Why We Didn't Build a Dashboard

The interface for AI work assumes a problem that no longer exists.

A few months into building Righthand, we sat down to design the screen our customers would open every morning. It was going to be a dashboard. Of course it was. Every B2B product I have ever loved has a dashboard. Stripe has one. Linear has one. Datadog is basically a religion built around them. When you build software for a company, the dashboard is the room you build first and the room your customer lives in.

We sketched it. Active agents on the left. Tasks in flight in the middle. A queue of items waiting for human review on the right. Filters across the top. A little graph somewhere to make it feel serious. It would be beautiful, dense, the place a manager opened their laptop to see what their AI team had been up to.

Then somebody on the team, I forget who, asked the question that killed it: "When would they actually look at this?"

We could not answer. Not honestly.

The room nobody walks into

The reason we could not answer is that we had built our entire mental model of the product around a habit that did not exist yet. Managers do not have a morning routine for checking on their AI agents because, until very recently, they did not have any AI agents to check on. We were designing a control room for an operation that runs at a speed and scale where, by the time you walk into the room, the work is done.

You can argue that the dashboard exists for oversight, not control. Fine. But oversight of what, exactly? If an agent has done four hundred things since you last logged in, you are not going to read four hundred rows. You will read a summary. And if you are going to read a summary, the dashboard is just a worse version of a message.

This is the moment the question got interesting. Not "what should the dashboard look like," but "why are we building a dashboard at all."

What dashboards were for

Dashboards are very good at one thing: packing a lot of information into a small space so a human can scan it, form a judgment, and decide what to do next. Information density for human consumption. That is the entire job.

For the last thirty years, every piece of business software has been built on a quiet assumption: the human is the end user. Not the customer, not the operator, the actual end user of the interface. The human is the entity that reads the screen, holds the context, makes the call, and takes the action. Every design decision flows from there. Tables are sortable because humans sort. Filters exist because humans cannot hold a thousand rows in their head. Charts exist because humans see patterns in shapes faster than in numbers. The dashboard is a love letter to human cognition.

We built it this way because we had no choice. There was exactly one kind of entity on the planet capable of looking at a screen, understanding what it meant, and doing something about it. So we optimized for that entity. Thirty years of design craft, poured into making screens that human brains could metabolize. It was the right answer to the question we were asking.

The question has changed.

The math nobody wants to do out loud

It takes a human being about twenty years to grow up enough to do paid work well. Twenty years of food, shelter, school, mistakes, and the slow accumulation of judgment before they can be trusted with something a stranger will pay for. That is not a knock on humans. It is what the path looks like.

An agent gets to the same point in three minutes and five clicks.

I am not saying that to be provocative, and I am not cheerleading it. It is uncomfortable to sit with, and I think we should sit with it anyway, because the math has consequences whether we like them or not. A company with five humans and a thousand agents is not a science fiction premise. It is a configuration some of our customers are already approaching, and a configuration the next five years will normalize. You do not have to want this future to notice it is the direction the arrow is pointing.

Once you accept that this is where the numbers go, a question lands in your lap that you cannot put down. What, in that world, is a human uniquely valuable for?

That is not rhetorical. It is the question. And until you have a real answer, you cannot design the interface, because the interface is downstream of the answer.

What humans were valuable for, and what they are valuable for now

For most of the history of work, humans were valuable because they could process information and act on it. A bookkeeper read the ledger and noticed the number that did not belong. A buyer read the sales reports and decided what to reorder. A manager read the dashboard and decided which fire to put out first. The valuable thing the human did was the perception-and-decision loop. Software's job was to feed that loop as efficiently as possible. Hence the dashboard.

That loop is no longer a defensibly human activity. Agents can read the ledger, the sales reports, and the dashboard. They can act on them. They can do it in parallel and at three in the morning, and they do not tire on the fourth report. So if the perception-and-decision loop is where you locate your value as a human, you are standing in the one place the machines are already strongest. That is not where your leverage is anymore.

What is left? What is the thing a human still does that an agent does not?

I think it is taste.

Rock and roll

Imagine you trained a music generator on every song recorded before 1940. Big band, swing, blues, gospel, classical, country, the whole archive. Then you asked it to generate the next great American genre.

Rock and roll is not in that distribution. It is not a small extrapolation from what came before. It is a strange, loud, vaguely scandalous mutation that drew on the blues and country and gospel but recombined them in a way the training data did not predict. The model could, in principle, generate it. If you generated a million songs, one of them might sound like Chuck Berry. It would also generate nine hundred and ninety-nine thousand, nine hundred and ninety-nine other things, most of which would sound like competent extensions of the 1939 charts.

Here is the question that matters. How would you know which one was the Chuck Berry?

Not by looking at the waveform. Not by counting the notes. Not by running it through any statistical test, because by definition the thing you are looking for is the one that does not look like the others. You would only know by hearing it and feeling something. You would know because some human, somewhere, said "that one. Play that one again."

That is taste. The thing humans actually did when rock and roll showed up was not generate it. The genre was sitting there in the latent space of American music, waiting. What humans did was pick it. They listened to a million things and pointed at the one that mattered. They turned it into a movement, a market, an era.

In a world where generation is cheap, selection is the work. The scarce resource is not the ability to produce options. It is the judgment to recognize which option is worth keeping. Taste is the human contribution when machines can produce anything.

This is not a poetic claim. It is a practical one. The customers I talk to are not asking their agents to produce more outputs. They are drowning in outputs. They are asking how to look at the outputs and decide which ones are right, which ones are wrong, which ones are interesting, which ones to send, which ones to throw away. The bottleneck has moved. It used to be production. It is now discernment.

What we built instead

Once we said it out loud, the dashboard stopped making sense as the front door of the product. A dashboard is an information-density tool for a human whose job is to perceive and decide. The perception part is increasingly the agent's job. The deciding part, the part that needs taste, does not happen well in a grid of rows. It happens in conversation. It happens in the moment a human looks at a specific thing the agent produced and says "yes, that, more like that," or "no, not that, here is why."

So we built into the channels people already use to make those decisions. Slack. Email. The places where work actually gets discussed, where context already lives, where the manager is already exercising taste on the things their human team produces. The agent shows up there, not in a separate room you have to remember to visit. It brings the work to you, in the shape of a message you can react to with the same instincts you use for everyone else on the team.

This is not a design flourish. It is the consequence of the argument. If the human contribution is taste, the interface should put humans in the seat where they exercise taste, and get out of the way. Taste is exercised on specific things, not on aggregates. You do not develop it by scrolling through a dashboard. You develop it by saying yes and no to one thing at a time, with the context of that one thing fully loaded into your head.

The product decision is the proof of the thesis, not the point of it. We are not arguing that messages are the right interface forever. We are arguing that whatever the right interface is, it has to be designed around what humans are actually good for in this new arrangement. Dashboards are designed around what humans used to be good for.

The honest part

I do not know what the long-term interface for human-and-agent work looks like. Nobody does, and anyone who tells you they do is selling something. The space is too new. The shape of the work is still forming. Our current answer is a starting point, not a destination. We will be wrong about parts of it, we will change our minds, and the product in two years will look different from the product today.

What I am more confident about is the negative claim. The interface is not a dashboard. Not because dashboards are bad, they are not. They are a beautiful solution to a problem we used to have. But the problem they solve is the problem of a human processing information at the center of the work. That human is not at the center anymore. The interface has to follow.

If you are building anything for a team that is starting to operate with agents, the question I would sit with is not "what should our dashboard show." It is "what is my user uniquely good at now, and where do they actually do it." Build for that place. The dashboard will not be missed.

I do not know what the answer is. I just do not think it is dashboards.