agent spend can be forecast with a driver model: agents x runs per agent x tokens per run x blended price per token, plus separate seat and platform lines.
- Each driver can move by about 2x in a year, so forecasts should include price-drop, model-mix and adoption scenarios.
- In an illustrative example, 21 average agents at 20,000 runs and 40,000 tokens per run at $3 per million tokens cost $604,800 a year.
- Seat licences should be a separate budget line because they do not scale with usage.
- Opencontroller's LLM Gateway attributes cost per agent and per version.
- Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you attribute LLM costs to individual AI agents?](https://opencontroller.lyzr.com/blog/how-to-attribute-llm-costs-to-ai-agents.md)
- [How do you set budgets and alerts for AI agent spend?](https://opencontroller.lyzr.com/blog/set-budgets-and-alerts-for-ai-agent-spend.md)
- [How do you combine AI seat licences and token spend into one cost view?](https://opencontroller.lyzr.com/blog/combine-ai-seat-licence-and-token-costs.md)
- [More on Gateways and cost](https://opencontroller.lyzr.com/blog/cluster/gateways-and-cost)
