Don't get carried away by leaderboards. What matters is how a model holds up in your company.
You deploy the model that won the leaderboard. Customers wait a long time for answers. Some answers are off. The bill is higher than you planned. The leaderboard tested one thing. Your operations test everything else. That's exactly why many AI projects end at the pilot stage.
You always pick two of three: accuracy, speed, cost
You can draw it as a triangle. Accuracy at the top, speed and cost in the corners. You can get two properties at once, the third suffers. A large, accurate model serving hundreds of people at once will be expensive. A cheap, fast model will be less accurate.
Testing is how you find out where your solution really sits in that triangle. And find out before your customers do.
With a person
A candidate has excellent grades from university. Whether they can handle your toughest client at four on a Friday, you only find out during the probation period.
With AI
The leaderboard is the grades. Your data, your customers and your busiest hours are the probation period. Nobody can do it for you.
First question: does it answer correctly in your field?
Leaderboards work like a standardized test. The model gets questions on chemistry, history or programming, and the answers are compared with the correct ones. But most business questions don't have a single correct answer. There are a hundred good ways to answer a customer's question.
That's why open-ended answers are evaluated with a method called LLM as a judge: another strong model rates the answers against criteria set in advance. Is the tone appropriate? Did the answer help? Did AI make anything up about your products?
One condition applies. The judging model can't be trusted blindly. First, an expert from the company rates a sample of answers. The judging model is calibrated to that, and only then rates hundreds more.
“An online shop's customer line. The head of support rates 50 real AI answers: good, bad, and why. Based on her ratings, the judging model then checks hundreds of new answers every week and flags the problematic ones. She only reviews what was flagged.”
Second question: can it cope when many people ask at once?
A model can be accurate and still unusable. That's why four things are measured, which we translate into operational terms:
- How long a person waits for the first word of the answer. In a chat, that's the difference between a smooth conversation and a closed window.
- How fast the rest of the answer arrives.
- How long the whole answer takes.
- How many people the system serves at once.
It's important to test with the kinds of questions that really come in. A chat with a short question and a long answer loads the system completely differently from a long document that needs a short summary. If you test the wrong type of questions, you get completely different numbers from what you'll see in production.
The result of the test is a breaking point. Up to a certain number of simultaneous users, response time is stable, then it rises sharply. That point is the number you use to plan capacity and budget.
With a person
Reception copes with a quiet morning. Whether it copes with Monday morning, when twenty people call at once, you only find out by trying.
With AI
We test load the same way. Not an average day, but the busiest hour, when the most people ask at once.
With an agent, you test every step, not just the result
An agent isn't one answer. It's a chain of decisions: did it understand what the person wants? Did it pick the right tool? Did it find the right material in the documents? Every link in the chain is a place where something can go wrong, so each one has its own test.
Testing can be shown as a pyramid. Many teams start at the top, with expert tests. But the pyramid stands on its foundations:
| LEVEL | WHAT WE TEST |
|---|---|
| 1. Operations | Is the system fast enough and can it handle the number of users? |
| 2. Format | Does the answer arrive in the shape the next system or person expects? |
| 3. Security | Is the answer free of personal data that shouldn't be there? |
| 4. Facts | Is the answer based on your data, not a guess? |
| 5. Expertise | Does it answer like an experienced person in your field? |
Ask yourself
If your AI started answering worse today, would you notice, or would your customers?
How we test so AI still works a month later
Testing doesn't end at launch. Models get updated, documents change, customers ask differently. That's why, for every solution, we build a set of check questions from your practice and run it regularly. When results get worse, we know before you do.
Before launch, we also measure speed and cost under your expected load. The model decision then rests not on a leaderboard, but on numbers from your company.
A leaderboard is a good start. The real test happens on your data, with your number of users and your measure of a good answer.
