What does cost-per-decision actually mean?
Cost per decision is what you spend, in euros, every time someone in the business evaluates a case and decides what to do with it. It isn't abstract: it's the sum of the time the person uses, the tools they need, the errors they occasionally make, and how long it takes them to act.
I used the example of a lead reaching my agency. Someone has to research the company, gauge the size of the business, qualify the interest, and write the first message. This isn't top-level strategic work, it's a decision unit repeated hundreds of times a month. If you don't isolate that cost, it stays hidden inside the payroll and you never see where the waste actually is.
If the client can't say where sales are coming from, that's where we start: first isolate the repeated decision, only then decide what to do about it.
Breaking down the formula: time, infrastructure, error, and delay
The formula I proposed combines four parts: human time multiplied by hourly cost, plus infrastructure cost, plus a percentage for error cost, plus a percentage for latency cost. I was explicit that the numbers I chose are an exercise and that everyone should adjust the percentages to their own reality.
In the example, a sales rep spends thirty minutes handling one lead, at a cost of 25 € per hour (salary, taxes, and other overhead tied to that person's seat). That's 12.50 € in time alone. Infrastructure, tools like a CRM or automation platforms, I spread out at 1 € per decision. Then come the two harder-to-see percentages: error, which I set at 20% because of poorly qualified leads and calls scheduled with the wrong people; and latency, another 20%, because a lead that waits days for a reply cools off or has already closed with a competitor.
Adding it all up, 0.5 hours × 25 € + 1 € + 20% + 20%, you get 19.44 € per decision. As I said in class, this is an exercise and the percentages can vary; building the habit of calculating it matters more than the exact figure.
How to use AI in business beyond chatbot prompts
The lead-triaging example: before and after adding AI
In the simulation, I compared the same process with and without AI support. In the manual scenario, the rep researches, qualifies, and writes everything from scratch, the same thirty minutes. In the assisted scenario, the system automatically gathers data about the company that submitted the form, compares it against past clients to score the lead, and prepares a draft reply.
The human then reviews that draft in about six minutes (0.1 hours), instead of doing all the work.
Keeping the same 25 € hourly cost and the 1 € infrastructure allocation, labor drops to 2.50 €. And because qualification now follows criteria set by whoever trained the system, rather than each rep's personal interpretation, error drops to 5% and latency to 3% (the reply goes out much faster). The final total comes to 3.79 € per decision.
I ended up with a decision cost roughly 15 € lower for each lead that came in. That difference, roughly 15.65 € per lead in this specific exercise, is what compounds with volume.
Volume and the trap of counting the same gain twice
If a business handles 500 leads a month, a saving of 15.65 € per decision comes to roughly 7,825 € in avoided cost that month, following the exercise's arithmetic. This is a cost reduction, not net profit, since other costs still apply. This saving only becomes real money if the freed-up capacity is actually used.
If the team now spends six minutes instead of thirty per lead but keeps the same salary without taking on more volume or higher-value work, nothing new has actually entered the company's account. What happened is free time that needs to be directed somewhere, either toward handling more leads without hiring, or toward reps spending more attention on negotiations that actually matter. That freed time has practical value even in small blocks, but it only turns into a result if someone decides what to do with it.
The simulation also lets us consider the commercial effect. In a hypothetical example with 500 leads, moving from a 5% close rate to 6% means closing 30 deals instead of 25. At a sale value of €1,000, that adds €5,000 in revenue from those sales, before delivery costs. If payments recur, future value depends on how long the relationship actually lasts. Calculate that separately from monthly operational savings so that you compare equivalent periods and avoid adding several months of expected revenue to a single month’s gains.
The four steps: observe, automate, add intelligence, scale
The cycle I proposed has four steps and always starts with observation, never with the tool. First you observe: you build an honest inventory of repetitive tasks and recurring decisions, and look at where the team is spending the most time without proportional return.
Then comes automation, which often doesn't even need artificial intelligence. If the goal is moving a lead from a form into a CRM, that's simple integration, using tools like Make or n8n. I insisted on this point: poorly collected data makes any layer of intelligence useless, because the AI ends up guessing instead of deciding based on real context.
The third step is where intelligence itself comes in: classifying, prioritizing, summarizing, and suggesting the next step, always based on history and rules the business has defined, not generic model opinion. The fourth step is scaling with human curation, batch approvals, alerts when something falls outside the pattern, and the team validating instead of manually doing everything.
Where the exercise has limits, and what it leaves out
I stressed that an experienced person brings things a machine doesn't have: prior knowledge of companies and people, sensitivity to unwritten context, a network of contacts. People are worth much more than AI on that nuance, even though, for producing more and generating more revenue, I prefer automation's numbers.
Give me the data, explain the rule, show me where it failed is, at its core, the spirit of the exercise: artificial intelligence serves as a gathering and drafting assistant, not the final decision-maker on financial commitments or reputation. Calculating cost per decision isn't an end in itself, it's how you figure out where technical investment pays off and where it still makes sense to keep a person deciding with proximity.
If you want to apply this to your business, the starting point is the same one I suggested in class: list the repetitive tasks in your day-to-day, measure how much time each one takes, and only after having that list ask which tool or automation makes sense. You might also read our take on margins, teams, and AI in online businesses and on triaging and qualifying B2B opportunities with AI.
Source note
This article is based on the class I gave in Sextas Ímpares #128, "The Era of Business Automation Intelligence," a recording of a talk I delivered at Ad Summit, available at https://www.youtube.com/watch?v=XTWg_GjK35Q. The section on cost per decision and the lead simulation starts around 14:00, the full calculation with the numbers appears near 27:30, and the four-step cycle of observe, automate, add intelligence, and scale is explained starting at 41:40. The values and percentages used in the class are an illustrative exercise, not a measured saving from any specific business.