Edit dashboards with AI: change a live report by asking for it

Written by Uptimal, the team that builds and operates Upbuild and runs marketing analytics for multinational clients across the globe. Everything below reflects how the product actually works, including the parts that are deliberately boring.

The demo everyone shows is an AI building a dashboard. The demo nobody shows is an AI editing one that four people rely on, because that is where "the model hallucinated" stops being funny. Upbuild Enterprise treats dashboard editing as a change-management problem that happens to have a conversational interface.

An agent per dashboard, not a general chatbot

In the app, every dashboard carries its own assistant thread, and so does every data stream underneath it (a data stream is one saved platform query: spend by campaign from Google Ads, say). You ask in plain language: swap the spark line to an eight-week ROAS trend, add Snapchat to the blend, split the CPA tile by market. The agent knows the page it is standing on, its queries, and what your organization lets it reach.

Its answer is not applied. It is drafted, as a new version of the page or the stream, exactly the way a colleague would open a branch. The thread survives reloads because history is stored server side and re-read on every question; a chat that loses its memory teaches people to stop asking.

The two guardrails that make it safe

First, the live test. A proposed query change runs against your real accounts before it can be published, and the proof of that run travels with the proposal. A filter that returns nothing, a dimension the platform rejects, a window that empties the answer: all of it surfaces in draft, not on the live page at 9 am.

Second, a rule the product calls never-less: a change cannot silently remove a field, metric or dimension the live dashboard currently reads, and the guard checks the data actually arriving, not just the query text, because a column can stop arriving while the definition still names it. Steering is free (dates, windows, filters, ordering can change on a live stream); shrinking what the page depends on is blocked unless a person does it deliberately.

Publish is human, rollback is boring

The agent proposes, previews and explains; a person with edit rights publishes. Every publish is an immutable version, fifty are kept, and restoring an older one is an explicit action that itself creates a version, so the audit trail never lies. When a platform refuses a query, the app reduces the refusal to its one useful sentence instead of pasting raw JSON at you, and asks the model for a way forward rather than showing a failed attempt as if it had happened.

Why this beats regenerating from scratch

The naive alternative, "just ask the AI to rebuild the dashboard", discards every guarantee the old page had earned: its history, its tested queries, its share links, its provenance. Editing as versioned proposals keeps the page continuous while letting the conversation move fast. It is the difference between an AI-built dashboard and an AI-run reporting practice.

Frequently asked questions