Every analytics product I’ve used hands you the same thing on day one: a dashboard. Charts, counters, a date picker, and the unspoken assumption that you’ll work out what it all means. I built one too. It took me an embarrassingly long time to see that the dashboard wasn’t the feature — it was the homework.
What changed my mind was watching people use ours. They’d open it, scan the panels, and then just sit there. The data was right in front of them and they still didn’t have their answer. That gap is the thing large language models actually close, and it isn’t the part most people expect.
The dashboard makes you do the work
A dashboard answers exactly one question: “show me everything.” Almost nobody arrives with that question.
People show up with something specific and human — did this change, who’s new, what did I miss while I was gone. A dashboard makes them translate that into filters and date ranges, read the result, then translate it back into an answer. Power users do this so fast they forget it’s happening. Everyone else feels the tax.
Most good dashboard design is an effort to lower that tax. A lot of Nielsen Norman Group’s research on dashboard comprehension reads as a catalog of techniques for helping people decode charts with less effort. That work is genuinely useful. But it’s optimizing the workaround — a chart is still an answer you have to decode yourself.
What we built
I run IGDetective, a tool that tracks public activity on Instagram accounts: recent follows and unfollows, who an account engages with, that sort of thing. The first version was a dashboard, like everyone else’s. People stared at panels of follower changes trying to assemble the story themselves.
So we added a feature we call Gossip Chat. You ask it, in plain words, “what has this account been up to this week?” and it answers in a sentence or two. Same data underneath. A completely different relationship to it.
What surprised me wasn’t the technology. It was the behavior change. The same people who had ignored the dashboard used the chat constantly.
Conversation collapses the distance to an answer
With a dashboard, the path runs: load, scan, hold four numbers in your head, infer a story. With a chat layer, the model holds the numbers and hands you the story. One step instead of four.
This is the real unlock of language models for consumer analytics, and I think it usually gets mis-stated. The win isn’t smarter analysis. The win is deleting the interpretation step that dashboards quietly charge every non-expert user.
A concrete version: someone wants to know if an account has gone quiet. On a dashboard, that’s cross-referencing a few panels and doing the math. In chat, it’s one question and a one-line answer. Nothing got smarter. The distance to the answer just went to zero.
If your users are analysts, keep the dashboard — they want the raw numbers and they move fast through them. If your users are ordinary people, a sentence beats a chart most of the time.
On a free tier, cost and latency are features
Here’s the part the “just add an AI feature” pitches skip. The moment a model sits behind a free consumer feature, inference cost and response time stop being backend concerns and turn into product decisions.
A dashboard is close to free to serve and renders instantly. A model that summarizes activity costs money on every single call and takes seconds to think. Across a free tier, careless choices there get expensive quickly, and slow answers quietly kill the feeling of a conversation.
We ended up treating model selection as a product call rather than a benchmark contest: smaller models tuned to the specific job, aggressive caching whenever the underlying data hasn’t moved, and an interface where a couple of seconds of thinking reads as deliberate instead of broken. If you’re putting AI into a consumer product, your unit economics and your latency budget are part of the user experience — whether or not you decided to treat them that way.
Ground it, or it will lie to you with confidence
The fastest way to wreck a tool like this is to let the model guess.
People forgive a dashboard that shows nothing; the blank is honest. They do not forgive a chatbot that states something false with total confidence. Models doing exactly that — producing plausible, wrong details in a fluent voice — is a well-documented failure mode the field calls hallucination. In an analytics product, a single fabricated “fact” about real activity is worse than no answer at all.
Our rule is simple: the model can only talk about data we actually hold. Everything Gossip Chat says is grounded in the public signals already on the screen — the same information you’d see if you opened our anonymous Instagram Story viewer or scrolled a public profile yourself. It summarizes and connects what’s there. It is not allowed to reach past it, and when it doesn’t know, it says so.
That constraint isn’t a limitation we put up with. It’s the thing that makes the conversation trustworthy enough to replace the dashboard in the first place.
The honest trade-off
Conversation isn’t simply better than visualization, and I’d be selling you something if I claimed it was.
Charts still win when you need to scan many dimensions at once or compare a trend across twelve months. A sentence is a poor tool for “show me the shape of this over a year.” The real answer for most products is both: visualization to explore, conversation for the specific question someone actually arrived with.
What AI changes is the default. For years the dashboard was the front door because it was the only door we knew how to build. Now the front door can be a question, and the dashboard becomes the thing you open when the question needs a chart to answer it.
Where this goes
I think almost every consumer analytics product grows a conversational layer over the next few years, and most will botch the first attempt. They’ll bolt a chatbot onto a dashboard, let it hallucinate, ignore the cost curve, and conclude that AI didn’t work for them.
The teams that get it right will treat the conversation as the primary interface, ground it ruthlessly in real data, and budget for inference like any other core feature. They’ll stop asking “how do we add AI to our dashboard” and start asking “what question was the dashboard always a stand-in for?”
The dashboard was never the product. The answer was. For the first time, we can just hand people the answer.
Steven Duncan is the founder of IGDetective, an Instagram activity tracker. He writes about consumer analytics, product design, and applied AI.
