Generative AI development: features that write, summarize and answer, with quality checks

Where generative AI earns its place

Generative AI is strongest when a first draft is valuable and a human can check it quickly. It is weakest when a single wrong detail is costly and nobody checks. We sort use cases by that test.

The last row is usually a no for a fully automatic feature. We build assistants for the professional, not replacements.

Generative AI versus agents

Generative AI produces content. An agent uses content-producing models to take actions across tools. Many projects need only the first. Our write-up on agentic AI vs generative AI explains the line, and the AI agent development page covers the second kind.

How we measure quality

Without measurement, every prompt change is a guess. We use three layers.

• A golden set of 30 to 100 real inputs with approved outputs, re-run after every change.
• A written rubric with four to six criteria, such as accuracy, tone, completeness and format. Reviewers score each output from one to five.
• Live sampling: after launch, a share of outputs is reviewed each week and scored the same way.

A simple example: a product description feature is scored on accuracy against the supplier sheet, house tone, length and banned words. If accuracy averages below the level you set, the feature stays in draft mode. Our LLM evaluation checklist lists the checks in more detail.

Hallucination, copyright and disclosure

Models sometimes state wrong things with confidence. See what is AI hallucination. We reduce it by supplying source text, asking for citations, requiring structured output and refusing when sources are missing. We cannot remove it, so review steps stay where errors cost money.

Other points we plan for:

• Transparency: the EU AI Act contains duties about labeling AI-generated content and telling people when they interact with AI. We list which apply to your feature. See what is the EU AI Act.
• Copyright: rules on AI-generated content and training data differ by country and are still moving. We avoid generating content that imitates named people or brands, and we keep records of sources.
• Privacy: prompts can hold personal data. We minimize what is sent and check the provider's data terms against GDPR, UK GDPR, CCPA/CPRA or the Australian Privacy Act as they apply to you.

Your legal adviser makes the final call on all three.

Our process

1. Pick one featureA single output with a named owner, such as weekly report summaries.
2. Gather examples and write the rubricReal inputs, approved outputs and the scoring rules.
3. Prototype on two or three modelsWe compare quality, speed and running cost on your golden set, then pick.
4. Build review into the flowAn approval screen, edit tracking and a way to flag bad outputs.
5. Pilot with real usersWeekly scoring and prompt changes, always re-tested on the golden set.
6. Hand overCode, prompts, golden set, rubric and a runbook for model changes.

What you get, timeline, reporting and related work

• A working generative feature inside your site, app or internal tool.
• The golden set, rubric and a script to re-run both.
• Edit tracking, so you see how often people change what the model wrote.
• A data and disclosure note for your review.

A single feature with clear examples often reaches a pilot within a few weeks. Several features, many languages or tight integration take longer. Your quote has dates for your scope, available through the contact page. During the pilot you get a weekly note with average rubric scores and edit rates.

We have no published generative AI case study, and no real project in our portfolio that fits closely enough to borrow, so we will not stretch one. The AI development page shows how we build in general.