Why we start with questions, not technology
When a COO reaches out to us, they almost always arrive with a formed idea: "I need to automate reporting," "I want an agent to handle internal inquiries," "I've been told AI can do this."
The idea isn't the problem. The problem is that the right technology applied to the wrong process produces no results. It produces a project no one uses.
That's why, before proposing any agent, we ask eight questions. They are not technical questions. They are business questions. And the answers determine whether it makes sense to move forward, what makes sense to build, and what result is reasonable to expect.
Question 1: How many times does this process occur per month?
Frequency is the first filter. A process that occurs twice a month does not justify the same investment as one that occurs two hundred times.
There is no universal threshold, but as a reference point: processes with fewer than twenty monthly occurrences rarely generate sufficient ROI to justify a dedicated agent. They can be incorporated into a broader workflow, but they are not the right starting point.
Question 2: How much time does each occurrence consume?
Volume without duration tells you nothing. You need both data points to estimate potential impact.
A process that occurs one hundred times a month and takes four minutes each time represents roughly seven monthly hours. If an agent resolves it in seconds, the savings are real but modest. If that same process takes forty minutes per occurrence, you're looking at more than sixty monthly hours. At that point, the calculation changes.
Question 3: Who handles it today, and what is their hourly cost?
We don't ask this to reduce headcount. We ask it to understand the opportunity cost.
If the process is executed by a financial analyst with a total cost of forty euros per hour, and that analyst spends thirty monthly hours on tasks an agent can handle, the opportunity cost exceeds EUR 1,200 per month. That figure makes it possible to evaluate whether the implementation investment makes sense over a six-to-twelve-month horizon.
Question 4: How much variability does the process have?
This question separates viable cases from those that are not.
A highly standardized process — same inputs, same rules, same expected output — is a direct candidate for automation. A process where each case requires distinct judgment, negotiation, or context that is not documented in any system is a weak candidate, at least at this stage.
We don't discard variable processes. We defer them until the team has the capacity to supervise them properly.
Question 5: Are the necessary data available and structured?
An AI agent does not generate information. It processes it. If the data it needs to operate are scattered across emails, unstructured spreadsheets, or the institutional knowledge of three people, the first task is not to build the agent. It is to organize the data source.
This question prevents projects that fail in production because the agent has nothing to work with.
Question 6: What happens if the agent makes an error?
The level of error tolerance defines the system's architecture.
An agent that answers internal queries about HR policies can operate with a reasonable error margin: if it fails, someone corrects it. An agent that processes payments or generates documents with legal implications requires human validation at critical checkpoints.
No AI agent operates with zero errors. What exists are systems designed so that errors are detectable, correctable, and low in impact.
Question 7: Who will operate and supervise the agent?
This is the question most often omitted in projects that fail.
An agent in production needs someone to supervise it, detect when it begins to degrade, and update it when process rules change. If no one is assigned with the time and judgment to do this, the agent deteriorates within weeks.
We don't ask for a technical team. We ask for a business-side person who understands the process and has two to four hours per week to supervise the system.
Question 8: What result would define success at ninety days?
If there is no agreed metric before the project begins, there is no way to evaluate whether it worked.
The metrics we use most frequently: hours recovered per week, error rate before and after, process cycle time, volume of cases resolved without human intervention. Any of these works. What doesn't work is "improving efficiency" without a concrete number behind it.
A concrete example: diagnostic at a distribution company
A distribution company with one hundred and twenty employees contacted us to automate supplier incident management. The process consumed, by their own estimate, roughly forty monthly hours across two members of the procurement team.
When we applied the diagnostic, we found that sixty percent of incidents followed a repeatable pattern: same category, same response, same supplier. The remaining forty percent required judgment and negotiation.
The proposal was not to automate the entire process. It was to build an agent that classified and resolved the standardized sixty percent, and routed the remaining forty percent with context already prepared so the team could resolve it in less time.
Estimated impact in the first quarter: between twenty and twenty-five monthly hours recovered, with an implementation cost amortizable in four to six months. The agent has been in production for three months, supervised by one member of the procurement team at two hours per week.
What we do with the answers
The eight questions don't produce a report. They produce a decision.
If the process has volume, the data are available, there is reasonable error tolerance, and someone exists to supervise the system, we move forward. If any of those conditions is not met, we say so before proposing anything.
That is the difference between a useful diagnostic and a sales presentation.
If you'd like to apply this process to one or two specific cases in your operation, you can request it at no cost. It is not a sales call. It is a forty-five-minute working session where we evaluate together whether it makes sense to move forward and, if so, where to start.
Request the complimentary diagnostic using the form below.
