Checklist
Before an AI or Machine Translation Decision
The evidence to request before accepting a provider claim about AI quality, savings or human-review routing.
TL;DR
Require a named baseline and comparable content for any AI quality or productivity claim, examine human correction, routing thresholds, and commercial units alongside the generated output, and separate deployed results from proposed assumptions before specifying go/no-go conditions for a rollout or pilot.
Separate a proposed AI benefit from a measured one. For a deployed workflow, obtain examples showing actual human review, routing and correction effort alongside throughput claims. Keep the date, version and measurement basis with each provider claim.
Evidence behind the AI claim
Record the provider's claimed benefit in writing: the comparison baseline, content, languages, measurement period and exclusions.
For every relevant workflow, identify the model or service, version, routing rules, fallback behavior and review or pass thresholds.
Request paired raw machine output and final reviewed text by language and content type when the workflow is deployed.
For those same paired samples, request human correction effort and time, including discarded, retried or escalated work.
Document what gets no review, light review and full review, who can override automation and how that decision is logged.
Capture the billing unit for each route (word, token, task or subscription), minimums, human-review fees and additional technology consumption.
Keep dated changes to models, prompts, context, thresholds and routing rules, and the buyer notices for them.
Request technology-vendor terms covering data use, training, retention, region and subprocessors; legal and security specialists assess compliance.
For a proposed workflow, document planned content and languages, baseline assumptions, go/no-go conditions for a controlled rollout or pilot and accountable human owner.
What the record can establish
Paired output and effort may establish a deployed-workflow comparison when sampling, content and timing are comparable. A proposed workflow is a forecast or test design, not measured savings.
The buyer decision
Ask which content can safely auto-pass, which findings require intervention, which costs shift to the buyer and what changes would reverse the decision.
If workflow, evaluation, or human-review evidence is missing, name what must be validated before recommending a wider AI rollout.