Scope the no-AI baseline you are trying to beat

An operations director at a restaurant chain told me about the question that stopped an AI project one meeting short of approval. The plan was to forecast how much of each item to prep at each location, so kitchens would stop over-prepping and throwing food away. It was a real problem, the waste was visible, and a demand model was the obvious answer. The room was ready to move.

Then she asked what the current method got right. Not what it got wrong, everyone could list that. What it got right. How close were the kitchen managers already, prepping the way they prepped now, on instinct and a paper guess sheet? Nobody knew. And without knowing, nobody could say what the AI needed to do to be worth building.

Better than what

Here is the thing the question exposed. You cannot justify an AI feature until you know what the thing it replaces already achieves. The current approach, however crude, produces some result. Kitchen managers guessing from experience hit their prep numbers with some accuracy. That accuracy is the bar. If the managers are already close most of the time, a model has very little room to improve, and the project might cost more than the waste it removes. If the managers are wildly off, there is real room, and the project is easy to justify. Same feature, opposite decision, and the only thing that tells them apart is a number nobody had measured.

Skipping the baseline is easy because the current method is not a system. It is people doing their jobs, so it does not feel like a thing you measure. But it is producing an outcome every single day, and that outcome is exactly what you are proposing to beat. If you do not know what it is, you are betting a build against an unknown, and “the AI is pretty accurate” means nothing without it. Accurate compared to what.

The baseline also sizes the prize. Once the operations director measured how much the current guessing already got right, she could estimate how much waste even a perfect model could remove. That is the ceiling on the project’s value. If the ceiling is low, no model is worth it, because you cannot beat a baseline that is already good by enough to pay for the build. She would rather learn that before funding it than after.

A colleague who runs support at a telecom described the same discipline from his side. His team wanted AI to route tickets, and before scoping it he measured how well the existing rule-based routing already did. It was right most of the time. The room to improve was small, and the honest conclusion was that a model was not worth building over a rule that mostly worked. The baseline made the decision for him.

Measure the current approach before you build

The move is to write down, in numbers, what the no-AI method achieves today, before you scope the AI. Human judgment counts. A simple rule counts. Whatever is running now is the incumbent, and the incumbent’s performance is the number the AI has to beat by enough to justify its cost.

Questions worth answering before funding the build:

What does the current method get right, beyond what it gets wrong, and how often?
If a perfect model replaced it, how much would actually improve?
Is that improvement large enough to pay for building and running the model?
Is the incumbent a rule or a habit that could just be tuned instead?

If you cannot state what the current approach achieves, you cannot state what better means, and better is the entire case for the project.

How we handle it at Density Labs

Measuring the no-AI baseline is a standard part of the AI Opportunity Assessment, our two-week fixed engagement at $2,500. Before we scope a model, we pin down what the current human or rule-based approach already delivers, so the client knows what the AI has to beat and by how much. Now and then the finding is that the incumbent is good enough that a model would not earn its cost, and that is a cheaper answer to reach during scoping than after a build that never had room to pay off.

You cannot know if AI is better until you know what you already have. Measure the no-AI baseline first, because better is meaningless without it.