
It all started with a request from a customer: “Can you help us figure out if our new AI-driven chatbot is ready to go live?”
They were replacing a rules-based support chatbot with an LLM. The new system could understand and respond much more flexibly, but that also made it harder to evaluate. It was the first time I’d encountered a request like that and there wasn’t exactly a rulebook to follow. Which, to me, made the challenge all the more enticing.
We tested the LLM-powered assistant across a wide range of real support scenarios and saw all sorts of outcomes. Some interactions worked beautifully. Others broke down in ways that traditional usability measures couldn’t fully explain.
That experience stayed with me. I’ve always been drawn to research challenges that require us to look beyond the obvious. Mixed-method research has always appealed to me because numbers and people tell different parts of the same story. And this was a particularly interesting challenge: understanding not just whether an AI assistant could complete a task, but what the experience of working with it actually meant for the person using it.
That customer request became one of the starting points for AI Relationship Quality, or ARQ.
ARQ is a research framework designed to help us understand the quality of people’s interactions with AI-powered experiences. It examines five elements of the human-AI relationship: control, understanding, trust, outcome, and affinity. Together, they help us look beyond conventional usability metrics to understand what it actually feels like to work with an AI system, how well it supports what people are trying to achieve, and whether AI adds meaningful value to that interaction. Does its involvement make sense? Is it useful? Is it worth the effort?
As part of our preparations for ARQ’s full global rollout, we’re experimenting with different use cases. Currently we’re using ARQ to evaluate a selection of real-world, AI-powered tools and experiences.
Along the way, we’ve started noticing something rather intriguing: how people experience AI and how well it actually performs don’t always align. A positive experience doesn’t necessarily mean AI has delivered everything someone needs. Equally, an assistant can fail to complete a task while still leaving the person feeling informed, supported, and in control.
In one recent study, we examined a popular consumer AI assistant through the ARQ framework. We directed 200 users to have the assistant help them with three common activities and report back on their experience. Another 10 users completed the same actions while thinking aloud on camera so we could view their full, unfiltered reactions.
Two tasks were very simple: plan a day out and find a restaurant for dinner. The middle task deliberately asked the assistant to do something it couldn’t do (call an internet provider to downgrade their plan) so that we could gauge its response to failure.
Overall, the results were very positive. The assistant scored 86 out of 100 on ARQ and 89% of participants were satisfied with their experience. But looking beyond those numbers revealed something interesting.
For example, 95% of participants received a plan for their day out. But some people reported that their plans were missing essential details that they would have to find on their own, like specific venues, prices, addresses, or links. The assistant had done the thinking, and people liked it, but it hadn’t necessarily removed all the work.
The forced-failure task also revealed something intriguing. Many participants were perfectly fine with the assistant not being able to complete the task, especially when it clearly explained the limitation and offered something useful instead. (It might provide a phone number, prepare a script for the conversation, or suggest another way forward.)
The relationship became shaky when the assistant’s boundaries weren’t clear. People were uncomfortable when the assistant appeared ready to act without explicit permission or directed them toward account access.
Being helpful isn’t always about doing more. Sometimes it’s about knowing when to stop, explaining why, and letting the human decide what happens next.
We saw a related pattern in a second ARQ study of another AI assistant. Like the first assistant, it was overall very successful at the normal requests. But when it encountered the impossible task, this assistant didn’t handle the failure quite as elegantly. Generally, it was able to avoid inventing information or taking unwanted actions, which users appreciated. But it was much less consistent about explaining why it couldn’t help and offering a useful next step, which was reflected in its ARQ score.
These early studies have helped us refine ARQ and reinforced the value of looking at AI experiences from different perspectives. We’re learning that a successful outcome doesn’t always mean a good experience, and a failed task doesn’t necessarily mean a poor one. How AI communicates its limitations, handles uncertainty, and helps people move forward can be just as important as what it actually delivers.
For me, these findings show why ARQ needs to look at the human-AI relationship as a whole, rather than focusing on individual measures of success. I don’t expect the framework to stay the same. Different audiences, use cases, and stages of product development bring different questions. As AI evolves, so will people’s expectations of it, and it’s only natural for ARQ to evolve alongside them.
We aren’t replacing traditional usability research or technical evaluations, we're bringing these perspectives together to help us understand what makes AI experiences work for humans. We are using ARQ as a diagnostic tool to see where attention is most needed. Ultimately, AI should make our lives easier, more productive, and perhaps even more enjoyable, without taking away our sense of control and agency.