Meryem Ghatrif
Meryem Ghatrif

Contents

[email protected]

Pensées · Agile & AI

Agile, UX and AI: what speeds up, what stays human

AI speeds up the making, not the learning. How UX fits into agile teams, what to hand to AI, what to keep human, and what never to paste into a chatbot.

7 min read

In November 2022, Jake Moffatt’s grandmother died. Looking for a flight on Air Canada’s website, he asked the airline’s chatbot about bereavement fares. The answer was clear and reassuring: he could travel now and claim the reduced fare afterwards, “within 90 days of the date your ticket was issued”.

He booked. When he later applied, the airline refused: its policy did not allow retroactive claims. Before British Columbia’s Civil Resolution Tribunal, Air Canada argued it could not be held liable for information given by its representatives, chatbot included. In effect, wrote tribunal member Christopher Rivers in February 2024, the airline was presenting its chatbot as “a separate legal entity that is responsible for its own actions”. He called it “a remarkable submission” and ordered Air Canada to pay.

The sums were small, about 650 Canadian dollars in damages. The principle was not.

For teams that work in sprints, the case points to a bigger question. AI makes almost everything quicker to produce: what should go faster, and what must stay human? My short answer: AI can speed up the making. It cannot do the learning for you, or take the responsibility off your hands.

Seventeen people and a ski lodge

Agile itself began as a conversation. In February 2001, seventeen people met at The Lodge at Snowbird, a ski resort in Utah, “to talk, ski, relax, and try to find common ground”, as Jim Highsmith recalls in the manifesto’s history. They left with a short manifesto for building software. Its first line still reads: “Individuals and interactions over processes and tools.”

Designers then had to find their place in short sprints. Three answers stand out.

The first is Lean UX. In 2011, Jeff Gothelf wrote that “Lean UX gets designers out of the deliverables business and back into the experience design business.” Instead of long specifications, the team turns its beliefs about users into hypotheses and tests them with the smallest thing that could prove them wrong. His book Lean UX followed in 2013, later revised with Josh Seiden.

The second is the design sprint, which Jake Knapp began running at Google in 2010 and brought to GV in 2012: five days to map a problem, sketch, decide, build a realistic prototype and, on Friday, watch real customers react to it.

The third is research inside the sprint: small, quick studies, shared by the whole team. Three answers, one idea: shorten the distance between a guess and a real person’s reaction.

What AI genuinely speeds up

AI enters this story as an accelerator of the making. In her guide for the Interaction Design Foundation, Laia Tremosa maps where it helps: clustering messy notes, drafting interview guides and hypotheses, generating ideas and edge cases, writing copy and error-message variants, drafting documentation. A sprint can now hold far more drafts and prototypes than it used to. I think that is good news, especially for small teams.

But look at which part of the loop it speeds up. Lean UX and design sprints run on the same cycle: make something, show it to people, learn, decide. AI shrinks the first step. The others still move at human speed, because they involve humans.

The trap is that faster making looks like faster learning. A generated persona, a tidy affinity map, a polished prototype: they look finished. That is the danger. As Tremosa warns, “AI can be fast and persuasive, even when wrong.” Laura Klein’s advice to agile designers, published by the IxDF, reads differently today: “We need to stop designing faster and learn how to start designing less.”

People, not personas

In 2024, Maria Rosala and Kate Moran of Nielsen Norman Group tested AI-generated “synthetic users” on studies they had already run with real people. Asked about online courses, the synthetic learners said they had completed all of them. Many real learners had not. The researchers’ verdict: “Real user research is essential.”

An AI can tell you what people usually say. Only people can show you what they actually do: on a slow connection, in their second language, with a child on their lap. I came to UX because I care about people, and time spent with them is the part of the sprint I would protect first.

Judgement: the past is not a plan

In 2014, a team at Amazon began building a tool to rate job applicants from one to five stars. It learned from ten years of CVs sent to the company, mostly by men. By 2015, as a 2018 Reuters report revealed, Amazon saw that the system was not rating candidates for technical jobs in a gender-neutral way: it penalised CVs that included the word “women’s”, as in “women’s chess club captain”. The programs were edited, but nothing guaranteed the machine would not find other ways to discriminate, and the team was disbanded by early 2017. Recruiters looked at its rankings, Reuters’ sources said, but never relied on them alone.

The tool did what it was built to do: find the pattern in the past. Deciding whether that pattern deserves to be repeated is judgement, and judgement belongs to people. In a sprint, it shows up every day: which problem matters, which idea to drop, what not to build at all. The manifesto’s principles call simplicity “the art of maximizing the amount of work not done”. No tool will make that choice for you.

Accountability: your chatbot is still you

Back to Mr Moffatt. The tribunal’s reasoning was plain: “It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.”

If an AI feature speaks to your customers, someone on the team answers for what it says, and the interface should be honest about its limits. Microsoft’s Guidelines for Human-AI Interaction start exactly there: “Make clear what the system can do” and “Make clear how well the system can do what it can do.”

Accessibility and ethics work the same way. A generated layout does not know who will use it with a screen reader; the IxDF advises checking every AI output against real user needs and accessibility guidelines before trusting it. Giving people ways to “opt in, adjust, or override AI recommendations” is a design decision, and people make it.

What you paste leaves the room

In April 2023, a Samsung engineer accidentally uploaded sensitive internal source code to ChatGPT, Bloomberg reported. From 1 May, Samsung temporarily restricted generative AI on company devices and asked staff using these tools elsewhere “not to submit any company-related information or personal data”. Its worry: such data is hard to retrieve or delete, and could reach other users.

Most of us don’t guard industrial secrets. We hold something more delicate: what people have told us. Picture it: you have just interviewed eight patients of a small clinic, and a tool could summarise the transcripts in two minutes. The law already has a name for the right reflex: data minimisation. In Europe, the GDPR asks for personal data that is “adequate, relevant and limited to what is necessary” for its purpose, and for data protection by design and by default. In Morocco, law 09-08, overseen by the CNDP, asks for data that is adequate, relevant and not excessive and, as a rule, for the person’s unambiguous consent.

And replacing “Fatima” with “Participant 3” is not anonymisation: if the notes can still be linked back to her, they remain personal data.

Where I draw the line

I have studied agile methods for UX and AI for designers with the Interaction Design Foundation, and one conviction stays with me: someone who agreed to talk to a designer did not agree to talk to a server. My line is simple. No names, faces, voices or client documents go into an AI tool unless the people concerned have agreed and the tool’s data settings have been checked. Anything identifying comes out before anything goes in. If a study will use AI, participants should know before they say yes. When in doubt, summarise by hand: it is slower, but it is how you remember what people said.

AI in the sprint: a checklist

Delegate, then review:

  • First drafts: interview guides, hypotheses, user stories, release notes.
  • Variations: copy, error messages, layouts, edge cases to test.
  • Clustering notes that are already anonymised.
  • Quick prototypes, so real people see an idea sooner.

Keep human:

  • Talking to users, and watching them use the product.
  • Choosing which problem matters, and what not to build.
  • Answering for what any AI feature tells your customers.
  • Accessibility checks with real assistive technologies and the people who rely on them.
  • Ethical calls: fairness, consent, and a way to say no to the machine.

Never paste:

  • Passwords, keys or access tokens.
  • Anything a client shared in confidence, without their written agreement.
  • What identifies research participants: names, contact details, faces, voices.
  • Sensitive data: health, religion, political opinions, ethnic origin.
  • Anything you would not read aloud in front of the person it is about.

Look closer

My garden is my refuge and my lab. Nothing in it grows faster because I check on it more often; what changes is what I notice. Nature is infinite once you look closely.

The heartsease, the flower of this site, draws fine dark lines on its lower petals. Botanists call them nectar guides: they lead a visitor to the centre. A good sprint needs the same lines. Every draft, every prototype, every clever summary should lead back to a person. AI can widen the view. The loupe stays in our hands.

Sources