Quality Gates

Quality gates are structured checkpoints built into AI workflows that verify output quality before it moves to the next stage or reaches a human reviewer. Each gate targets a specific type of error, creating layered defence against AI mistakes.

The five gate types

1. Format gate: Checks structure, length, and formatting requirements. "Does this have all five required sections? Is it under 300 words?"

1. Fact gate: Flags unverified claims and potential hallucinations. "List every factual claim in this output. Flag any you are less than 90% confident about."

1. Tone gate: Verifies voice and style compliance. "Rate this on formality (1-10). Does it match our brand voice guide?"

1. Completeness gate: Compares output against the original brief. "Compare this output against the brief. List anything missing."

1. Consistency gate: Checks for contradictions with previous outputs. "Compare this with the previous version. Flag any contradictions."

How gates work in practice

Gates can be implemented as follow-up prompts in the same conversation, as separate AI calls in an automation pipeline, or as programmatic checks (regex for word count, PII detection, etc.).

The simplest implementation is the "two-pass technique": after any important output, add a follow-up prompt asking the AI to review its own work for specific error types. This single follow-up catches 60-70% of errors because AI is better at reviewing than generating — the second pass applies a critical lens that the first pass, focused on creation, often misses.

The Swiss cheese model

No single gate catches every error. Like the Swiss cheese model in safety engineering, each gate has holes — but the holes are in different places. Stack multiple gates and the probability of an error passing through all of them becomes very small.

A practical quality pipeline might look like: 1. AI generates output 2. Format gate: checks structure and length 3. Fact gate: flags uncertain claims 4. Tone gate: verifies brand voice 5. Output passes to human for final review

The human reviewer at the end catches whatever slipped through the automated gates — but they are reviewing much cleaner output, making their review faster and more focused.

When to use gates

Not every AI output needs quality gates. Quick internal notes, brainstorming sessions, and exploratory research can skip them. But any output that will be seen by clients, published publicly, used for decisions, or included in reports should pass through at least one gate.

The rule of thumb: if the cost of an error exceeds the cost of checking, add a gate.