The artefact was never the design.
When large language models arrived, much of the industry concluded that conversation design was finished: no more intents to map, no more flows to draw, the model writes the dialogue now. That conclusion mistakes the artefact for the discipline.
Maaike has been arguing against that confusion since long before ChatGPT. Back when the field debated "flow or no-flow", her answer was: neither, both. There is no such thing as a happy path, only users and the paths they want to take. The work was never to script one line through a diagram. It was to design the conversational space: the framework of conversations to be had, the context that follows the user through it, and the rules that keep it useful, safe and on-brand.
“It is easy to fall into the trap of thinking that the artefact is, in fact, the design. It is not – design artefacts are there to communicate design intent.”
Seen that way, generative AI didn't destroy conversation design. It destroyed the illusion that the flowchart was conversation design. Users have taken the reins and steer conversations with their own words. The designer's job moves to the space itself: blending the freedom of LLMs with the structure of dialogue management, so the experience meets users where they are and businesses where their legal and brand constraints live.
Prompts are the new flowcharts.
Every era of this field is tempted to redefine itself around its most visible artefact. Yesterday it was the decision tree; today it's the prompt. When the industry rushed to rebrand as prompt engineers, Maaike's warning was blunt: prompt engineering, essential as it is, revolves around the present, not the future.
Prompts are today's material, and she works in them daily, alongside personas, guardrails, retrieval and evals. But a discipline that defines itself by its current artefact pigeonholes itself out of the next shift and abandons its real job.
“We are not just language producers; we are designers with a capital D.”
The capital D means working up-stack, shaping how an organisation and its users understand each other. Purpose over prompts.
Should you even build that agent?
The most valuable question in an AI programme is the one asked least: does this need to be an agent at all? As Maaike puts it: “AI is not a spectacle, it is a potential solution to a user problem.” Fancy AI features, for all their glitter and glow, do nothing for your users unless they solve an actual problem.
Skipping that question feeds what she calls the infernal cycle of innovation: an agent ships without validated adoption, users ignore or distrust it, the failure discredits the technology itself, and the next, better-conceived initiative gets buried in the AI-experiment-cemetery.
Most AI projects, in her analysis, fail on foundations rather than technology: executive AI literacy (low literacy paired with high risk appetite is the most dangerous combination in a boardroom), organisational readiness, the state of the product the AI lands in, and whether users are even asking for it. She is fond of reminding leadership teams that AI will not fix a broken user experience. Her rule of thumb: scope an agent the way you would scope a new employee's job, with specific responsibilities, clear boundaries and measurable value.
The basics we keep forgetting.
Voice assistants, chatbots, LLMs, agents: every wave arrives with the same amnesia. Maaike has called the field a community that keeps forgetting its own fundamentals. The one it forgets first is user research, qualitative and quantitative, aimed at what people are actually trying to accomplish.
- Research before design. Who is the audience, what are they really asking, how do they feel about the brand? Those questions have opened every project she has run, from Alexa skills to enterprise agent suites. A demo answers none of them.
- Task completion over likeability. Since 2022 she has tracked the field's shift away from charming personas and affective trust toward one blunt measure: did the user get the thing done?
- Language is culture. Translating the words in your script is not enough. She calls the real job transcreation: rebuilding tone, references and persona for each culture. Her book devotes a chapter to multilingual scriptwriting for the same reason her responsible-design principles insist on informed empathy, inclusive word choice and flexibility for neurodiverse users.
- You can't design conversations without having them. Table-reads, role-play, Wizard-of-Oz testing: her Voice Design Sprint and its serious games exist because teams only align, and designs only improve, when people say the conversation out loud.
What this looks like in practice.
Pulled together, her method is less "write better dialogue" than a product discipline for conversations:
- Foundations first. Assess the points of influence (leadership, organisation, product, users) and establish problem-solution fit before any model choice. Choosing the technology first is deciding you'll take the plane before deciding to visit your neighbour.
- Design the space, not the script. Conversational maps instead of linear paths; LLM freedom blended with dialogue-management structure; context carried across turns and agents; multilingual from the start.
- Have the conversations, early. Prototype by talking: table-reads, role-play, Wizard-of-Oz sessions with real users, before the engineering hardens around unvalidated assumptions.
- Prove it, continuously. Task-completion metrics, evaluation frameworks and testing that turn "it seems to work" into evidence: for users, for the business and, in regulated industries, for the people who audit you.
Where this is heading.
Maaike's path runs from award-winning voice apps, through the French reference book on conversation design, to multi-agent systems in production in insurance and healthcare. The throughline has never been the artefact of the moment. It has always been the same conviction, restated for each new wave of technology:
“Embrace the shift towards user-centered conversations, focus on strategic design, and prioritize purpose over prompts.”