Enterprise AI Engineering · AI Evaluation & Assurance

    Know whether your AI actually works.

    Evaluation and assurance for GenAI, RAG, machine learning and AI agents: golden datasets, task-success metrics, factuality and groundedness checks, tool-call correctness, prompt regression, safety testing, red teaming and production monitoring.

    Our position

    AI systems need acceptance criteria just like software systems.

    AI output is probabilistic. The same input can produce a different answer, and a correct answer on Monday can be wrong after a model update on Friday. That does not make AI untestable. It means testing needs a different apparatus from unit tests.

    Testing a probabilistic system requires all seven of these, not just unit tests:

    • Datasets

      Realistic inputs with known good outcomes, covering the common cases and the rare ones that matter.

    • Rubrics

      Written criteria for what a good output is, so two reviewers and an automated judge score the same way.

    • Business criteria

      The thresholds the sponsor will accept, agreed before the build rather than negotiated after the demo.

    • Human baselines

      How well people do the task today, so the system is compared with reality rather than perfection.

    • Automated evaluation

      Scoring that runs on every change, fast enough to be a release gate.

    • Production feedback

      Sampled real traffic, user corrections and escalations flowing back into the datasets.

    • Continuous evaluation

      The same checks running after launch, because the data, the model provider and the users all change.

    In brief

    AI Evaluation & Assurance is the discipline of proving that a Generative AI, RAG, machine learning or agent system meets agreed acceptance criteria before release and keeps meeting them in production. It combines golden datasets, task-success and factuality metrics, retrieval and tool-call checks, human review, safety testing and red teaming with continuous production monitoring. It is needed whenever an AI system's output is probabilistic, which is every AI system. It solves the question every sponsor eventually asks: does this actually work?

    How evaluation runs

    From test cases to a release decision to production monitoring.

    Six stages. The acceptance stage is a decision against criteria agreed in advance, and the monitoring stage runs the same evaluation on real traffic so the decision stays true.

    1. 01

      Test cases

      Golden datasets built from real inputs, edge cases and adversarial prompts.

    2. 02

      Model or agent

      The candidate release, with every artefact versioned.

    3. 03

      Automated evaluation

      Metrics, rubric-based judges and tool-call checks, run on every change.

    4. 04

      Human evaluation

      Calibrated reviewers on the samples that automation cannot judge reliably.

    5. 05

      Acceptance

      A release decision against the criteria written down in advance.

    6. 06

      Production monitoring

      Sampled traffic scored the same way, with drift and regression alerts.

    Capabilities

    What we evaluate, and how.

    • What we evaluate

      • LLM evaluation
      • RAG evaluation
      • Agent evaluation
      • Retrieval evaluation
    • How we evaluate

      • Golden datasets
      • Task-success evaluation
      • Factuality
      • Groundedness
      • Tool-call correctness
      • Prompt regression
    • Risk and safety

      • Safety testing
      • Red teaming
      • Prompt-injection testing
      • Data-boundary checks
    • Production

      • Latency
      • Cost
      • Human escalation
      • Production monitoring

    What to measure depends on the system.

    A single accuracy number hides more than it shows. These are the properties we score for each kind of system, chosen so that a regression points at the component that caused it.

    Evaluation properties by system type
    SystemWhat we measureHow it is done
    RAG and knowledge assistantsRetrieval precision and recall, groundedness, answer correctness, citation accuracy, refusal on unanswerable questionsRetrieval is scored separately from generation so a regression is attributed to the right component
    AI agentsTask success, tool-call correctness, steps to completion, unsafe-action rate, escalation appropriatenessEvaluated in a simulated environment with realistic tool responses, including failures
    Copilots and draftingAcceptance rate, edit distance, time to complete the task against a human baselineMeasured with real users where possible; benchmark scores are an input, not the result
    Classification and predictionPrecision, recall, F1, ROC/AUC, MAE or RMSE by segment, calibrationMetric chosen from the cost of each kind of error, then tracked for drift
    Every systemLatency at the required percentile, cost per request or task, escalation rateQuality that is unaffordable or too slow does not ship

    Answer first

    The evaluation questions every AI sponsor should ask.

    How do you test a Generative AI system?

    With a golden dataset of realistic inputs and expected outputs, written acceptance criteria, automated scoring for the properties that can be scored, calibrated human review for the ones that cannot, safety and adversarial testing, and the same checks running on sampled production traffic after release.

    Unit tests still apply to the software around the model. They cannot tell you whether the model's answer was right, which is what the rest of this page is for.

    How do you evaluate RAG?

    In two stages. First retrieval: for each question in the evaluation set, were the right passages retrieved, scored for precision, recall and rank. Then generation: given the retrieved context, was the answer correct, grounded in that context and cited. Scoring them separately is what lets you fix the right component.

    How do you measure hallucination?

    By decomposing each answer into claims and checking whether every claim is supported by the retrieved context (groundedness) and, where a reference exists, whether it is true (factuality). The result is a rate per answer and per dataset, judged automatically and verified by human sampling. We do not promise a fixed hallucination percentage; we measure it for the specific use case.

    How do you evaluate an AI agent?

    With a set of tasks that have known correct outcomes, run in a simulated environment that returns realistic tool responses including failures, scoring task success, tool-call correctness, steps taken, unsafe actions and whether the agent escalated when it should have. Production traces are then sampled and scored the same way.

    What is an AI golden dataset?

    A curated set of inputs with agreed correct outputs or rubric scores, drawn from real usage and deliberately extended with edge cases and adversarial examples. It is the fixed reference every release is measured against, and it grows from production feedback.

    Building it is a joint effort with the business, because deciding what a correct output is requires the people who do the job.

    Can LLM-as-a-judge be trusted?

    Partly, and only when calibrated. A model can score outputs against a rubric at scale, but its agreement with human reviewers has to be measured on the same samples, its known biases (toward length, toward its own style) controlled, and its judgments spot-checked continuously. Used that way it is a useful instrument; used blindly it is a second opinion from the system under test.

    What should be human evaluated?

    Anything where correctness needs domain judgment, where the cost of a wrong output is high, where the rubric is new or contested, and a rolling sample of everything else to keep the automated judge honest. Human evaluation is expensive, so it is aimed where automation is least reliable.

    What should be automated?

    Everything with a checkable property: retrieval metrics, schema and format compliance, tool-call correctness, safety classifiers, regression against the golden dataset, latency and cost. Automation is what makes evaluation a release gate rather than a quarterly review.

    How do you test AI before production?

    By writing the acceptance criteria first, building the golden dataset with the business, running automated and human evaluation on every candidate release, red teaming the system for unsafe outputs and actions, and rehearsing the failure and rollback paths. A system that passes ships with monitoring that keeps running the same tests.

    How the operating layer runs these checks after launch

    Assurance in context

    Evaluation is the evidence that governance runs on.

    Risk classification, oversight decisions and audit requests all end in the same question: what does the evidence say about this system? Evaluation produces that evidence continuously. We structure it so a governance function can use it directly.

    The NIST AI Risk Management Framework organises AI risk work into four functions: Govern, Map, Measure and Manage, with measurement treated as an ongoing activity rather than a one-off test.
    Source: National Institute of Standards and Technology, AI Risk Management Framework (AI RMF 1.0) (2023). A voluntary framework. NeoIntelli uses its structure as a reference when shaping evaluation and monitoring plans; it does not certify compliance with it.

    Where the governance framework lives

    Policy, risk classification and oversight for an AI capability center are covered under Responsible AI governance in AI GCC. This page covers how evaluation produces the evidence that framework needs, for a specific system. We do not claim regulatory compliance on a client's behalf.

    Metrics we measure

    What an evaluation programme reports.

    • Evaluation pass rate

      Share of release candidates meeting acceptance criteria on the golden dataset.

    • Task success

      Completion to the agreed outcome, on the evaluation set and on sampled production traffic.

    • Groundedness and factuality

      Share of claims supported by sources, and share of factual statements verified as correct.

    • Retrieval precision and recall

      Whether the right passages are retrieved for the questions the workflow asks.

    • Tool-call correctness

      Right tool, right arguments, right sequence, measured per step.

    • Judge agreement with humans

      How closely automated scoring tracks calibrated human review on the same samples.

    • Unsafe-output and unsafe-action rate

      Findings from safety testing and red teaming, before and after mitigation.

    • Regression rate

      Share of changes that lowered a tracked metric and were caught before release.

    • Production drift

      Movement in quality signals on sampled traffic against the release baseline.

    These are the metrics we instrument and report. Targets are agreed per system against a measured baseline; we do not publish universal improvement percentages.

    Next step

    Find out whether your AI actually works, with evidence.

    Bring the system, the questions the business keeps asking about it, and whatever testing exists today. We will design the acceptance criteria, the golden dataset and the evaluation pipeline, and tell you what the numbers currently say.

    A 30-minute working session with a senior AI engineer. Bring the workflow, the data it depends on and the constraint that has stopped it so far.