# Don’t Ask Global Agent Anything, Ask These Three Things AI isn’t built for AMA (ask me anything). Get recommendations for outcome-based questions for the best results with Amplitude AI. Source: https://amplitude.com/en-us/blog/ask-ai-these-three-questions --- [Enzo Avigo](/blog/author/enzo-avigo) [Principal Product Manager](/blog/author/enzo-avigo) [](https://www.facebook.com/sharer/sharer.php?u=https%3A%2F%2Famplitude.com%2Fblog%2F%2Fblog%2Fask-ai-these-three-questions)[](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Famplitude.com%2Fblog%2F%2Fblog%2Fask-ai-these-three-questions)[](https://twitter.com/intent/tweet?url=https%3A%2F%2Famplitude.com%2Fblog%2F%2Fblog%2Fask-ai-these-three-questions\&text=Don%E2%80%99t%20Ask%20Global%20Agent%20Anything%2C%20Ask%20These%20Three%20Things)[](mailto:?subject=Checkout%20this%20Amplitude%20Article\&body=Check%20this%20out%3A%20https%3A%2F%2Famplitude.com%2Fblog%2F%2Fblog%2Fask-ai-these-three-questions) If a genie granted you three wishes, would you leave any of them up to chance with a broad, indirect request? Of course not. You’d be prescriptive. You’d start with a specific outcome in mind and tailor your words carefully to get exactly that outcome. AI is similar. Though the blank chat bar invites you to ask any question any number of different times, the most productive answers will come from the questions that are thoughtfully and carefully worded. Amplitude [Global Agent](https://amplitude.com/ai-agents) is designed to [think and act like an expert analyst](https://amplitude.com/blog/ai-thinks-like-analyst). It’s great at discovering insights in data, but even the best analyst (despite years of business knowledge and context) can’t read your mind or create exact answers from a blank slate. To get the best answers, you need to set the AI up for success. The best AI analyses should start with an outcome-based question and use the conversational abilities of an agent to make sequential discoveries that hone in on your ultimate answer. That’s how Amplitude AI Agents are designed to work. It’s easy to get answers from them, but to get the most valuable information, here are three ways I recommend starting your conversations. ## “Show me the \[buzzword].” AI is great at finding stuff. Whether it’s a chart, event, or metric, AI can comb through mountains of data, run analyses, and surface the right information in seconds. The best way to approach AI is to use buzzwords that describe key business outcomes and tell it to find relevant data. Great questions to ask include: - “Show me conversion” - “Show me traffic” - “Show me retention” - “Show me engagement” Our AI will find the right data (and the best chart) to discuss that outcome. It prefers using high-quality charts (judging by criteria, like most viewed, recently used, relevant to your team, etc.) that already exist. If no existing charts are a match, it will create a new one using your taxonomy of question types. AI Agents are trained on billions of data points about how analysts use our platform to uncover insights. They understand your business outcomes and the language you use to describe events. More importantly, they understand how specific visualizations chart the data necessary to answer business questions, including the relevant filters and dimensions along the way. This means users don’t need to know event names, chart titles, or where something lives to get answers. They just need to have an **outcome** in mind. Don’t ask for a specific chart or report. Even if that’s the way you’ve always done it, this is a new age. That’s a limiting, implementation-focused question, not an outcome-based question. Don’t limit the AI by telling it what you think the next step should be. Tell the agent where you’re going and let it help chart the course. Though you can ask broader questions if you want to explore, we recommend adding time windows and segments to get decision-ready, practical charts right out of the gate. For example: - “Show me retention for self-serve customers in EMEA after the October release.” - “Show me conversion from sign‑up to first value for new user.” - “Show me our most used trackable user interaction event.” You can ask follow-up questions from there to dig deeper and hone results, such as “refine this by …” or “compare this to …” ## “What’s causing this anomaly?” It’s useful to spot an anomaly, such as a spike or drop. It’s far more important to understand why that change happened. Human analysts typically need to zoom in and out to understand surrounding factors to diagnose the root cause. Things like seasonality, cohorts, experiments, marketing campaigns, product releases, etc., all come into play. Our AI Agents find those variables automatically. They synthesize thousands of surrounding data points, analyze patterns, and test possible explanations against your data to decide which are the strongest. It incorporates business context to connect dots across relevant sources into a clear, data-backed story. It picks relevant properties, like region, customer type, products, UTM parameters, and takes into account: - **Seasonality and trends** to identify if the observation is actually unusual vs last week, month, or year - **Sample size and noise,** filtering out tiny or inconsequential cohorts that look dramatic but don’t matter, and identifying suspected bot traffic - **Internal factors, **such as product releases and experiment history - **Data quality, **considering its impact on the results, before assuming the anomaly is the result of user errors or feature issues You can get to the root cause easily from anywhere in the Amplitude platform by: - Asking Global Agent why any anomaly is happening - Clicking into your chart, [hitting AI Insights](https://amplitude.com/blog/introducing-automated-insights), and viewing, accepting, or refining a suggested “What’s causing this spike?” question - Highlighting something on a dashboard and asking, “What’s causing this anomaly?” Anomaly detection workflows that would have historically taken hours now take seconds. But don’t stop at just figuring out the root cause, take action. You can ask AI for recommendations to turn your anomaly explanation into real change, such as rolling back a problematic release, tightening bot filters, fixing a data issue, or launching a follow-up experiment. ## “What’s the best performing \[X]?” Humans love to compare things. Who or what’s the biggest, most efficient, or best? AI loves that type of structure, only it processes information faster and better than we ever could. Answering comparison-based questions requires us to put things into perspective. Manually, this means looking at different data sets, each composed of lots of columns, lines, and filters. Humans usually can’t even fit all the data we need on a single screen, so there’s lots of scrolling, toggling, and squinting involved. Moreover, determining what’s the “best performing” isn’t as simple as comparing a single number. “Best” implies a well-rounded answer and requires a complicated calculation with various inputs, plus a standard to benchmark against. Answering a question like, “Which channel has the highest traffic?” is easy. Just look at the “traffic” column, and you have your answer. However, understanding “Which is the best performing channel?” is not as straightforward. In this scenario, let’s say you have 15 channels and each has 5 metrics associated with it: top-of-funnel metrics like engagement, bottom-of-funnel metrics like conversion, plus revenue, average order value, and retention. If you ask yourself, “Which is my best-performing channel?” your brain will probably freeze because it’s a very hard, nuanced question to answer. Some channels might bring tons of TOFU traffic but low conversion, so you can’t just go off of conversion. Others might convert well but bring in low-quality, low-retention customers. You need a calculation, maybe something that calculates volume x conversion rate. Running that calculation at scale would result in a big number that would compound in complexity when you multiply it again by every channel. It’s too much for our brains to compute efficiently. AI, on the other hand, lives for these kinds of comparison calculations. It consumes data and runs complex calculations in seconds, taking into account all these nuanced facets and benchmarking results against industry standards. Here are some of the best “best” prompts to get you started: - “What’s the best performing onboarding path for new B2B workspaces?” - “Which experiment variant is the overall best when you balance conversion and retention?” - “What’s the best performing acquisition channel for high-LTV users in the last 90 days?” - “What’s the best performing email subject line for >10k sends in the last 30 days?” - “What’s the best performing pricing tier for new US customers, excluding enterprise?” ## Put Amplitude AI to work today With AI, as with all things, you get out of it what you put in. Put in vague questions, and you won’t get meaningful answers. Ask Global Agent these three types of questions—”Show me the \[buzzword],” “What’s causing this anomaly?” and “What’s the best performing \[X]?”—and you’ll absolutely get a meaningful, data-rich conversation that fuels discovery and action. For more AI analytics best practices, check out these [blog articles.](https://amplitude.com/blog/tag/artificial-intelligence) [Get started](https://app.amplitude.com/signup?siteLocation=blog) with Amplitude AI today. ##### Ready to use AI to transform your product? Amplitude Agents help you understand your users more easily than ever. [Get started now](/signup?source=blog-ai\&topic=ai\&siteLocation=blog-inline-cta) About the author Enzo Avigo Principal Product Manager [More from ](/blog/author/enzo-avigo) [Enzo](/blog/author/enzo-avigo) Enzo is a Principal Product Manager at Amplitude, driving the transformation of analytics for product and marketing teams in the AI era. Topics [AI](/blog/tag/artificial-intelligence) [Agents](/blog/tag/agents) [Amplitude AI](/blog/tag/amplitude-ai) #### Recommended Reading [Read ](/blog/stop-chasing-churn) [Insights](/blog/stop-chasing-churn) ###### [The Problem with Chasing Churn](/blog/stop-chasing-churn) [May 18, 2026](/blog/stop-chasing-churn) [11 min read](/blog/stop-chasing-churn) [Read ](/blog/hostinger-conversion-lift-experimentation) [Customers](/blog/hostinger-conversion-lift-experimentation) ###### [How Hostinger Achieved a 20%+ Conversion Lift Through Experimentation](/blog/hostinger-conversion-lift-experimentation) [May 18, 2026](/blog/hostinger-conversion-lift-experimentation) [7 min read](/blog/hostinger-conversion-lift-experimentation) [Read ](/blog/stage-streams-smarter-with-data) [Customers](/blog/stage-streams-smarter-with-data) ###### [How STAGE Streams Smarter by Putting Data at the Center](/blog/stage-streams-smarter-with-data) [May 15, 2026](/blog/stage-streams-smarter-with-data) [6 min read](/blog/stage-streams-smarter-with-data) [Read ](/blog/building-the-validation-stack-for-ai-product-development) [Company](/blog/building-the-validation-stack-for-ai-product-development) ###### [Building the Validation Stack for AI Product Development](/blog/building-the-validation-stack-for-ai-product-development) [May 14, 2026](/blog/building-the-validation-stack-for-ai-product-development) [7 min read](/blog/building-the-validation-stack-for-ai-product-development)