A treasury team once walked me through a programme dashboard with thirty-odd charts on it. Spend by quarter, beneficiaries by district, a map shaded five ways. The work was clean. The team was proud of it, and they had every right to be.
Then a director asked the only question that mattered: did the spending change anything? The room went quiet. Not because the data was missing — the data was everywhere — but because no chart on the screen had been built to answer that question. The dashboard described the programme. It could not judge it.
That silence is the most common failure in analytics, and it is not a technical one.
The gap is structural, not technical
Most analytics functions are very good at producing insight and very bad at producing impact. The instinct is to blame the tooling, or the data quality, or the model. But the gap sits somewhere more boring and more fixable: in the chain between knowing something and doing something about it.
A logic model — the staple of programme evaluation — makes that chain explicit. Inputs lead to activities, activities to outputs, outputs to outcomes, outcomes to impact. Read it left to right and you can see exactly where analytics work tends to stop. Teams are rewarded for outputs: the dashboard shipped, the report delivered, the model deployed. Impact lives two links further down the chain, and almost every time, the chain breaks at the same join — the leap from Insight to Action.
This is why a beautiful analysis that no one operationalises is not a neutral artefact. It consumed analyst time, compute, and the attention of the people who reviewed it, and it changed no decision. It is a cost wearing the costume of a capability.
Insight that does not change a decision is the most expensive kind of trivia.
Working backwards from the decision
The fix is not more analysis. It is to design the analysis backwards — from the decision it is meant to inform, not forward from the data you happen to hold.
For your most important metric, ask three questions before a single chart is drawn:
- Who acts on this? Name the person and the seat, not "the leadership". A number with no owner has no action attached to it.
- When do they act? Tie the metric to a moment already on the calendar — the budget cycle, the board pack, the quarterly review. An insight that arrives after the decision is made is decoration.
- What changes as a result? If the number is high, what do they do; if it is low, what do they do differently? If both answers are the same, the metric is not decision-relevant.
If you cannot name the action, you have not found an insight. You have found a fact. The distinction is the whole job.
This is also the cleanest test of an analytics roadmap. Walk the list of dashboards and models a team maintains and strike out every one that cannot survive those three questions. What remains is the function's actual value. In most institutions, the list gets shorter than anyone is comfortable admitting — and the budget that was funding the struck-out work is suddenly available for the work that moves something.
What it looks like when the chain holds
The counter-example is worth holding in mind, because it proves the gap is bridgeable, not inevitable.
When Togo needed to move emergency cash to its poorest citizens in 2020, the country had no income register to work from — the kind of missing-data problem that usually ends a programme before it starts. Instead of stopping at the gap, the Novissi programme fused what it did have: mobile-phone usage patterns and satellite imagery, run through a model that ranked the poorest cantons and, within them, the most likely poorest individuals. The insight was a ranking. But the ranking was built to trigger an action — a mobile-money transfer — to a named recipient, on a defined timeline, with a measurable result at the other end.
Cash transfers targeted by fusing mobile-phone and satellite data where no income register existed — insight built to trigger an action, not a report.
Source: World Bank, 2021
Notice what made it work. The analysis was not more sophisticated than the treasury dashboard I opened with — if anything, less so. What it had was a decision wired into its design from the first line. Every link in the logic model was accounted for: the data was an input, the model an activity, the ranking an output, the transfer the action, and a household with cash the outcome. Nothing was produced that did not move something downstream. That discipline — not the algorithm — is what separates an analytics function that produces impact from one that produces charts.
The Togo case is unusually visible because the stakes were unusually high. But the principle scales down to the most routine quarterly review. The question a leader should be able to ask of any dashboard is the one that silenced the room: did this change a decision? If the honest answer is no, the work is not finished — it has not yet started.
This is how every Tensō engagement is built — from the decision backwards, closing on a metric that moved. If your analytics function produces insight but struggles to produce impact, that is the gap we close.