AI Confronts Response Time Challenge: A Crucial Overlooked Threshold

The Doherty Threshold and Its Crucial Importance

In the field of artificial intelligence, the concept of the "waiting problem" is well known. Since 1982, the Doherty threshold, which sets the ideal response time at around 400 milliseconds, has been a key reference point. This threshold is crucial for maintaining user engagement: below it, users remain focused and involved, but beyond it, their attention begins to wane. Unfortunately, in 2023 and 2024, none of the leading AI products managed to meet this threshold. Worse still, products launched in 2025 and 2026 failed even more dramatically, with response times stretching into several minutes.

A Well-Documented Threshold

The Doherty threshold is a well-established concept in the field of human-computer interaction (HCI) for over four decades. Research has shown that response times under 400 milliseconds allow users to "think with" the machine rather than passively waiting. A study conducted by Brad Myers in 1985 also highlighted the importance of progress indicators, which can nearly compensate for perceived wait time and improve task completion rates.

Behind the Scenes of AI Products

When you interact with an AI product, several scenarios can unfold between the moment you hit "Enter" and when the response appears:

• In the best-case scenario, you receive a continuous stream of tokens. The first token typically appears within one to two seconds, followed by others over a period of ten to thirty seconds, creating an impression of ongoing progress.

• In a median case, you see a simple ellipsis or a blinking dot, with no indication of remaining time or completion percentage. This lack of feedback leaves users in uncertainty.

• In the worst-case scenario, in agent mode, no useful information is provided. The agent updates its status every twenty to thirty seconds, then remains silent for several minutes.

User Strategies for Coping with Waiting

In the absence of clear feedback, users have developed strategies to cope with this waiting:

• Checking a secondary tab: Users submit a task, then open a new tab to engage in other activities, periodically returning to check the result.

• Refreshing for verification: During a long generation, users refresh the page to ensure that the connection is still active, although this can sometimes interrupt the ongoing response.

• Prompt "are you stuck or just slow?": Users send a message to check if the system is still responsive.

• Screen recording: Advanced users record their screens during long operations to analyze what the agent did afterward.

The Impact of Agents on Response Times

AI products from 2023 and 2024 already exhibited response times ranging from a few seconds to several tens of seconds, which was frustrating for users. However, the agents introduced in 2025 and 2026 exacerbated the problem, with response times extending from several minutes to several hours, placing them in a category comparable to compilation or video rendering operations.

Towards a Better User Experience

To design an effective waiting user experience in the field of AI, it is essential to leverage forty years of HCI research. Such UX should include at least four fundamental characteristics:

• Continuous progress, rather than a simple blinking dot. Even for short operations, the product should show in real-time what the model is accomplishing.

• Estimated remaining time, continuously updated to inform the user.

• Completion notification, visible through the operating system to signal the end of the process.

• Action log, accessible during and after execution to allow for detailed analysis.

These elements would help reduce user anxiety and improve their overall experience with AI products.