A research process run by AI and automation
With three researchers on sixty designers, and no option for additional capacity anytime soon, we needed to identify opportunities for improving craft quality and impact visibility. So we employed AI and automations to assist us in the research process.
- ING
- Head of UX Research
- Three researchers, me included
- Built in stages, still running
The problem
Three researchers, a bank, and a design centre of excellence of more than sixty designers. That is before anyone counts the stakeholders outside design. Far more demand than three people can take on. More capacity would have been welcome. It was not on offer. And the bigger the company, the more process sits around every single study. That process is its own drag on how much a team can carry. So the question became what we could improve without adding people. We have done several things about that. This is one of them.
Requests arrived as direct messages. Whoever held the relationship with that stakeholder, or happened to be online, got the information, and the rest of the team never saw it. A lot of it was simply lost.
Setting a project up meant making the same pages and folders by hand, every time. Notes were written up after a session if somebody got round to it, and when they were, they often sat on one person’s laptop where nobody else could find them. And the insights, the actual point of the whole thing, went into a report, got presented once, then sat in a slide deck nobody opened again.
How do we take on more research, without adding people and without dropping the standard?
The plan
Two pillars, two projects in each. One pillar bought time. The other built the argument for more people, because headcount needs a business case and a business case needs data we did not have.
Do more research with three people
- Tooling The things that make a single study quicker to run. Bought where they existed, built where they did not.
- Operations This study The process wrapped around every study, from the request landing to the insight being found again later.
Earn the headcount
- Quality reviews A standard we could show someone, so the argument rested on the work.
- Research metrics A dashboard putting the value of research in front of the people who sign off budgets.
What we tried
I drew the whole process first, end to end, before touching a tool. Then I built it in pieces. I built every piece myself. That is slower than handing it to engineers. It still fits the way researchers actually work.
- Draw the whole process, end to end Request through to delivered report and presentation, before touching a single tool. Drawing it is also how you find the steps that exist only out of habit.
- Design the flow, then build it in pieces Each piece designed, built, and put in front of the people who have to live in it. Never one big system, released once.
- Test with my own team, then with stakeholders Half of this system belongs to people who do not work in my team. They will not read a manual.
The design
Eight of the fourteen steps run without anyone touching them. Four need a person: the ask, the decision that it is worth doing, the notes a researcher writes in the session, and the research itself. Twice in the whole run the system stops and tells a researcher it is their turn. That split was the decision. We went through the process asking where AI could take admin off a researcher’s day and where the craft had to stay human. Automating the admin bought time for the craft. The work got faster. It also got better.
-
A designer asks for research A person does this. -
It lands on the research board Runs by itself. 
The researcher is notified The system tells a researcher it is their turn. - The researcher reviews it and approves A person does this. An intake is scheduled with the product team Runs by itself.
-
The session is recorded Runs by itself. 
It is transcribed Runs by itself. The researcher takes their own notes A person does this. - The notes and the transcript are merged Runs by itself. A research plan is drafted Runs by itself.
- The researcher is pinged to review the plan The system tells a researcher it is their turn.
- The researcher runs the research A person does this.
- Insights are logged Runs by itself. They are shared with stakeholders Runs by itself.
Logged insights feed back into the next research plan, so a new study starts from what we already know.
Outcome
What got faster was the admin. Researchers were not hired to do admin. It is still running. It is still only in the Netherlands. The piece that would gain most from crossing a border is the insight database. Every country has its own reasons to build its own version of that. The boards and prompts are not mine to publish, so this describes the system without showing it. I am happy to walk through it properly in a conversation.
Per researcher, up from one to three. Where it lands depends on the researcher and how big the projects are.
Down from three to four weeks, and only if the request reached the right person at all.
Down from one to two weeks of writing from an empty page. The researcher edits a draft instead.
Down from two weeks after the final report, when it happened at all. Now it always happens.