The Dashboard Layer, Eighteen Months On
A dashboard is not the end of a pipeline, it is the only part most people ever see. Which makes one design decision more important than every visual feature: the numbers on it are never written by a language model.
For most of the people affected by your data work, the dashboard is the product.
They will never see the pipeline that fed it, the checks that guarded it, or the eleven judgements someone made about how to read a supplier's export. They see a number, in a box, and they make a decision.
That asymmetry is worth sitting with, because it means the last layer carries the credibility of everything behind it — and it is usually the layer built last and thought about least.
The decision that matters most
Start with the one that is easy to miss and hard to reverse.
Numbers are never emitted by the language model.
When someone asks a question of a dashboard in plain language, the model does not answer it. The model produces a validated query plan. The server then computes the result deterministically, and every answer block carries its provenance, a CSV export of what it used, and the ability to be pinned onto the dashboard as a tile.
The distinction is everything. A language model that generates prose about your data will, eventually and confidently, generate a number that is not in your data. Not maliciously — that is simply what the technology does at the edges. Once one wrong figure reaches a board pack with your name on it, every other figure becomes a question.
Separating deciding what to compute from computing it removes that failure mode by construction rather than by care. The model is good at the first job. It is structurally unsuited to the second, and it is not asked to do it.
Q&A is for the person who cannot build anything
The commercially interesting part of that architecture is who it serves.
A Q&A page inside a dashboard lets a viewer — somebody with no licence, no training, and no interest in building anything — ask a follow-up question of a shared artifact and get a computed, sourced answer.
That is the difference between sending someone a chart they have to interpret and sending them something they can interrogate. The follow-up question is where the value has always been, and until recently it required a request, a queue, and three days.
Sessions are private by default. The owner can share one, and it becomes a replayable artifact others can read, comment on, and fork. Because every message stores its rendered blocks, a shared session reads exactly as it did when it ran.
And exports respect row-level security: a viewer clicking Export to CSV or PDF receives only their scoped data. The permission model does not stop at the edge of the screen.
Styles, and why they are not themes
Presentation runs on a Style System — six stock styles plus custom ones, built as token bundles rather than the older idea of a theme.
The difference is not cosmetic. A theme is a look you apply to a thing. A token bundle is a set of decisions — colour, type, spacing, density — that every artifact resolves against. Change the bundle and everything moves together, because nothing had a copy of the values.
It is the same principle as parameters in a pipeline: define once, reference everywhere, and never discover that four dashboards disagree because somebody set a colour by hand in March.
One artifact, several shapes
Content is not welded to its format. Pick a different one and it reshapes.
A document carries its own controls — length (brief, standard or full), surface (web or PDF), and shape (page or poster) — because a thing meant to be read is not a thing meant to be scanned on a wall, even when the underlying numbers are identical.
Exports follow the same logic rather than flattening everything into one file type. A dashboard prints client-side. Paged documents render server-side onto real pages. A published motion story exports to 1080p MP4. Posters go to PNG. And CSV comes out of the underlying data, for the person whose next step is a spreadsheet regardless of how beautiful the board was.
Publishing is the point
The reason to care about this layer at all is that it is where the work stops being yours.
Everything upstream — the connections, the recipes, the checks, the schedules — exists so that a number can be put in front of someone who will act on it. A pipeline nobody sees the output of is a hobby.
Which is why the design decisions worth arguing about here are not the chart types. They are: can a viewer ask a follow-up without asking you; does the answer they get carry its provenance; does their export respect what they are allowed to see; and is the number on the screen something a computer calculated rather than something a model composed. That last one is worth its own argument.
Get those four right and the visual library is a detail. Get them wrong and no amount of chart types will save the meeting.
