An AI Agent That Re-Reads Its List Every 15 Minutes Pays for It 96 Times a Day

Ask what it costs to run a fleet of AI agents and you will usually get a model price per million tokens. That number matters less than people think. The bigger driver is a habit: how often your agents re-read things they have already decided. This explainer shows why, with a worked example built on round, made-up numbers so you can redo the sums for your own setup.

Label first. The example below is hypothetical. It describes no real firm and no real bill. Every figure is an assumption, shown so you can swap in your own.

How agents run up a token bill

AI agents are billed in tokens, the small chunks of text a model reads and writes. Every time an agent wakes up, it reads its instructions and whatever it needs to look at before it acts. Read a lot, and often, and the bill climbs even when little gets done. Reading is usually billed per million input tokens. Writing is billed separately, often at a higher rate, but most agents spend far more of their day reading than writing.

So the cost of an agent team is roughly: how much each agent reads per wake, times how many times it wakes, times the price of reading. The model price is only one of those three numbers.

The worked example: a lead list read 96 times a day

Picture a small firm with a team of agents finding and following up sales leads. The shared lead list lives in a notes page with 2,000 rows. Assume reading the whole list costs about 40,000 tokens, plus about 5,000 tokens of instructions each time an agent wakes.

Now add a timer. A send clock wakes the pipeline every 15 minutes to check whether anything is ready. That is 96 wakes a day, and on each wake the agent ranks the list by reading it end to end.

• Before: 96 wakes × 45,000 tokens = about 4.3 million tokens a day. At an assumed $2 per million input tokens, that is about $8.64 a day, roughly $260 a month, before any writing.

Most of those wakes find that nothing has changed. The agents are paying, every 15 minutes, to rediscover yesterday's decisions.

The change: score once, at the source

Now change one habit. Each lead gets a score, say 0 to 10, once, when it first arrives. The list moves into a simple database, so a shortlist of leads scoring 7 or more is one query that returns a few rows instead of a long read. The timer goes, and agents wake only when a new shortlisted lead appears.

• Scoring on arrival: assume 50 new leads a day at about 1,000 tokens each = 50,000 tokens.
• Waking on real work: assume 10 wakes a day at about 5,000 tokens each = 50,000 tokens.
• After: about 100,000 tokens a day, or about $0.20 a day at the same assumed price, roughly $6 a month.

In this example daily token use falls by more than nine-tenths, and the model price never changed. Nothing about the work got lazier either. The agents still find, score, research and draft. They just stop re-reading what they have already decided. Your own numbers will differ, but the shape is common: the waste sits in the reading pattern, not in the model.

Hypothetical example. Token counts and the $2-per-million input price are illustrative assumptions in line with the mid-range model prices quoted in the guide to what AI really costs a small firm; output tokens are excluded for simplicity.

Why the token bill is still the wrong thing to stare at

Even in the "before" case, $260 a month may be smaller than the cost of the time a person spends checking the agents' work, or the cost of a bad email going out under your name. Tokens are the raw material. What you are buying is finished work. That is the case made in price 1 finished task, not 1 million tokens: count the tokens, the subscriptions and the minutes people spend prompting, checking and fixing, then divide by the jobs that were actually finished to your standard.

Rework is the other hidden line. MarketScale's summary of a July 2026 McKinsey report put response refinement, agents checking, revising and regenerating their own answers, at about 60% of agentic AI costs. That is covered in most agentic AI spend goes to fixing answers. A cheap token bill on an agent that produces work no one can use is still expensive.

What to check if you run agents

• Find the re-reads. List what each agent reads on every wake. If the answer doesn't change between wakes, store it instead of rediscovering it.
• Score at the source. Decide once what deserves attention, and let only that move downstream.
• Kill timers that check for nothing. Wake agents when there is work, not on a clock.
• Move shared memory into a real database once the list gets big. A notes tool is fine while it is small.
• Measure cost per finished job, with a person's review time included.

If you are not sure an agent is the right tool at all, the test on AI agents for business helps you decide between an agent, plain automation or nothing. For how a team of narrow agents divides the work, see one giant AI assistant or twenty narrow ones.

Frequently asked questions

What does it cost to run AI agents?

It depends on three numbers: how much each agent reads per wake, how often it wakes, and the model's price per million tokens. In the hypothetical example here, the same agents cost about $8.64 a day when they re-read a 2,000-row list every 15 minutes, and about $0.20 a day when leads were scored once on arrival and agents woke only for real work.

Why are my AI agent token costs so high?

Usually because agents re-read large, unchanged material on a schedule: whole lists, long documents or full chat histories. Check what each agent reads on every wake, and whether it needs to.

Is a cheaper model the best way to cut agent costs?

Not first. Cutting how much and how often agents read can save more than switching models, and it doesn't risk quality. Then compare models on cost per finished job at your quality standard.

Want a second pair of eyes on your agent costs?

Tell Ada about one loop of work your agents, or your people, repeat every day. She'll help you find where the re-reading and rework sit, and put it in your free report. Talk to Ada, or see how BuildBrain approaches AI agent setup for small firms.