Beyond the Toggle Tax: A Design Principle for Friction-Free AI
A single work task is rarely confined to one place. For instance, a customer update might begin in a dashboard, take shape in a document, gather feedback in Slack, and reach its final audience through an email. Each individual tool supports part of the outcome, yet they are rarely connected.
The strain at work comes from the extra steps required to toggle between each tool. Someone has to remember where information lives, locate the right source, reconstruct the context, and find their place in the original task. And they repeat that sequence all day.
Over the course of a workday, those small detours add up. A study of 137 employees across 20 teams and three Fortune 500 companies found that workers switched between applications and websites nearly 1,200 times each day. The time spent reorienting after those switches added up to almost four hours a week. After 65% of switches, the person moved to another application in less than 11 seconds.
AI is entering this already fragmented environment, but it’s rarely doing anything to reduce the fragmentation. It can help people find information, draft content, and take action more efficiently, but that help often requires a detour. People may still have to leave their work, find the tool, explain the task, provide relevant context, and transfer the output back to where they were working.
But that experience is a narrow representation of how to potentially use AI at work. Friction-free AI removes those avoidable detours. It follows work already underway, understands the current context, and provides help where the task is happening.
The tax lives in the reorientation
Cognitive load, the toggle tax, and the workaround economy have given teams language to describe the friction characteristic of modern work. Each term describes the same experience: People spend much of the day managing how work gets done rather than doing the work itself.
“Every time you switch, you have to re-remember what you were working on and rebuild the context,” said Noah Silverstein, Head of Product Marketing for Superhuman Go.
Disruption begins when the person has to find a tool, or multiple tools, before they can complete a task. For each new tool, they must restate the task, copy relevant information into a prompt, and carry the answer back.
“I reach a good flow state when there aren’t obstacles to obscure my train of thought, and I can keep going in whatever direction that takes me,” Silverstein said. That direction may cross tools and lines of thinking. Continuity comes from letting the person pursue the task without constantly needing to read a new tool into what’s happening. As a comparison, it would be understandably disruptive to bring a new team member up to speed on a midstream project multiple times a day.
Capability still needs a path to the work
AI tools can already perform far more tasks than most people use them for. “There is a large chasm between AI’s capabilities and AI’s adoption,” said Noam Levinsky, Chief Product Officer at Superhuman. “It’s often referred to as the capability overhang.”
Part of the gap comes from uncertainty. People may not know which tasks to delegate, how to explain the work, or whether they can rely on the result. Product delivery creates another barrier. A capable system provides limited value when people have to recognize an opportunity, remember which tool can help, and redirect their work toward it. “I think we often underestimate the user experience in how automation is delivered,” Levinsky said.
Closing the capability gap requires both reach and integration:
- Reach makes AI available when an opportunity for help appears.
- Integration gives AI enough context and access to contribute within the existing workflow.
A system with reach may sit one click away in a browser extension, the detour mentioned earlier in this piece. Integration, on the other hand, allows the AI to participate in the work itself. “We’re not trying to give you just another place where there’s more work to do,” Levinsky said. “We’re trying to bring AI into your existing applications and workflows.”
More capabilities will not close the gap on their own. Product design determines whether those capabilities reach the work.
Follow the path already underway
A person writing a memo might pause because they cannot remember a statistic or where it came from. The usual path requires leaving the document, finding the right dashboard or study, locating the number, and returning to the sentence.
“What we’re really trying to do with Superhuman Go is build a tool that follows you,” Silverstein said. That requires more than an AI assistant you have to go to. Superhuman Go can understand the surface someone is working on, draw on connected information, and help within the current task.
In the memo example, an integrated agent can surface the statistic and its source while the person is still writing. The same principle applies across a workday:
- While someone drafts a Slack message, an agent can bring forward the relevant customer update.
- While someone schedules a meeting, an agent can supply appropriate availability.
- Inside a document, an agent can highlight text, leave a comment, or suggest an edit inline.
- During a project, an agent can draw on connected email, calendar, and project information without requiring someone to collect it first.
These interactions combine three elements:
- Proximity places AI within reach.
- Context awareness helps AI understand what the person is trying to accomplish.
- In-surface action allows AI to contribute where the work is happening.
The person continues pursuing the original outcome. The information and assistance come to where the work is taking place.
Remove the detours
A practical test for friction-free AI is whether someone can receive the benefit without having to find the tool. Teams can evaluate an AI experience by asking:
- Does the AI appear where the work is already happening?
- Can it understand the current task without requiring the person to rebuild the context?
- Can it contribute inside the workflow rather than returning an answer that must be transferred elsewhere?
- Can the person inspect, accept, revise, or dismiss the help without leaving the task?
- Does the output remain connected to the people, information, and systems involved in the work?
- When the AI acts proactively, is the intervention relevant enough to justify the interruption?
The final question becomes increasingly important as AI shifts from responding to prompts to recognizing opportunities for help. “Blanket proactivity can be much more annoying than helpful,” Silverstein said. Useful proactivity requires enough context to understand the task, the person’s preferences, and the value of intervening at that moment.
To apply the principle, examine one workflow at a time: Observe the path people already take; identify where they stop to search, switch, or transfer information; bring AI into those moments with the right context; keep the output where people will use it; and measure the friction removed.
“I get to follow the path of what I was trying to get done, and I don’t have to take these detours,” Silverstein said. That is the standard for friction-free AI: capable assistance that reaches the work without pulling people away from it.
Learn more about how Superhuman is applying this principle across Go, Docs, writing, email, agents, and connected work.
