The Written Accuracy Bar: Defining "Good Enough" Before You Build AI

An AI system is ready for production when it meets a standard that was written down before anyone saw the results. The Written Accuracy Bar is that standard: the metric, the threshold, the sample, the confidence level, the severity rules, and the people who sign. This page explains how to set each element and provides a complete template.

Why the bar must be written first

When the threshold is set after the results arrive, it tends to settle wherever the results happen to be. A bar written in advance protects the business owner, the delivery team, and the people affected by the system from that drift. It also turns evaluation from opinion into evidence. OpenAI's evaluation guidance names "vibe-based evals," using "it seems like it's working" as a strategy, as an anti-pattern, and recommends task-specific evals that reflect real-world distributions.¹ Anthropic's guidance asks for success criteria that are specific, measurable, achievable, and relevant, and offers a model of a measurable criterion: "less than 0.1% of outputs out of 10,000 trials flagged for toxicity."²

The stakes are practical. In McKinsey's 2026 State of AI survey, inaccuracy was the most commonly experienced AI-related risk among respondent organizations.³ Regulation points the same way: for high-risk systems, the EU AI Act requires an "appropriate level of accuracy" and requires that "the levels of accuracy and the relevant accuracy metrics" be declared in the instructions for use.⁴ The NIST AI Risk Management Framework places measurement at the center of its Measure function, and its Generative AI Profile (NIST AI 600-1) identifies confabulation, the confident generation of false content, as a risk specific to generative systems.⁵⁶

Step 1: Choose metrics that fit the task

Different tasks fail in different ways, so they need different metrics. Choose one primary metric that decides pass or fail, and a small number of secondary metrics that you monitor.

Prefer automated grading where it is reliable. Anthropic's guidance ranks code-based grading as fastest and most reliable, model-based grading as flexible and scalable, and human grading as the highest quality but slowest.² When a model is used as a grader, calibrate it against human judgments before relying on it; OpenAI lists failing to do so as an anti-pattern.¹

Step 2: Classify errors by severity

A single accuracy number treats every error as equal. In practice, a misspelled supplier name and a wrong bank account number are not the same event. Define severity tiers and set a separate threshold for each.

Build the severity list from the specific harms of the workflow. For generative and agentic systems, the OWASP Top 10 for LLM Applications (2025) is a useful prompt for security-related harms, including prompt injection, sensitive information disclosure, improper output handling, and excessive agency.⁷ Microsoft's red-teaming guidance recommends maintaining a harms list, extending it as testing uncovers new harms, and following red teaming with systematic measurement, since red teaming reveals the risk surface but does not measure prevalence.⁸

Step 3: Size the sample and state the confidence

An accuracy figure from 50 examples and the same figure from 2,000 examples carry very different certainty. The bar should specify both the sample size and the confidence level, and the pass rule should use the confidence interval, not only the point estimate.

Sample size for a target margin of error

For a proportion such as accuracy, the standard approximation is n = z² × p(1 − p) / E², where z is 1.96 for 95% confidence, p is the expected accuracy, and E is the desired margin of error. Using p = 0.5 gives the most conservative (largest) sample. When you expect high accuracy, the required sample is smaller, as the table shows.

Two practical notes. First, the normal approximation is weak near 0% or 100% and at small samples. Report intervals using the Wilson score method, which NIST's Engineering Statistics Handbook describes and which, unlike the common formula, cannot produce a lower limit below zero.⁹ Second, if you will report accuracy by segment (for example, by document type or language), each segment you gate on needs its own adequate sample.

Confidence intervals in practice

Suppose a system gets 470 of 500 evaluation cases correct: 94.0%. The 95% Wilson interval is approximately 91.6% to 95.8%. If the bar requires the lower bound to be at least 92%, this result does not pass yet, even though the headline figure looks strong. With 1,910 correct out of 2,000 (95.5%), the interval narrows to approximately 94.5% to 96.3%. A larger sample does not make the system better; it makes the evidence clearer.

Bounding rare, critical errors

For S1 errors, the question is usually "how confident are we that the rate is below a ceiling?" If you observe zero critical errors in n independent cases, the exact one-sided 95% upper bound on the true rate is 1 − 0.05^1/n, which is very close to 3/n (often called the "rule of three"). In practice:

• Zero S1 errors in 300 cases supports a rate below about 1%.
• Zero S1 errors in 600 cases supports a rate below about 0.5%.
• Zero S1 errors in 3,000 cases supports a rate below about 0.1%.

When the required ceiling demands a sample larger than you can label, that is a strong signal to route the S1-prone cases to people by design rather than attempting to prove a rate statistically.

Step 4: Compare against the human baseline

The right reference point is usually the current process, not perfection. Measure the human error rate on the same sample, with the same grading rules, before setting the threshold. Then state the comparison explicitly: for example, "the system's S2 error rate must be no higher than the human baseline, and its overall accuracy must be within two points of the baseline or better." Record who graded the human baseline and how disagreements between graders were resolved, and measure inter-grader agreement where judgment is involved. If two experts agree only 90% of the time, a 98% bar against a single grader's labels is not meaningful.

Step 5: Define the exception route

An accuracy bar applies to the cases the system handles. The exception route defines which cases it does not. Specify the routing triggers (low confidence, missing data, policy flags, S1-prone categories), the queue, the human role, and the turnaround target. Report two numbers together: the share of volume handled automatically, and the accuracy on that share. A system that handles 70% of volume at 99% accuracy can be more valuable than one that handles 100% at 93%. For high-risk systems, the EU AI Act requires that people overseeing the system can disregard, override, or reverse its output and can stop it.¹⁰

Step 6: Set the review cadence

The bar is not only a launch gate. It is the standard the system is held to in operation. Define how often a fresh production sample is graded, which signals trigger an unscheduled review (input distribution shift, a rise in exception rates, a model or prompt change, a new data source), and what happens on a breach. The EU AI Act's post-market monitoring provision describes this kind of active, systematic collection and analysis of performance data over a system's lifetime.¹¹ OpenAI's guidance makes the same point plainly: evaluation is a continuous process.¹

Step 7: Sign it

The bar is signed by the business owner, the person accountable for risk or compliance in the workflow, and the delivery lead. Changes after signature are dated amendments with a stated reason. This record supports internal governance and aligns with the documented, continually improved management system that ISO/IEC 42001 describes.¹²

The template

Illustrative example

Example: inbound claims-document classification for an insurer. This example is illustrative. The organization and all figures are fictional and do not describe any Quantabase client.