Episode transcript The transcript is generated automatically by Podscribe, Sonix, Otter and other electronic transcription services.
Hello everyone, here is Ricardo Vargas, and this is the 5 Minutes Podcast.
Let me describe a scene that is happening in thousands of organizations right now. A team runs a proof of concept with artificial intelligence. The vendor gives free credits, the pilot costs almost nothing, the results look impressive, and everyone in the steering committee is delighted. The project gets approved. And then it goes into production, and an invoice starts arriving every month, forever.
I have spent most of my career around capital projects and business cases, and this is the part of the AI conversation that we are not having. We talk about capability, about ethics, about jobs. We almost never talk about the bill.
So let me offer three things I think we should be much more careful about.
The first is that artificial intelligence is rarely a project cost. It is an operating cost wearing the clothes of a project cost. We approve it inside a project budget, with a start and an end, and then it lives in the operating expenses of the company for as long as the solution runs. The project closes, the team disperses, and the invoice stays. So the question we should ask before approval is very simple. Who owns this line in year three, when none of us is around anymore? If we cannot name that person, we have not finished the business case.
The second is that this cost behaves in a way most of our budgeting was never designed for: a traditional license is fixed and predictable. Consumption of an AI service is variable, and it scales with usage. Which means the more successful we are, the more we pay. Let that land for a moment, because it inverts something we are used to.
In most projects, the success scenario is the cheap scenario. Here, the success scenario is the expensive one. If adoption triples, the bill triples, and our business case assumed a flat number.
The third is the set of costs that never appear in the vendor's proposal. Preparing the data and keeping it prepared. Integrating with the systems that actually run the company. The human review of what the model produces, which is real work done by real people who are paid. And the cost of change that we do not control, when a vendor adjusts pricing, retires a model, or replaces it with a new one that behaves differently and forces us to test everything again.
So what would I suggest we do?
Put a running cost line in the business case, and express it as a unit rate rather than a lump sum. Cost per document processed, cost per customer interaction, cost per report generated. A lump sum hides the risk. A unit rate lets us multiply it by tomorrow's volume, which is the number that will actually hurt us. Ask the vendor, in writing, what happens when our volume triples. Not what it costs today, at pilot scale, with promotional credits. What it costs at the volume we are promising the board we will reach. Name the owner of the bill before go-live, not after. And model the exit while we still have leverage, because the cost of leaving is part of the cost of entering.
None of this means we should slow down on artificial intelligence. On the contrary. The organizations that will get the most out of AI are the ones that can keep funding it in the third year, and that is a financial discipline question much more than a technology question.
Here is the question I would take into our next steering committee.
Do we know what the AI in our project can do? Do we know what it will cost when it is doing it at full scale? Who pays for it after the project ends, and whether the value we measured is still larger than that number?
If we cannot answer those three, we do not have a business case yet. We have a pilot with good manners.
I hope you enjoyed this episode, and see you next week with another 5 Minutes Podcast.
