The intelligence of AI models does not prevent enterprises from adopting them. The problem is that quality cannot be defined and measured in the first place.
Now, AI judges are playing an increasingly important role. In AI evaluation, "judge" is an AI system that scores the output from another AI system.
Judge Builder is Databricks’ framework for creating judges, first introduced as part of Company Components. agent brick The technology was introduced earlier this year. The framework has evolved significantly since its initial launch in response to direct user feedback and adoption.
While early versions focused on technical implementation, customer feedback revealed that the real bottleneck was organizational alignment. Databricks currently offers a structured workshop process that guides teams through three key challenges: getting stakeholders to agree on quality standards, gaining expertise from a narrow set of subject matter experts, and implementing evaluation systems at scale.
"Model intelligence is usually not the bottleneck. The model is very smart;" Jonathan Frankl, Chief AI Scientist at Databricks, told VentureBeat in an exclusive briefing. "Instead, what really matters is how we get the model to do what we want it to do, and how we know if it did what we wanted it to do."
The “ouroboros problem” of AI evaluation
Judge Builder refers to what Databricks research scientist Pallavi Koppol, who led its development, calls: "Ouroboros problem." Ouroboros is an ancient symbol depicting a serpent eating its own tail.
Using AI systems to evaluate AI systems creates the challenge of circular validation.
"You want the judge to see whether your system is good, whether the AI system is good, and in that case the judge is also the AI system." Coppol explained. "And now you’re saying, how do we know this judge is good?"
The solution is measurement "Distance to human expert ground truth" as a primary score function. By minimizing the gap between how AI auditors score output and how domain experts score output, organizations can rely on these auditors as scalable substitutes for human evaluations.
This approach is fundamentally different from traditional approaches guardrail system or evaluation of a single indicator. Rather than asking whether AI output passes or fails common quality checks, Judge Builder creates highly specific evaluation criteria tailored to each organization’s expertise and business requirements.
The technical implementation is also distinctive. Judge Builder is integrated with Databricks’ MLflow. Rapid optimization It has tools to manipulate any underlying model. Teams can manage versions of auditors, track performance over time, and deploy multiple auditors simultaneously across different quality dimensions.
Lessons learned: Building judges that actually work
Databricks’ work with enterprise customers has revealed three important lessons that apply to anyone building an AI judge.
Lesson 1: Experts don’t agree as much as you think. When quality is subjective, organizations find that even their subject matter experts disagree on what constitutes acceptable output. A customer service response may be factually correct, but use an inappropriate tone. A financial summary may be comprehensive, but it’s too technical for your target audience.
"One of the biggest lessons of this whole process is that every problem becomes a human problem." Frankl said. "The hardest part is getting the idea out of someone’s head and making it clear. And what’s even more difficult is that companies aren’t one brain, they’re many."
This modification is batched annotation with inter-rater reliability checks. The team annotates the examples in small groups and measures the match score before moving on. Detect discrepancies early. In one case, after three experts gave ratings of 1, 5, and neutral for the same outcome, discussion revealed that they had different interpretations of the rating criteria.
Companies using this approach have achieved interrater reliability scores as high as 0.6, compared to a typical score of 0.3 for external annotation services. The higher the match, the better the decision performance because the training data contains less noise.
Lesson 2: Break down vague criteria into concrete judges. Rather than a single judge evaluating whether a response is appropriate; "Be relevant, factual, and concise;" Create three separate judges. Each targets a specific quality aspect. This granularity is important. "overall quality" When I look at the score, I know something is wrong, but I don’t know what to fix.
The best results are achieved by combining top-down requirements such as regulatory constraints and stakeholder priorities with bottom-up discovery of observed failure patterns. One customer built a top-down correctness determination system and discovered through data analysis that the correct response almost always cited the top two search results. This insight has become a new production-friendly judge that can proxy for correctness without the need for ground truth labels.
Lesson 3: You need fewer examples than you think. Teams can create a strong jury from just 20 to 30 carefully selected examples. The key is to choose edge cases that highlight inconsistencies, rather than obvious examples that everyone agrees on.
"Some teams can do this process in as little as 3 hours, so it doesn’t take long to get a good judge." Coppol said.
Production track record: from pilot to 7-digit deployment
Frankle shared three metrics Databricks uses to measure the success of Judge Builder: whether customers want to use it again, whether they increase their AI spending, and whether their AI efforts progress further.
In the first metric, one customer created a panel of more than a dozen judges after the first workshop. "This customer built a jury of over 10 people after showing them how to do this in a rigorous way for the first time in this framework." Frankl said. "They actually went to town as judges and are measuring everything now."
For the second metric, the business impact is clear. "We have multiple customers who have taken this workshop and are now spending seven figures like never before with Databricks’ GenAI." Frankl said.
The third metric reveals the strategic value of Judge Builder. Customers who were previously hesitant to use advanced techniques such as reinforcement learning can now do so with confidence because they can measure whether improvements actually occur.
"Some clients have done very sophisticated things after previously resisting these judges." Frankl said. "They went from doing a little prompt engineering to doing reinforcement learning with us. Why should we spend money on reinforcement learning? And why should we spend energy on reinforcement learning when we don’t even know if it actually makes a difference?"
What companies should do now
Teams that successfully move AI from pilot to production treat judges not as one-time artifacts, but as evolving assets that grow with the system.
Databricks recommends three practical steps. First, we focus on high-impact judges by identifying one important regulatory requirement and one observed failure mode. These will be your first judge portfolio.
Second, create lightweight workflows with subject matter experts. A few hours of reviewing 20 to 30 edge cases is sufficient adjustment for most judges. Denoise your data using batched annotations and interrater reliability checks.
Third, schedule regular auditor reviews using production data. As systems evolve, new failure modes emerge. Your judge’s portfolio should evolve with them.
"Judges are a way to evaluate models, they’re a way to create guardrails, they’re a way to get metrics that allow you to do quick optimizations, they’re a way to get metrics that allow you to do reinforcement learning." Frankl said. "Once you have a judge that you know represents your human preferences in empirical form, and you can query it as much as you like, you can use it in 10,000 different ways to evaluate and improve your agent."
