Reference · 2026
Agent Spend as Percent of Revenue
How to work out what share of revenue your agents should consume — and why the honest answer is a calculation you run on your own numbers, not a benchmark someone hands you.
By the sipi.bot engineering team · Published 2026-07-18 · Last updated 2026-07-25 · Methodology
Why there is no honest benchmark here yet
A credible agent-spend-to-revenue ratio needs two numbers per company — inference spend and revenue — for a defined population. That data is not public, and sipi.bot has no telemetry from which to derive it. Any site quoting a precise median for this ratio should be asked where the revenue figures came from.
What you can do instead is calculate your own, which is the number that actually governs your decisions.
Working out your own ratio
- Total agent spend. Sum inference and tool-call spend across providers for a full month. Published per-model rates are in ai-model-costs-2026.csv.
- Divide by revenue for the same month. Use recognised revenue, not bookings, or the ratio will flatter you.
- Separate build from run. Development and evaluation spend behaves differently from production serving; a ratio that mixes them will move for reasons unrelated to your product.
- Track the trend, not the level. The useful signal is direction against revenue growth. A ratio rising while revenue is flat is the thing to act on, whatever the absolute number.
What actually moves the ratio
These are mechanisms, not measurements — each is something you can verify against your own bill:
- Model tier. Running a frontier model where a smaller one suffices is usually the single largest line. Compare published rates before optimising anything else.
- Retry behaviour. Unbounded retries multiply cost without changing output. A task costing $0.34 retried 40 times costs $13.60.
- Context size. Re-sending long context on every turn is billed on every turn.
- Unattended sessions. Spend that continues when nobody is watching is bounded only by rate limits and your balance.
A pre-spend firewall addresses the last two directly: a velocity cap bounds retry loops, and a daily total cap bounds unattended sessions.
Methodology
What this page is based on: published provider pricing (ai-model-costs-2026.csv) and the failure modes exercised by the open-source evaluation suite — 53 scenarios you can run yourself with pip install sipi-bot.
What it is not based on: any survey of companies, and any customer or deployment telemetry. sipi.bot collects none — that is a design property of the product — so it publishes no medians, ranges or percentiles for this ratio.