01Process reviewWe look at how your team works today and agree on how to measure success.
02Tasks and responsibilitiesWe break the workflow into tasks and identify which ones AI can handle and which decisions need your team’s approval.
03Model selection by taskWe select a model for each task based on output quality, response time and cost.
04Model instructions and output formatsWe write the instructions the model will follow, define JSON schemas for structured outputs and track changes by version.
05Data access and toolsWe define which sources the model can consult, which tools it can use and how access to personal data is restricted.
06Testing and evaluationWe compare outputs with reference examples, rerun earlier tests on each new version and submit the results to the designated reviewer.Reviewer sign-off
07Training and handoverWe show your team how to run the workflow and monitor its results. We hand over the operating guide and the current prompt library.
Sample release evaluation
Prompt set v1.4
Evaluation criterionAcceptance thresholdSample result
GroundednessScore indicating how well approved sources support the information in a response.≥ 0.950.97
Schema complianceShare of outputs that conform to the specified JSON schema.100%100%
Terminology complianceShare of terms used in accordance with the organisation’s glossary.≥ 98%99.2%
Personal data detectionNumber of personal data instances detected in the reviewed outputs and logs.00
Response time (p95)Time within which 95% of tasks completed.< 4 s2.8 s
Cost per taskProcessing cost per task compared with its allocated budget.Within budget18% below budget
All 6 criteria met.
Once all acceptance criteria are met, we review the results and release the version with the designated reviewer’s approval.
Method 10 principles
We begin every workflow with measurable targets and acceptance criteria agreed with the business owner.
Instead of one large prompt, we build steps with defined inputs and outputs, and measure and fix each one separately.
We define model outputs with JSON Schema and never pass unvalidated free text to the next step.
We keep prompts under version control and release no change that has not passed the evaluation set.
We re-run a test set of real, expert-labelled examples on every release to catch regressions before launch.
When a model scores outputs, we check its scores against human labels and re-validate whenever the rubric changes.
In steps that need facts, we make the model answer only from approved sources (RAG) and cite them.
We choose each step's model on quality, cost and latency, and route simple tasks to smaller models.
In line with KVKK and GDPR, we mask personal data, send no unnecessary data to models and define retention periods.
We release high-risk outputs only after human approval and log every model call with its cost, latency and error rate.