Let’s say someone on your marketing team notices that signups dipped on Tuesday, and they want to know why. For most organizations, you’ll find out in about a week, once the request works its way through the analytics queue and an analyst has time to dig in. By then the campaign has run, the moment has passed, and the answer is a postmortem instead of a decision.
I’m discussing Amplitude with companies every day, and the change I’ve been most glad to see is that this particular queue keeps getting shorter. More and more, the person with the question is also the person who gets the answer. Here's what that looks like in practice, and where I’d tell a client to start.
What changes when anyone on your team can query the data
The Global Agent is where most teams feel the difference first. Rather than building a chart, filtering it three times, and exporting the result into a deck, you ask the question the way you'd ask a colleague standing next to you. Think something like “why did activation drop for mobile users last week?” Then, Amplitude goes and investigates, builds the visual with related data, and hands back an answer with its reasoning attached. It can act on what it finds, too, spinning up a cohort or launching an experiment right out of the same conversation.
Two things make this workable for people who've never seen your schema. The agent carries context as you go, so a question you ask now gets connected to something it flagged several turns ago. And it translates vocabulary, which matters more than it sounds like it should. When a marketer asks to break a funnel down by "account type" and the property is actually named account status, the agent finds the right one instead of shrugging.
That same access travels through Amplitude's MCP support, which brings real charts and data into Slack, Notion, Figma, GitHub, and Claude. If your product managers are already asking you data questions in Slack instead of opening Amplitude, consider that your signal to turn it on.
How Amplitude’s AI agents monitor dashboards and session replays
The Global Agent answers when you ask. The four Specialized Agents run on a schedule and come find you, through an agent inbox, over email, or in Slack.
- The Dashboard Agent watches a dashboard you've already built, flags the movement that matters, and runs the root-cause analysis without anyone asking it to.
- The Session Replay Agent looks for friction patterns, including rage clicks (someone clicking the same thing over and over because nothing is happening) and dead clicks (someone clicking something that looks interactive but isn't). It hands back a curated playlist with the specific moments behind each finding, so you can go verify. This skips those classic “watch parties” of years past and puts AI to work reviewing thousands of sessions.
- The Customer Feedback Agent reads through survey responses and support tickets and connects those themes back to what people actually did in the product.
- The Web Experimentation Agent designs tests, launches them, and calls the winner.
Then, Automated Insights handles the follow-up question. When a number moves, it does the digging an analyst would do by hand, checking recent releases, campaign flights, and audience shifts, then comes back with a written explanation and a recommended next step. Your analytics team still owns the hard problems, which is exactly where you want them.
How event governance makes self-service analytics safe
Of course, opening analytics up to more people only works if the data underneath holds up. This is the part of the platform I find teams underuse most, and it's the one I’d nudge you toward first if you're nervous about handing the keys over.
Start with your project health score
Amplitude scores each project's health and gives you a working to-do list to improve it: assign event owners, add descriptions, set naming conventions. It turns governance from a thing you know you should get around to into a queue you can work through.
Let AI write the documentation
When you click into an undocumented event, Amplitude proposes a description inferred from how that event behaves. In practice, they come back accurate enough to review and approve in a sitting.
Catch drift before it reaches an analysis
Schema validation flags unexpected or missing properties per event, so instrumentation drift shows up while it's still a small fix. And when two developers ship start_trial and Start Trial, you can select both and merge them retroactively from the interface. Naming conventions can be enforced at the project level, separately for events and properties.
One taxonomy principle carries over no matter what you're running. Name your events for actions rather than interface elements (think "item added to cart" instead of "clicked CTA button") and reuse the same event anywhere the behavior is conceptually the same. You can always differentiate later with properties.
What item-level attribution tells you about product discovery
Another great thing about Amplitude is that a handful of analyses that used to start with a scoping conversation are now just configuration.
- Pathfinder answers what happened between two known points, so you can ask which five paths users most commonly took between starting a trial and upgrading.
- Persisted properties carry an event property value forward across a user's journey, across sessions and devices, with rules you set for whether the first or most recent value wins and how long it sticks around. Item-level attribution builds on that, and ecommerce teams tend to feel this one immediately.
| EXAMPLE Say you've got a $34 cart holding an $18 steak someone found through search, $10 of avocados from a popular-products module, and $6 of apples off a recommendation carousel. If you attribute at the order level, one discovery surface takes credit for the whole basket, which misprices every merchandising decision that follows. Item-level attribution instead splits the credit out to the surface that introduced each item. |
- Attribution modeling gets similar treatment. You can drop the same conversion event into a table three times and apply a different model to each column, so last touch, first touch, and data-driven credit sit side by side in one view.
Turning a behavioral finding into a live experiment
Every answer in the sections above lands somewhere: screenshotted into a deck for a presentation later in a week or dropped into an email update. But it may take people a few days to act on. The distance between knowing something and testing it is still measured in sprints, so by the time the change goes live, the context around it has moved.
With Zoning Insights, that distance gets a lot shorter. When you draw a box around any section of a live page, you get click-through rate, exposure, scroll reach, and revenue per click for that zone. Then, you can pull the session replays behind it and build a cohort straight from what you find. What you end up with is a ranked list of the places on the page where a change would move something.
Once you know what to test, Web Experimentation is where those hypotheses go, and Amplitude recently released some cool features on that front. Mutex groups keep overlapping tests from contaminating each other, while holdout groups let you measure the cumulative lift of everything you've shipped. And dependencies let you gate one test on another test's outcome, so a sequence of related tests can run without a person shepherding each handoff.
Where to start with Amplitude
If I had to pick three things to roll out this quarter when it comes to Amplitude, I’d start with:
- Make the Global Agent the default way people ask questions
- Turn on MCP so those questions get answered where your team already works
- Run Zoning Insights on your highest-traffic page for experimentation
Adswerve has more Amplitude-certified experts in the US than any other partner. The work we find most valuable sits where these capabilities meet a real measurement strategy, which usually means getting the event taxonomy solid enough that self-service holds up and connecting what you learn about behavior back to the media decisions that drove it. If you've recently started using Amplitude, we'd love to talk through where to begin. Reach out when you're ready.