# 2025 is the Year of Evals! Just like 2024, and 2023, and …

John Dickerson, Mozilla AI | AI Engineer World's Fair 2025 | 19:14

Source: https://www.youtube.com/watch?v=CQGuvf6gSrM
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/2025-is-the-year-of-evals-just-like-2024-and-2023-and
Published: 2025-08-06
Tags: enterprise, evals, observability

## TL;DR
- Evaluation becomes a C-suite concern when AI systems take actions for people and teams, rather than only producing inputs for another system.
- Enterprise AI spending moved from small GenAI experiments in 2023, to production applications in 2024, to scaled and increasingly autonomous systems in 2025.
- Useful evaluations need to connect model and agent behavior to business KPIs, while high-risk domains may still require experts to validate system outputs.

## Summary
John Dickerson traces enterprise AI evaluation through the past decade, drawing on his experience at Arthur AI and his role as CEO of Mozilla AI. He argues that monitoring and evaluation depend on the same underlying ability to measure system behavior. Evaluation was often difficult to sell because machine learning outputs disappeared inside larger, opaque systems and had little visible connection to business results. ChatGPT changed the buying conversation by making AI understandable to senior executives and unlocking discretionary budgets. Enterprise projects then moved from GenAI experiments in 2023, to production applications in 2024, and to scaled, more autonomous systems in 2025. Agents add complexity and risk because they perceive, reason, and act. Dickerson says evaluations now need to produce numbers tied to revenue, cost, risk, or other business outcomes. He also discusses expert validation for high-stakes work and the limits of using LLMs as judges.

## Key ideas
### Monitoring and evaluation depend on the same ability to measure behavior
[01:05](https://www.youtube.com/watch?v=CQGuvf6gSrM&t=65s)
Dickerson describes AI and machine learning monitoring as two sides of the same ruler. Monitoring or observability requires measurement, and measurement is the core function of evaluation. Before ChatGPT, this need was often hidden inside larger systems. A model might produce numbers that another system consumed, so decision-makers saw an opaque product rather than a model whose behavior needed direct evaluation. That made evaluation less visible outside the CIO or CTO's organization. Dickerson's point is practical: an evaluation program cannot be separated from the measurements used to monitor a deployed system.

### ChatGPT made AI understandable to executives and unlocked focused budgets
[01:20](https://www.youtube.com/watch?v=CQGuvf6gSrM&t=80s)
Two events changed the buying environment at the same time. ChatGPT made AI something CEOs, CFOs, and CISOs could understand through direct use. A budget freeze, driven by fear of a recession, also pushed enterprises to reserve spending for specific pet projects. GenAI became that pet project because executives could interact with it through a simple web interface and see an immediate result. Dickerson says the resulting discretionary budgets went toward GenAI even while wider IT spending was frozen or reduced. This explains why GenAI received attention before evaluation did.

### Enterprise AI moved from experiments to production and then to scale
[09:15](https://www.youtube.com/watch?v=CQGuvf6gSrM&t=555s)
Dickerson divides the recent enterprise cycle into three stages. In 2023, limited available budget went toward GenAI science projects. In 2024, applications such as internal chat tools and hiring tools started going into production. Once those systems were in use, business leaders began asking about return on investment, governance, risk, compliance, and brand optics. In 2025, those applications are shipping and scaling, with larger budgets and greater usage. Dickerson connects this spending to executive comfort with AI and to improvements in the technology. He says this progression brings evaluation closer to the people who control enterprise budgets.

### Agents make evaluation more visible because they take actions
[11:18](https://www.youtube.com/watch?v=CQGuvf6gSrM&t=678s)
Dickerson defines agents as systems that perceive their environment, learn, abstract and generalize, reason, and act. Reasoning models and systems that operate in virtual or cyber-physical environments add complexity and risk compared with traditional machine learning models that mainly output numbers for another system. Agents can make decisions and take complex steps autonomously or semi-autonomously. Humans will remain in the loop for many systems, but agentic systems are already moving into enterprise, small-business, and experimental deployments. Once a system acts for a person or team, its behavior becomes much easier for senior leaders to connect with operational and business risk.

### An evaluation must connect system behavior to a business KPI
[12:11](https://www.youtube.com/watch?v=CQGuvf6gSrM&t=731s)
For an enterprise product to be bought, Dickerson says its value must connect to a downstream business KPI. Possible measures include risk mitigation, revenue gains, or losing less money. Evaluation supplies the quantitative estimates needed for those conversations. He describes the CEO as someone who now understands enough about generative and agentic systems to discuss them and approve budgets. The CFO needs numbers for allocation and planning. The CISO sees security risks such as hallucination and prompt injection, while the CIO and CTO need standards and operational evidence. Evaluation becomes a first-class discussion because different executives can use its measurements for their own decisions.

### Monitoring has expanded from one model to the whole multi-agent system
[14:51](https://www.youtube.com/watch?v=CQGuvf6gSrM&t=891s)
Dickerson says evaluation, observability, monitoring, and security companies have shifted toward monitoring multi-agent systems. The object of monitoring should be the whole system, rather than only the model used by one agent. This view is increasingly accepted in industry and government. He also argues that revenue in the evaluation market has been changing quickly. Publicly reported startup numbers can lag the current situation by six or eight months, so older figures may not describe what companies are making now. His forecast is tied to the growth of deployed and autonomous systems, which create a larger need to understand behavior across the complete system.

### High-stakes evaluations still need domain experts and carefully built environments
[16:18](https://www.youtube.com/watch?v=CQGuvf6gSrM&t=978s)
For a financial multi-agent system doing discounted cash flow analysis, general machine learning expertise is not enough to determine whether the work is correct. Dickerson expects companies to hire specialists who sit alongside the system and validate its work. He cites expert rates of $50, $100, or $200 per hour from a leaked spreadsheet and says this cost can make sense when an error could lose substantial money or someone's job. He also says dataset and environment creation matter more than anything in the evaluation space. A well-built competitive DCF environment can become a company asset, although future systems may absorb the validation data.

### LLMs can judge outputs, but their judgments need human validation
[18:08](https://www.youtube.com/watch?v=CQGuvf6gSrM&t=1088s)
Dickerson says the LLM-as-a-judge approach is already used in practice because it can reduce the work of creating evaluation datasets. Giving an LLM a persona produces a low-cost approximation of human judging. The method has known problems, however. He refers to research on biases relative to human judgments, including preferences related to conciseness and helpfulness. Teams therefore need to validate the judge and check that it has not drifted toward an unexpected bias. LLM judges can help with scale, but they do not remove the need to establish trustworthy evaluation data and test the judging process itself.

## Notable quotes
- "You can't do monitoring or observability without being able to measure, and measurement is the core functionality for evaluation." (01:05)
- "We have systems that are now acting for humans, acting for teams, as opposed to just providing inputs into larger systems." (02:24)
- "At the end of the day, the thing that really matters, like I mentioned, is connecting when you're selling any product into an enterprise or SMB, is being able to attach your product into some sort of downstream business KPI." (12:11)
- "When you talk to anyone in the eval space, it's the data set creation and the environment creation that matters more than anything." (17:52)
- "You do need to make sure you're validating this and making sure that you're not going off in some weird bias direction." (18:58)

## Tools & references mentioned
- Arthur AI
- Mozilla AI
- Braintrust
- Arize AI
- Galileo
- OpenAI
- ChatGPT
- H2O
- Algorithmia
- Seldon
- WhyLabs
- Aporia
- Snowflake
- Databricks
- Datadog
- Amazon SageMaker
- Vertex AI
- Microsoft
- JPMorgan Chase
- Jamie Dimon
- Nature Machine Intelligence
- Mercor
- Google
- Meta
- ICLR
- Anthropic
- Any-agent

## Who should watch
- You are building evaluation or observability tools and need to explain their value to executives who control budgets.
- Your GenAI application has moved from an internal experiment into production, and you need measurements for risk, governance, or return on investment.
- You are evaluating an agent in a high-stakes domain and need to decide where expert review or an LLM judge can be trusted.

## Related talks

- [The 2025 AI Engineering Report](https://aietalks.com/talks/the-2025-ai-engineering-report) (Barr Yaron, Amplify Partners, 12:33)
- [Evals](https://aietalks.com/talks/evals) (Omar Khattab & Vitor Balogh & Rafal Wilinski, Zapier & Ido Pesal, Vercel & Randall, Bolt Foundry & Diego Rodriguez, Krea & Doug Guthrie & Ankur Goyal, Braintrust & Jeff Huber, Chroma & Jason Liu, 567 Studio & John Dickerson, Mozilla AI, 3:59:12)
- [How to Run Evals at Scale: Thinking Beyond Accuracy or Similarity](https://aietalks.com/talks/how-to-run-evals-at-scale-thinking-beyond-accuracy-or-similarity) (Muktesh Mishra, Adobe, 09:25)
- [Engineering Better Evals: Scalable LLM Evaluation Pipelines That Work](https://aietalks.com/talks/engineering-better-evals-scalable-llm-evaluation-pipelines-that-work) (Dat Ngo & Aman Khan, Arize, 24:46)
- [Why should anyone care about Evals?](https://aietalks.com/talks/why-should-anyone-care-about-evals) (Manu Goyal, Braintrust, 05:41)