Part of AI consulting

Implementing AI features, from requirements to tested answers

I work on AI features in websites and services, such as AI search, answer summaries and assistants: deciding what the feature should do, writing it down so it can be built or bought, testing the answers it gives, and fitting it into the content and interface around it. I advise, specify and test; your team or your vendor builds it.

My background in UX, information architecture and front-end work means I understand how an AI feature works as well as what it should do, and I can work with developers, designers, content teams and leadership on the same feature.

Questions I help answer

A product owner asks

“We want an AI assistant. What should it actually do?”

What I check
The questions people bring, which ones it should answer, and which it should hand to a person.
What you get
Requirements your team or vendor can build or buy against.

A service owner asks

“It goes live next month. How do we test it?”

What I check
Questions people really ask, with every answer checked against its source.
What you get
The results, the content fixes behind wrong answers, and a method to run again.

A content lead asks

“The answers are vague. Is it the model or our content?”

What I check
The content and structure it answers from, because they decide how good the answers can be.
What you get
The pages and structure to change, in order.

What AI implementation covers

  • Deciding what the feature should answer, what it shouldn't, and how you'll know it works
  • Writing requirements so the feature can be built in-house or bought from a vendor
  • Comparing AI and search platforms, vendor-neutral
  • Testing the answers with questions people really ask, judged the way they would judge them
  • Shaping the content and structure the AI answers from, because they decide how good the answers can be
The Claude Console playground: a system prompt that limits answers to Wellington City Council's dog fees page and keeps every condition, a question from an accredited owner paying late, and the answer: most likely $222, with the lost discount and a note that the page doesn't say exactly, so call the council
Specifying behaviour and testing it: a hard question from a test set, run against the Claude API (October 2026).

Where I've done it

At MBIE I wrote requirements for modern, AI-assisted search across MBIE's websites, and was part of a team that compared AI and enterprise search platforms.

Testing language models in my own demos keeps me close to what they do with real content, where they go wrong and what they cost to run. The Ask about Miro assistant on this site is one of them: it answers from my background and projects, and links to its sources.

Findability and AI search at MBIE

Need this for your agency? I help with it as AI project consulting: requirements, platform evaluation and testing the answers.

Common questions

What does AI implementation advice cover?

Helping an AI feature, such as AI search, an answer summary or an assistant, do what people need: deciding what it should do, writing requirements so it can be built or bought, testing the answers it gives, and fitting it into the content and interface around it.

How do you know whether an AI feature's answers are good enough?

Test it with the questions people really ask, judged the way they would judge an answer, and check each answer against its source. A demo isn't evidence.

Should an AI feature be built in-house or bought?

Either can work. Write the requirements so they hold for both, then compare platforms vendor-neutrally on real test questions and your own content before choosing.