Building Dashboards for Operational AI Oversight
A dashboard is only useful if it is honest about what it does not yet know. For systems running across devices with unreliable connectivity that is not a minor caveat but the central design problem, because a significant share of what the dashboard is supposed to show has not arrived yet and will not for hours or days.
Consider a dashboard tracking clinical activity across a network of rural health posts. A clean total on screen looks complete and usually is not, since some of those clinics only sync when a device reaches a signal, which might happen once a day or once a week. A dashboard that presents what has arrived so far as though it were the full picture understates reality, and more seriously, it leaves the person reading it unable to distinguish between nothing happening at a location and nothing having been heard from that location yet.
The fix is not complicated, though it does have to be deliberate. Alongside any number, the dashboard needs to show how much of the expected data has actually come in, so that a low figure can be read correctly as either genuinely low or simply incomplete. Without that, the dashboard is not giving people less information than they need, it is giving them wrong information in a form that looks reliable.
The same issue distorts comparisons between locations. A site with strong, constant connectivity will always appear busier than one that syncs rarely, even when the two are doing roughly the same amount of work, and left uncorrected that pattern leads to real decisions made on a false signal, such as assuming a quiet clinic needs less support when it is simply less visible. Showing sync frequency alongside activity data allows someone to separate a genuine difference in workload from a difference in how often a place manages to check in.
Order matters as much as completeness. Data from intermittent devices does not arrive in the order it was generated, and a record from three days ago can sync after records from yesterday have already landed. Every dashboard we build separates when something happened from when the system found out about it, because charts built on the second of those keep shifting and rewriting themselves in ways that make a trend almost impossible to read.
A district health office or a national response team can only act on what the dashboard tells them. Concealing the gaps does not make the picture cleaner, it makes the resulting mistake harder to catch until it has already been made.