chapter 1
Earlier this year, Pendo and Appinio asked 2,000 enterprise employees about their feelings towards and behaviors around AI. Some insights fell in line with our expectations: Nearly two-thirds of employees want AI to help them work faster, and one-fourth want to find new ideas.
But some data points surprised us: half of all employees are straight-up dissatisfied with the tools they're using, and confidence in AI accuracy is low—only 33% trust AI agents and their results.
Yet even when AI agents fail to deliver, employees aren’t giving up right away. Instead, they spend time re-prompting, until they switch tools and abandon AI altogether for its more traditional counterpart (miss us, Google?).
This is just one example of what we’re calling the AI Experience Gap: the distance between what users expect, and what they’re actually getting from AI tools.
It’s the reality for most AI agent users, and we’ve traced this back to a singular root issue: Organizations are measuring the wrong AI metrics.
When it comes to AI agents, users don’t expect novelty. They simply want practical value.
62% use AI to save time, 35% use it to improve productivity, and another 25% use it to find new ideas and inspiration. The hope is simple: make work easier. Help me do what I’m already doing, with less manual work and friction.
chapter 2
Sentiment data tells a more complicated story. Users aren’t AI skeptics, but they’re also not fully bought-in on its promises. When asked about how they feel about AI today, the responses point towards cautious optimism:
Users aren’t confident that agents will deliver the right results or information, and they’re expecting the worst-case scenario when it comes to hallucination. There’s a long road ahead for AI builders and their end-users.
chapter 3
Users aren’t confident that agents will deliver the right results or information, and they’re expecting the worst-case scenario when it comes to hallucination. There’s a long road ahead for AI builders and their end-users.
Who is responsible for making AI agents actually work? Is it the AI Product Manager, the engineer that’s coding the tool, or the company that buys the agent and rolls it out for its employees?
Users aren’t sure. When asked who plays the biggest role in ensuring AI tools are reliable and trustworthy, they said:
Although users may not agree on who is responsible for improving AI agents, there’s one thing users do agree on: 83% believe companies should provide training or guidance on how to use AI tools effectively and responsibly.
Users want to get better at using AI, and they’re seeking guardrails, guidance, and support.
Users are aware that AI can assist them, and they expect AI performance to improve.
In fact, 38% of users expect AI to make a positive impact on productivity and job satisfaction over the next year, whereas only 8% expect a negative impact.
chapter 4
The gap between what users expect and what they’re actually getting from AI tools comes down to one thing: Organizations are measuring the wrong things.
Activity indicators, such as logins and conversation volume, tell you if an agent is being adopted, but not if it’s actually working and improving the user experience.
Closing this gap requires four things:
The first step to improving agent impact is being able to answer questions, like “Did the user accomplish their goal?” “Did they save time compared to the traditional workflow?”, and “Did they come back, or did they give up?”
Re-prompting agents repeatedly, abandoning sessions, and user rage are leading indicators of a broken experience.
Track them, see which use cases they’re correlated to, and improve this.
You just read that most users are turning to AI for productivity and speed. Now, prove this by benchmarking agent task completion times against the traditional workflow.
Teams should also monitor agent retention by use case and user segment, so they can pinpoint which types of problems are driving the best outcomes. Some use cases don’t make sense for an agent, and that’s totally okay. Not every challenge needs an AI agent to solve it!
Users who trust AI will get more from it. Training, guidance, and transparency about AI capabilities (and limitations) will help companies buying and building agents drive adoption.
chapter 5
AI builders and buyers need greater visibility into how AI agents are actually performing. Today, organizations have three paths:
Dev-centric platforms, like Arize and LangSmith, are built for ML engineers and AI developers. They excel at model observability, tracking metrics like latency, token usage, hallucination rates, and prompt/response quality at the technical layer.
Dev-centric tools are best for teams debugging model performance, optimizing prompts, and monitoring LLM behavior in production.
However, these tools struggle to answer the question, "Is the user getting value?"
PMs, CX leaders, and business stakeholders don’t get a unified view into how the agent is performing in the context of broader product experience, and if it’s delivering value to the people using it.
PM-centric agent measurement tools focus on the user rather than the model. Instead of asking "Did the LLM respond correctly?" it asks: "Did the user get what they needed? Did they save time? Did they come back?"
While dev tools measure AI performance, PM tools measure impact. With a product-centric tool, you can measure:
This is the most popular option (and the most expensive). Organizations that don't measure agent performance don't know if their AI investments are paying off.
They can't identify friction until it becomes churn. They can't prove productivity gains to leadership. And they can't answer the question users are already asking: "Is this actually helping me?"
Every frustrated user, every abandoned session, every hour lost to re-prompting is a tax on the AI investment.
chapter 6
Users are abandoning AI because it keeps wasting their time. The 69% frustration rate, the endless re-prompting loops, the quiet switch back to Google: these are symptoms of organizations measuring adoption when they should be measuring outcomes.
Now, everyone is a builder. And to improve AI systems, teams need the right blend of agent measurement: tracking task completion, time saved, and agent retention.
Doing this will help close the experience gap, create agents that get prompts right the first time, and help businesses trace ROI back to their investments.
The findings are based on a survey of 2,000 adults from the UK and Germany conducted by Appinio on behalf of Pendo in October 2025.
The research explored usage frequency, satisfaction, trust, and emotional responses to AI-powered tools, providing insight into how people interact with emerging agentic technologies.
Earlier this year, Pendo and Appinio asked 2,000 enterprise employees about their feelings towards and behaviors around AI. Some insights fell in line with our expectations: Nearly two-thirds of employees want AI to help them work faster, and one-fourth want to find new ideas.
But some data points surprised us: half of all employees are straight-up dissatisfied with the tools they're using, and confidence in AI accuracy is low—only 33% trust AI agents and their results.
Yet even when AI agents fail to deliver, employees aren’t giving up right away. Instead, they spend time re-prompting, until they switch tools and abandon AI altogether for its more traditional counterpart (miss us, Google?).
This is just one example of what we’re calling the AI Experience Gap: the distance between what users expect, and what they’re actually getting from AI tools.
It’s the reality for most AI agent users, and we’ve traced this back to a singular root issue: Organizations are measuring the wrong AI metrics.
When it comes to AI agents, users don’t expect novelty. They simply want practical value.
62% use AI to save time, 35% use it to improve productivity, and another 25% use it to find new ideas and inspiration. The hope is simple: make work easier. Help me do what I’m already doing, with less manual work and friction.
Sentiment data tells a more complicated story. Users aren’t AI skeptics, but they’re also not fully bought-in on its promises. When asked about how they feel about AI today, the responses point towards cautious optimism:
Users aren’t confident that agents will deliver the right results or information, and they’re expecting the worst-case scenario when it comes to hallucination. There’s a long road ahead for AI builders and their end-users.
Users aren’t confident that agents will deliver the right results or information, and they’re expecting the worst-case scenario when it comes to hallucination. There’s a long road ahead for AI builders and their end-users.
Who is responsible for making AI agents actually work? Is it the AI Product Manager, the engineer that’s coding the tool, or the company that buys the agent and rolls it out for its employees?
Users aren’t sure. When asked who plays the biggest role in ensuring AI tools are reliable and trustworthy, they said:
Although users may not agree on who is responsible for improving AI agents, there’s one thing users do agree on: 83% believe companies should provide training or guidance on how to use AI tools effectively and responsibly.
Users want to get better at using AI, and they’re seeking guardrails, guidance, and support.
Users are aware that AI can assist them, and they expect AI performance to improve.
In fact, 38% of users expect AI to make a positive impact on productivity and job satisfaction over the next year, whereas only 8% expect a negative impact.
The gap between what users expect and what they’re actually getting from AI tools comes down to one thing: Organizations are measuring the wrong things.
Activity indicators, such as logins and conversation volume, tell you if an agent is being adopted, but not if it’s actually working and improving the user experience.
Closing this gap requires four things:
The first step to improving agent impact is being able to answer questions, like “Did the user accomplish their goal?” “Did they save time compared to the traditional workflow?”, and “Did they come back, or did they give up?”
Re-prompting agents repeatedly, abandoning sessions, and user rage are leading indicators of a broken experience.
Track them, see which use cases they’re correlated to, and improve this.
You just read that most users are turning to AI for productivity and speed. Now, prove this by benchmarking agent task completion times against the traditional workflow.
Teams should also monitor agent retention by use case and user segment, so they can pinpoint which types of problems are driving the best outcomes. Some use cases don’t make sense for an agent, and that’s totally okay. Not every challenge needs an AI agent to solve it!
Users who trust AI will get more from it. Training, guidance, and transparency about AI capabilities (and limitations) will help companies buying and building agents drive adoption.
AI builders and buyers need greater visibility into how AI agents are actually performing. Today, organizations have three paths:
Dev-centric platforms, like Arize and LangSmith, are built for ML engineers and AI developers. They excel at model observability, tracking metrics like latency, token usage, hallucination rates, and prompt/response quality at the technical layer.
Dev-centric tools are best for teams debugging model performance, optimizing prompts, and monitoring LLM behavior in production.
However, these tools struggle to answer the question, "Is the user getting value?"
PMs, CX leaders, and business stakeholders don’t get a unified view into how the agent is performing in the context of broader product experience, and if it’s delivering value to the people using it.
PM-centric agent measurement tools focus on the user rather than the model. Instead of asking "Did the LLM respond correctly?" it asks: "Did the user get what they needed? Did they save time? Did they come back?"
While dev tools measure AI performance, PM tools measure impact. With a product-centric tool, you can measure:
This is the most popular option (and the most expensive). Organizations that don't measure agent performance don't know if their AI investments are paying off.
They can't identify friction until it becomes churn. They can't prove productivity gains to leadership. And they can't answer the question users are already asking: "Is this actually helping me?"
Every frustrated user, every abandoned session, every hour lost to re-prompting is a tax on the AI investment.
Users are abandoning AI because it keeps wasting their time. The 69% frustration rate, the endless re-prompting loops, the quiet switch back to Google: these are symptoms of organizations measuring adoption when they should be measuring outcomes.
Now, everyone is a builder. And to improve AI systems, teams need the right blend of agent measurement: tracking task completion, time saved, and agent retention.
Doing this will help close the experience gap, create agents that get prompts right the first time, and help businesses trace ROI back to their investments.
The findings are based on a survey of 2,000 adults from the UK and Germany conducted by Appinio on behalf of Pendo in October 2025.
The research explored usage frequency, satisfaction, trust, and emotional responses to AI-powered tools, providing insight into how people interact with emerging agentic technologies.
