Assistant Output Acceptance Gate (QA)

Assistant Output Acceptance Gate (QA)

1. 01

   Freeze the client-facing deliverable spec before any prompt is written

   Name the format, length, tone, and factual constraints in one page the account lead signs off on. Without a frozen spec, reviewers argue about taste instead of defects.

2. 02

   Route the job to the model tier that matches its complexity, not the tier you pay for by default

   OpenAI's October 2, 2026 guide splits the GPT-6 family into prototyping, feature development, and multi-step orchestration variants; a 200-word ad variant and a 12-step research workflow should not run on the same tier.

3. 03

   Run every factual claim through a second model and log the disagreement

   Cross-model checks catch the failures a single assistant will confidently repeat. EuroEval's Dutch leaderboard shows top models separated by 2.5 points on fluency but 62 points on local knowledge, so a fluent answer is not a verified one.

4. 04

   Test the output against the client's own source documents, not the model's memory

   Upload the brand guide, prior campaign data, and product specs, then require the assistant to cite which file supports each claim. Anything it cannot attribute gets cut or flagged for the client.

5. 05

   Score the draft on a fixed rubric before a human edits a single line

   Use four gates: factual accuracy, brand voice match, structural compliance with the spec, and whether the piece answers a discrete question in its first 100 words. HubSpot's 2026 AEO guidance treats that first-100-words test as the baseline for answer engine visibility.

6. 06

   Log the prompt, model version, and reviewer initials against the deliverable

   Model behavior shifts between releases, so a retainer client asking why a March asset reads differently from a June one needs a traceable answer. This log is also the evidence base when a client questions an AI capability claim.

7. 07

   Escalate any output touching regulated claims to a named human owner

   Forrester's 2026 Consumer Benchmark found nearly nine in ten US and UK adults aware of AI but a persistent trust gap, sharpest in financial services. Client work in those verticals needs a documented human sign-off, not a reviewer checkbox.