Guides15 min read

How to Run LLM Evals: A Step-by-Step Guide for Product Managers

Learn how to design, run, and interpret LLM evaluations as a product manager. Covers eval frameworks, metric selection, dataset creation, and using results to ship better AI features.

By Tim Adair• Published 2026-02-09

Quick Answer (TL;DR)

LLM evals are systematic tests that measure how well your AI feature performs against defined quality criteria. As a PM, you do not need to write the evaluation code yourself, but you must own the eval strategy: deciding what to measure, building the test dataset, setting pass/fail thresholds, and interpreting the results to make ship/no-ship decisions. Think of evals as your AI product's test suite. Without them, you are shipping blind.

Summary: LLM evals are structured tests that let you measure AI output quality against defined criteria so you can make confident ship decisions.

Key Steps:

  • Define what "good" looks like for your AI feature with concrete scoring rubrics
  • Build a representative evaluation dataset of 50-200 test cases covering edge cases and core scenarios
  • Run evals on every model change, prompt change, or pipeline update and track results over time
  • Time Required: 2-4 days to set up your first eval suite; 1-2 hours per eval run thereafter

    Best For: PMs building or maintaining any product feature powered by an LLM


    Table of Contents

  • What Are LLM Evals and Why PMs Should Care
  • The PM's Role in Evaluations
  • Choosing What to Evaluate
  • Building Your Evaluation Dataset
  • Selecting Eval Metrics
  • Setting Up Your First Eval
  • Running Evals in Practice
  • Interpreting Results and Making Decisions
  • Automating Evals in Your Pipeline
  • Common Eval Mistakes
  • Getting Started Checklist
  • Key Takeaways

  • What Are LLM Evals and Why PMs Should Care

    An LLM eval is a structured test that sends a set of inputs to your AI system and scores the outputs against predefined criteria. It is the AI equivalent of a unit test suite, but instead of checking whether code returns the right integer, you are checking whether the model produces responses that are accurate, helpful, safe, and on-brand.

    Without evals, you have no reliable way to answer critical product questions:

  • Did the new prompt actually improve response quality, or did it just feel better in the three examples you tried?
  • Is the model performing equally well for all user segments, or does it degrade for certain input types?
  • When you upgrade from one model version to another, what breaks?
  • Evals replace gut feelings with data. They are the foundation of responsible AI product management and the mechanism that allows you to iterate on AI features with confidence.

    Why This Is a PM Responsibility

    Engineering owns the eval infrastructure. Data science may help design the scoring functions. But the PM must own the eval strategy because only the PM understands what "good" means from the user's perspective. A technically excellent response that misses the user's actual need is a product failure, and only the PM can define that boundary.


    The PM's Role in Evaluations

    Your job in the eval process is not to write Python scripts. It is to make four critical decisions:

    1. Define Quality Dimensions

    What attributes make a response "good" for your specific feature? Common dimensions include:

  • Accuracy: Is the information factually correct?
  • Relevance: Does the response actually address the user's question?
  • Completeness: Does it cover the key points without critical omissions?
  • Tone and style: Does it match your product's voice?
  • Safety: Does it avoid harmful, biased, or inappropriate content?
  • Conciseness: Is it the right length for the use case?
  • You will not evaluate all of these for every feature. A customer support chatbot might prioritize accuracy and tone. A code generation feature might prioritize correctness and completeness. Pick the 3-4 dimensions that matter most for your feature.

    2. Create the Gold Standard

    You decide what a great response looks like. This means writing or curating reference answers that represent the quality bar you want the model to hit. These reference answers become the benchmark for automated scoring.

    3. Set Thresholds

    You decide the pass/fail criteria. "We need 90% of responses to score 4 or higher on accuracy" is a product decision, not a technical one. Setting these thresholds requires balancing quality, cost, latency, and user expectations.

    4. Make Ship Decisions

    Eval results tell you whether a change improved or degraded quality. You decide whether the trade-offs are acceptable and whether the feature is ready to ship.


    Choosing What to Evaluate

    Not every aspect of your AI feature needs formal evaluation. Focus your eval efforts on the areas with the highest risk and the highest user impact.

    High-Priority Eval Targets

    Core task performance: Can the model do the main thing users expect? If your feature summarizes documents, evaluate summarization quality. If it answers questions, evaluate answer accuracy.

    Edge cases and failure modes: What happens with unusual inputs? Empty inputs, extremely long inputs, adversarial inputs, inputs in unexpected languages, ambiguous queries. These are where models fail most visibly.

    Safety and compliance: Does the model ever produce harmful, biased, or legally problematic outputs? This is non-negotiable for any user-facing feature.

    Consistency: Does the model give substantially different answers to the same question on repeated runs? High variance is a product quality issue even if the average quality is acceptable.

    Lower-Priority (But Still Important)

    Latency and cost: Track these as operational metrics alongside quality evals. A response that takes 30 seconds is a product failure regardless of how accurate it is.

    Format compliance: Does the model follow the output format you specified? JSON when you asked for JSON, bullet points when you asked for bullet points.


    Building Your Evaluation Dataset

    Your eval dataset is the single most important artifact in your eval process. A bad dataset produces misleading results that lead to bad ship decisions.

    Dataset Composition

    A strong eval dataset has three tiers:

    Tier 1: Core scenarios (60% of dataset) represent the most common user inputs. Pull these from production logs if you have them, or generate representative examples based on user research. These cases should be straightforward and represent the bread-and-butter use case.

    Tier 2: Edge cases (25% of dataset) represent tricky or unusual inputs: very short queries, very long queries, ambiguous requests, queries with typos, queries that require the model to say "I don't know." Edge cases are where quality differences between model versions become most visible.

    Tier 3: Adversarial cases (15% of dataset) represent inputs designed to break the model: prompt injection attempts, requests for harmful content, attempts to get the model to contradict its instructions, inputs that try to extract system prompts. These test your safety boundaries.

    Dataset Size

    For most product features, you want:

  • Minimum viable dataset: 50 test cases (enough to spot major regressions)
  • Recommended dataset: 100-200 test cases (enough for statistically meaningful comparisons)
  • Comprehensive dataset: 500+ test cases (for high-stakes features like medical or financial applications)
  • Start with 50 and grow over time. Every production bug or user complaint should generate a new eval case.

    Creating Reference Answers

    For each test case, you need a reference answer or scoring rubric. There are two approaches:

    Reference-based evaluation: You write the ideal answer and score model outputs by comparing them to the reference. Best for factual tasks where there is a clearly correct answer.

    Rubric-based evaluation: You write scoring criteria and either human raters or an LLM judge scores outputs against the rubric. Best for creative or subjective tasks where multiple good answers exist.


    Selecting Eval Metrics

    Automated Metrics

    These can be computed programmatically without human involvement:

    Exact match: Does the output exactly match the expected answer? Useful for classification tasks and structured outputs.

    String similarity (BLEU, ROUGE): How much word overlap exists between the model output and the reference answer? Useful as a rough quality signal for summarization and generation tasks, but can be misleading because two very different sentences can convey the same meaning.

    LLM-as-judge: Use a separate, stronger LLM to grade outputs against a rubric. This is increasingly the standard approach for evaluating open-ended generation. The judge model scores each output on your defined dimensions and provides reasoning.

    Regex and format checks: Does the output conform to the required format? Valid JSON, correct number of bullet points, within length limits.

    Factual accuracy (retrieval-based): For RAG systems, check whether the model's claims are supported by the retrieved source documents.

    Human Metrics

    Automated metrics cannot fully capture quality for subjective or nuanced outputs. Include human evaluation for:

  • Initial baseline: Have 3-5 people rate 50 outputs to calibrate your automated metrics
  • Periodic audits: Sample 20-30 outputs weekly for human review to catch drift
  • Disagreement cases: When automated metrics are uncertain, route to human reviewers
  • The Metric Stack

    For most AI features, use this combination:

  • Automated format checks (fast, catches structural failures)
  • LLM-as-judge scoring on your 3-4 quality dimensions (scalable, catches quality issues)
  • Weekly human audit of a random sample (catches things LLM judges miss)

  • Setting Up Your First Eval

    Step 1: Write Your Eval Spec

    Before anyone writes code, document your eval strategy in a one-page spec covering the feature description, primary quality dimensions, dataset size target, scoring method, pass threshold, eval frequency, and ownership split between PM and engineering.

    Step 2: Curate Your Initial Dataset

    Spend one focused day building your first 50 test cases. Pull from real user queries from production logs or customer support tickets, edge cases identified during product design, adversarial cases from your security or trust team, and scenarios from user research and customer interviews.

    Step 3: Define Your Scoring Rubric

    Write a clear rubric for each quality dimension. Be specific enough that two different raters would give similar scores. A bad rubric says "Rate accuracy from 1-5." A good rubric says "Accuracy: 5 = all claims factually correct and verifiable. 4 = all major claims correct, minor details may be imprecise. 3 = mostly correct but contains one significant error. 2 = contains multiple errors that could mislead users. 1 = fundamentally incorrect or fabricated information."

    Step 4: Run a Calibration Round

    Before automating, do a manual eval run. Have 2-3 team members score 20 outputs independently. Compare scores. Where scores diverge, discuss and refine the rubric. This calibration step prevents weeks of wasted effort from a poorly defined rubric.

    Step 5: Automate

    Work with your engineering team to set up the eval pipeline: a script that runs all test cases through your AI feature, an LLM-as-judge or other automated scorer that grades each output, a results dashboard that shows scores broken down by dimension and test case category, and alerting when scores drop below your threshold.


    Running Evals in Practice

    When to Run Evals

    Always run evals when:

  • You change the system prompt or any prompt template
  • You switch model versions (e.g., GPT-4 to GPT-4o, Claude 3.5 to Claude 4)
  • You change the retrieval pipeline, context window, or tool configuration
  • You modify post-processing or output formatting logic
  • Run evals on a schedule when:

  • Model providers push updates to hosted models (they do this silently)
  • Your knowledge base or retrieval corpus changes
  • Weekly, as a baseline health check
  • Reading Eval Results

    An eval run produces a results matrix showing the overall pass rate against your target, pass rates broken down by quality dimension (accuracy, relevance, tone, completeness), pass rates by test category (core scenarios, edge cases, adversarial), and regression data compared to the previous version. The PM's job is to assess whether dimension-level trade-offs are acceptable for users.

    Debugging Failures

    When eval scores drop, follow this process:

  • Look at the failing cases: Read the actual inputs and outputs. Patterns usually emerge quickly.
  • Categorize failures: Are failures concentrated in a specific input type, topic area, or output format?
  • Trace the root cause: Is the model producing bad content, or is the prompt not giving it enough guidance? Is the retrieval pipeline returning irrelevant context?
  • Fix and re-eval: Make targeted changes and run the eval again. Compare to the previous run.

  • Interpreting Results and Making Decisions

    The Ship Decision Framework

    Use this framework to translate eval results into product decisions:

    Green (ship it): All dimensions meet or exceed thresholds. No regressions from the previous version. Adversarial test pass rate is above your safety floor.

    Yellow (ship with monitoring): Core scenarios pass. Some edge case degradation. You have monitoring in place to catch production issues. The improvement in primary dimensions outweighs minor edge case regressions.

    Red (do not ship): Any safety or adversarial test failure. Core scenario pass rate below threshold. Significant regression in any primary dimension without a corresponding major gain elsewhere.

    Communicating Results to Stakeholders

    Executives and cross-functional partners do not want to see raw eval scores. Translate your results into business language:

  • "Our new model version answers customer questions correctly 95% of the time, up from 91%. We are confident this will reduce support ticket escalations."
  • "We found that the model struggles with pricing questions at 72% accuracy. We are holding the launch until we fix this, because incorrect pricing responses could erode customer trust."
  • "The eval suite caught a regression that would have affected 15% of users. We fixed it before launch."

  • Automating Evals in Your Pipeline

    The Eval-Driven Development Workflow

    Integrate evals into your development workflow the same way you integrate unit tests:

  • Pre-merge: Run the eval suite on every pull request that touches prompts, model config, or the AI pipeline. Block merging if scores drop below threshold.
  • Post-deploy: Run the full eval suite after every deployment to catch environment-specific issues.
  • Scheduled: Run weekly eval sweeps to detect model drift from provider-side updates.
  • Monitoring Production Quality

    Evals on test data are necessary but not sufficient. You also need production quality monitoring:

  • Sample and score: Randomly sample 1-5% of production requests and run them through your eval scoring pipeline. This catches distribution shifts where real user inputs diverge from your test dataset.
  • User feedback signals: Track thumbs up/down, regeneration rates, and support ticket mentions of the AI feature. Correlate these with eval scores.
  • Regression alerts: Set up alerts when production quality scores drop below thresholds or when user feedback signals shift negatively.

  • Common Eval Mistakes

    Mistake 1: Evaluating on too few examples

    Instead: Start with at least 50 test cases and grow to 200 over the first month.

    Why: With fewer than 50 cases, a single flaky test result can swing your pass rate by 2% or more, making it impossible to distinguish real quality changes from noise.

    Mistake 2: Not including adversarial cases

    Instead: Dedicate at least 15% of your eval dataset to adversarial and safety-related inputs.

    Why: Models can score 95% on happy-path cases while being trivially jailbroken. If you only test the happy path, you will miss your biggest risks.

    Mistake 3: Using the eval dataset to tune prompts

    Instead: Maintain a separate development set for prompt iteration. Only run the eval dataset for final scoring.

    Why: If you optimize your prompts against your eval dataset, your scores will look great but will not generalize to real users. This is the AI equivalent of teaching to the test.

    Mistake 4: Treating eval as a one-time setup

    Instead: Add new test cases monthly from production failures, user complaints, and edge cases discovered in the wild.

    Why: Your eval dataset should evolve with your product. The inputs users send in month six will be different from month one. A stale eval dataset gives you false confidence.

    Mistake 5: Not calibrating your LLM judge

    Instead: Validate your LLM judge against human ratings on at least 50 examples before trusting it.

    Why: LLM judges have their own biases. They tend to be generous raters and can miss subtle quality issues that humans catch. Calibration ensures your automated scores are meaningful.


    Getting Started Checklist

    Day 1: Strategy

  • Identify the AI feature you will evaluate first (pick your highest-risk feature)
  • Define 3-4 quality dimensions that matter most for this feature
  • Write your one-page eval spec
  • Align with your engineering lead on who owns what
  • Day 2: Dataset

  • Pull 30 representative inputs from production logs or user research
  • Write 10 edge case inputs based on known product limitations
  • Write 10 adversarial inputs (prompt injections, harmful requests, boundary-pushing queries)
  • Create reference answers or scoring rubrics for each test case
  • Day 3: Calibration

  • Run 20 test cases through the AI feature manually
  • Have 2-3 team members independently score the outputs
  • Compare scores and refine your rubric where raters disagreed
  • Document the final rubric with examples of each score level
  • Day 4: Automation

  • Work with engineering to set up the eval pipeline (test runner, scorer, results dashboard)
  • Run the full eval suite and review initial baseline scores
  • Set your pass/fail thresholds based on the baseline
  • Integrate eval runs into your PR and deployment workflow
  • Ongoing

  • Add 5-10 new test cases per week from production feedback
  • Run weekly scheduled evals to detect model drift
  • Conduct monthly human calibration audits (20-30 examples)
  • Review and update thresholds quarterly as your quality bar evolves

  • Key Takeaways

  • LLM evals are your AI feature's test suite. Without them, you are shipping blind and hoping for the best.
  • The PM owns the eval strategy: what to measure, how to score it, and what thresholds to set. Engineering owns the infrastructure.
  • Build a diverse eval dataset with core scenarios (60%), edge cases (25%), and adversarial cases (15%). Start with 50 and grow to 200.
  • Use LLM-as-judge scoring for scalable automated evaluation, calibrated against human ratings.
  • Run evals on every prompt change, model change, and pipeline change. Automate them into your CI/CD workflow.
  • Translate eval results into ship/no-ship decisions using clear green/yellow/red criteria.
  • Next Steps:

  • Pick the AI feature with the highest user-facing risk and write an eval spec this week
  • Build your first 50-case eval dataset by pulling from production logs and user research
  • Run a manual calibration round with your team before automating anything

  • Prompt Engineering for Product Managers
  • Specifying AI Agent Behaviors
  • Red Teaming AI Products
  • AI Product Monitoring and Observability

  • About This Guide

    Last Updated: February 9, 2026

    Reading Time: 15 minutes

    Expertise Level: Intermediate

    Citation: Adair, Tim. "How to Run LLM Evals: A Step-by-Step Guide for Product Managers." IdeaPlan, 2026. https://ideaplan.io/guides/how-to-run-llm-evals

    Free Resource

    Want More Guides Like This?

    Subscribe to get product management guides, templates, and expert strategies delivered to your inbox.

    No spam. Unsubscribe anytime.

    Want instant access to all 50+ premium templates?

    Put This Guide Into Practice

    Use our templates and frameworks to apply these concepts to your product.