Algeria AgencyArtificial intelligence
A dashboard that says no
A dashboard earns its place through what it refuses to assert. Three refusals to insist on, and the hole never to fill in.
The instinct when commissioning a dashboard is to ask for more: more indicators, more sources, finer granularity. That is the opposite of what makes one useful, and most of the screens we take over suffer from excess rather than shortage.
A useful dashboard does three unwelcome things. It declares which of your sources governs when they contradict each other. It shows a forecast with its uncertainty, or no forecast at all. And it shows the share of your activity it does not measure, instead of filling it in.
This article describes those three refusals, why they are harder to obtain than a feature, and the attribution hole specific to a market where a large share of sales ends in cash and offline.
The problem is not a shortage of numbers
Almost nobody arrives saying "I have no data". People arrive with four tools producing it, disagreeing about the same month, and a monthly meeting spent debating which figure is right instead of what to do.
The gaps almost always have dull, explicable causes: attribution windows differ, one counts orders and another payments, a time zone shifts a whole day, one source de-duplicates and another does not. None of those is a bug; they are different definitions of the same word.
So the useful work begins by documenting those gaps rather than making them disappear. A dashboard showing one number where four tools give four has made a choice, and if that choice is written nowhere it will be contested at every meeting by somebody with another source open in front of them.
There is a methodological consequence: the first useful meeting is not a tooling meeting but a definitions meeting. What counts as a sale — the order, the payment, the delivery? Which date is it attached to? Three questions of that kind explain most of the gaps, and they are settled among yourselves, without a supplier, in an hour.
A declared source of truth
For each indicator, one source governs and the others are observations. That is not a quality ranking, it is a necessary convention: without it, no decision can be defended three weeks later.
The choice is made by use, not preference. For revenue it is the accounts, even if they arrive late. For campaign effectiveness it is the source with the same definition of a conversion on both sides. For stock it is the system that decrements, not the one that displays.
That document fits on a page and is the only deliverable of this kind of project that keeps its value when the tooling changes. We hand it over even when the rest of the project does not happen, because it represents most of the thinking and because it is yours.
A clarification that avoids a common misunderstanding: declaring a source of truth does not mean ceasing to display the others. The screen gains from showing both values when they diverge markedly, noting which one governs. Hiding the gap does not remove it, it moves it into the head of whoever has the other tool open and who will stop trusting the dashboard.
One screen per decision
The right design unit is not a theme but a decision. "Should we raise the budget on this product", "which orders will cause trouble this week", "where does the money that returns nothing go": each translates into three or four numbers and a comparison.
A screen of forty indicators is not a dashboard, it is a report, and a report is read once. The practical difference shows after a month: screens built around a decision get opened, screens built around a theme do not.
There is a simple test for whether a screen is well designed: ask the person meant to use it what they would do differently depending on what it shows. If they cannot answer, the screen is not for deciding, it is for informing — which is legitimate but is called a report and gets read once a month rather than daily.
A dashboard is read standing up
The internet market observatory published by ARPCE counts, for the second quarter of 2025, some 59.10 million internet subscriptions in Algeria, of which 88.71% are mobile and 11.29% fixed.
That has a direct consequence for design, not merely for compatibility. Real consultation happens standing up, between two meetings, on a six-inch screen — not seated at a large display with time to scroll a twelve-column table.
A screen designed for the large format and then "adapted" for mobile becomes a screen consulted once a month, when somebody is at their desk and remembers it. The constraint is taken at the drawing stage: three or four numbers, one comparison, no horizontal scrolling.
That constraint has a useful side effect: it forces the trade-offs. When there is only room for four numbers, the discussion about which to keep is the discussion that should have happened from the start. Screens designed without a space constraint never have it, which is why they grow until they are unreadable without anybody having taken a decision.
ARPCE, internet market observatory, second quarter of 2025
Alerts on deviation, not on thresholds
A threshold is a round number picked a year ago by somebody who may no longer be there. It fires when activity changes level without anything abnormal happening, and it stays silent when something abnormal happens inside the range.
A deviation alert compares the value with its own usual behaviour for that day of the week and that period. It needs some history and it has one property worth the difference: people stop ignoring it, because it does not fire every Monday.
An alert also needs a named recipient and an expected action. ‘Sales are down thirty percent’ sent to a distribution list produces nothing, because everybody assumes somebody else is looking. The same alert sent to one person, with the question to settle, produces an answer. It is less a tooling question than an organisational one, and it is where most alerting systems die.
The forecast and its uncertainty
A forecast is a claim about the future drawn from a past under no obligation to repeat itself. That is not an objection: it is a reason to show it with its range, the number of observations behind it, and the date of the last recalculation.
A single point with no interval is an opinion presented as a fact, and it will be read as a fact. The difference is not cosmetic: "between 40 and 90" and "65" lead to different decisions, and only the first is what the data says.
It is also the feature most readily removed for you, because a range is less impressive in a demo. Asking for it explicitly at scoping is the only way to have it.
It is also worth saying what a range is not. It does not measure commercial risk, it measures the spread of what past data allows you to assert. An event the history has never seen — a closure, a supply break, a competitor opening opposite — is in no interval. Saying so avoids the opposite error, which is treating the range as a guarantee.
One full cycle, minimum
Local seasonality does not resemble the patterns these tools were calibrated on. Ramadan shifts demand and the hour of purchase, the school year concentrates whole categories, holidays create short peaks, and summer empties some wilayas and fills others.
A model trained elsewhere reads those movements as anomalies and corrects them, which is precisely the wrong thing to do: they are the most regular movements of the year. So at least one full cycle of history is needed for a forecast to be honest here, and often two.
Below that, the correct position is to show no forecast. A screen saying "not enough history to conclude" is more useful than a confident number drawn from eleven observations, because the second will be believed and used to decide a stock order.
There is a partial workaround when history is missing, and it is worth knowing: borrow the shape without borrowing the level. If your sector has a known and documented seasonality, it can be used to say when demand will rise, without claiming to say by how much. It is less than a forecast and it is usable for planning stock, which is often the real question.
There is a common error to avoid in the other direction: treating each year as a repeat of the last. Ramadan moves about ten days each year in the Gregorian calendar, which is enough to slide a peak from one month to another over three or four years. A model learning seasonality on months rather than on the real calendar learns a drift rather than a cycle, and it gets it more wrong the more history it has.
The attribution hole
A significant share of activity ends offline and in cash. Somebody sees a post, travels, buys at the counter, and no tool connects those three events. That is not a configuration gap: no measurement of that journey exists.
There is, however, a permanent temptation to fill it with a model. A coefficient that "estimates" offline sales attributable to a campaign, applied uniformly, produces a plausible number and a smooth curve. It is the only figure on the screen that nothing measured, and once displayed it is indistinguishable from the rest.
The position we argue for is to draw the hole. What is measured is displayed, what is not is displayed as unmeasured and at its approximate size, and the total does not claim to reach a hundred percent. It sells less well and it is the only version you can decide on without lying to yourself.
Honest partial measurements exist and should be preferred to the model. A discount code specific to a campaign, a question asked at the counter, a dedicated phone number: each captures a fraction of the offline journey, and that fraction is measured rather than estimated. It is smaller than what a coefficient would display, and it has the property of not degrading when you look at it closely.
Broken series
A change of tracking tool, a rebuilt shop, a recreated ad account: each produces a curve restarting from zero. With no marker, that reads as a commercial collapse, and the reading always arrives at the worst moment, in a meeting, in front of somebody seeing the chart for the first time.
The correction is graphic and cheap: mark the break explicitly on the axis, with its date and cause. That marking has to be data in the dashboard, not an annotation somebody adds by hand and forgets on the third occurrence.
There is a more insidious variant of the break: the one that does not reset the curve but silently changes the definition. A tool that starts counting orders instead of payments produces a continuous and false series from a certain date, and nothing on screen flags it. That is why the sources document has to be dated and re-read, rather than written once.
What is not read gets removed
A month after go-live, look at what has been opened and what has not. Unread screens cost maintenance, degrade silently when a source changes, and dilute attention away from the ones that work.
Removal is a healthy and unpopular operation, because every screen was asked for by somebody. Doing it once, openly, beats accumulating for two years until nobody trusts any of it.
Removal happens more easily if it was announced at the start. Saying at go-live that an unread screen will be removed after a month turns deletion into a rule rather than a judgement, and nobody takes it personally. It is a sentence to place at scoping, and it saves a difficult conversation later.
The data stays exportable
A dashboard must not become the place your numbers are held. At any moment, the data feeding it should be exportable in an ordinary format, readable without the tool and without us.
Ask that before choosing, because it cannot be asked afterwards. It also has immediate practical value: whoever has to prepare a file for a bank or a lender needs the figures in a spreadsheet, not a screenshot.
There is a concrete check to run before signing: ask for an export of the current month and open it. If it arrives in seconds, in a format your accountant can read, the question is settled. If it has to go through the supplier's support desk, you have your answer — and you have it while you still have a choice.
What we do, and the hole we will not fill
We start from three decisions you make regularly without the data you would need, we audit the sources and document their gaps, we put one screen live, and we come back a month later to remove whatever is not being read. The sources document stays yours in every case.
What we will not build is the attribution model that closes the offline gap. It is technically simple, we know how to do it, and it would produce exactly what this kind of project is supposed to produce: a complete screen, a continuous curve, a total reaching a hundred percent. The problem is that one value on that screen would be an assumption rather than a measurement, and once displayed beside the others nothing would distinguish it. It would be used to settle a budget, then to justify the settlement, and the assumption would become a fact through repetition alone. We would rather deliver an incomplete, honest screen with the hole drawn at its size.
The counterpart is that you will see, permanently, the share of your activity nobody measures. That is uncomfortable and it is the point: that share is a fact about your market rather than a defect in your tooling, and knowing it changes how you read everything else on the screen. If you want a supplier who makes it disappear, they exist; we only ask that you know what disappeared was the measurement, not the uncertainty.
There is a practical consequence of this position better announced than discovered: your screens will not compare with your competitors’. A business displaying a complete attribution rate and a business displaying a hole are not talking about the same thing, and the second will look less impressive in a meeting. The difference is not in the results, it is in what each one accepts as measured.
Frequently asked questions
How much history does a forecast need?
At least one full seasonal cycle, often two. Below that the right answer is to show no forecast, and the screen should say so.
Can you connect our advertising to our in-store sales?
Partly, and honestly. We show what is measurable and mark the rest as unmeasured rather than estimating it.
Do we need to change our tracking tools?
Rarely. The work is about reconciliation and definitions, not tooling, unless a tool produces no usable data at all.
Who updates the screens?
They update themselves. What needs a person is the monthly check on gaps between sources, described in the sources document.
What if we do not like the numbers?
They stay displayed. A screen whose function is reassurance stops being usable for deciding, which was its only reason to exist.
How long for the first screen?
A few weeks, much of it on the source audit — the stage everyone wants to skip and the one that determines the value of the rest.
Where we come in
The same month read in four tools gives four different answers. The gap between them is your real subject, and no dashboard shows it.
- We audit the four sources before building a fifth one.
- We start from the decisions you take without a figure, not from the figures available.
- We leave visible the share of your business nobody measures.
The model that fills the offline gap is easy to produce and reassuring, and it manufactures what it claims to measure: we will deliver none.
Read next
A dashboard is read or it dies: the ten-minute review
Most dashboards are not bad, they are abandoned. What keeps one alive is a habit rather than a feature.Training a model: the three cases where it is worth it
Training changes the shape of answers, almost never their content. What you must supply, and the three cases where it is the right tool.The impact assessment, before connecting an assistant to your customer data
It is done before go-live, and that alone is what distinguishes it from a report. Afterwards it describes a risk that has already either materialised or not.
Let us talk about your project
A free audit, no commitment: we look at your online presence and tell you what is holding it back.