Why AI costs us about 1% of revenue

At Oliverlist, AI costs about 1% of revenue. Yet it has been part of the product since May 2023: it classifies every prospect reply, drafts answers and writes call scripts. There is nothing magic about that number. It comes from a few simple choices you can make from your very first project.

Since July, I have added a second layer: agents that handle part of my own work as a CEO. Same logic, same discipline on cost. Here are the choices that matter.

AI only works when there is work

Each of my agents is a set of written instructions plus a schedule. The principle I apply is simple: every run starts with a check that involves no AI at all. Is there anything to process? A new reply, a new bank statement, a new report. If nothing has changed, the agent stops without calling a model. That run cost next to nothing.

The product works the same way. Sorter runs every 5 minutes, but a model is only called when a reply is waiting to be classified. The bill follows activity, not the clock.

This has a practical consequence: an agent can run often, every 5 or 15 minutes, without the bill taking off. You set the frequency by what the business needs, not by the budget.

It sounds obvious. It is not always the case: a badly configured tool re-reads everything on every pass, just in case. Over a month, it shows on the invoice.

The right model for each task

Not every task needs a top-tier model. Putting a reply into one of a handful of categories is short and repetitive. Writing a financial analysis or reviewing code is a different job. So each agent has its own model, chosen for its task.

The habit to build: test a lighter model on real cases, and only move up when quality requires it. Lighter models answer faster and cost a fraction of the price. On a task repeated thousands of times, that is where the bulk of the gap comes from.

Revisit those choices regularly. Models change quickly, and so do their prices. A task that needed an expensive model a year ago often runs on a lighter one today.

Three habits push the bill up without anyone noticing: sending the model the whole history when the latest items would do, retrying a failed task in a loop, and giving a simple sorting job to a model built for analysis. So track the cost agent by agent, every month. A bill that rises while the volume of work stays flat often points to one of those three causes.

Automation does the heavy lifting

The main reason for that 1% is not technical. It is about sequence: at Oliverlist, automation came before AI. Since 2020, the chain has been automated from contact import to invoicing, with one invoice per meeting generated automatically. Plain code like that costs almost nothing per run.

That is where the big productivity jump came from: a threefold increase between 2019 and 2023, before generative AI. I covered this in Automate before you talk about AI. AI came afterwards, and only where something has to be read or written: understanding a reply, drafting a message, spotting a meeting agreed in an email thread.

For the agents behind my CEO work, the foundation is just as modest: a small server and a database, for a few dozen dollars a month on top of the AI subscription.

The right benchmark

One percent sounds low. But compared with what? The wrong benchmark is the price of software. The right one is the fully loaded cost of a person doing the same work. That means salary, social charges and onboarding, plus the management time around the role. Run the numbers with your own figures, for a single task: sorting incoming email, preparing a weekly report, following up on overdue payments. The gap speaks for itself.

Run the calculation the other way too: how many hours a week does this task take today, and whose hours are they? The answer often surprises, because the work is scattered across several people, in small pieces, and nobody has ever added it up.

Two caveats, in fairness.

First, the AI bill is not the whole cost. The real cost is the time spent breaking work into tasks, writing instructions, measuring and monitoring. An unmonitored agent costs far more than its bill on the day it goes wrong and nobody notices.

Second, AI does not replace a person wholesale. It takes over tasks. Do the calculation task by task, and think at the same time about what stays with the person, often the part that calls for judgment.

Key takeaways

• Make AI work on demand, not on a timer: a check without AI first, a model only when there is work.
• One model per task, chosen on real cases and reviewed regularly.
• Automate with plain code first. Keep AI for reading and writing.
• Benchmark against the fully loaded cost of a person, task by task, including the time spent monitoring.