Reporting for whom, to decide what
Most reporting exists because someone once asked for it. It persists because removing things is harder than adding them, and it accumulates until an organisation produces far more reporting than anyone reads.
The corrective question is the one from the first lesson, applied to reporting rather than to analysis: who is this for, and what decision does it inform?
A report built for a defined audience and a defined decision has a natural structure, a natural cadence and a natural length. One built without them contains everything anyone might want, arrives on a schedule nobody chose, and is skimmed.
The practical exercise is to take an existing reporting set and ask of every element which decision it supports and who takes it. The proportion that survives is usually small, and the exercise is uncomfortable because most of it was somebody's request.
Cadence
The rate of reporting should follow the rate of decisions.
Daily suits operational monitoring where something might require action today: payment acceptance, system availability, alert queues, live incident indicators.
Weekly suits operational management, campaign performance and anything reviewed in a weekly cycle.
Monthly suits commercial performance, which is the rhythm most operator decisions actually follow.
Quarterly suits strategic measures: cohort quality, concentration, market contribution, structural indicators.
Ad hoc suits everything driven by a specific question.
Reporting more frequently than decisions are made produces a specific harm. A metric reviewed monthly but reported daily presents twenty-nine opportunities to react to random variation, and in a business with this much natural variance the reactions will be to noise.
The related discipline is that measures with long feedback loops should not be reported on short cycles at all. Cohort contribution at twelve months does not move daily, and reporting it daily invites treating movement in a partially formed figure as information.
What belongs in a reporting set
A structure that works for most operators.
A small operational set, reported frequently, covering things that might require action today.
A management set, reported weekly or monthly, covering performance against expectation with enough segmentation to locate a problem.
A strategic set, reported quarterly, covering the measures identified in the metrics lesson as important and usually absent: cohort quality, value concentration, contribution by market, exposure concentration, protective coverage.
Exception reporting, which surfaces things outside expected ranges rather than reporting everything and leaving the reader to find them.
Self-service access for people who want to explore, which reduces ad hoc requests and does not substitute for curated reporting, since a person given a query tool and no guidance produces their own definitions.
The principle throughout is that a short set read carefully beats a comprehensive set skimmed, and that the difference between them is what gets removed.
Visual choices and honest representation
Presentation carries interpretation, and several common choices misrepresent this data specifically.
Bar charts of averages are the standard default and are misleading for skewed measures, since the bar represents a value few observations are near. Adding a distribution, a median marker or a box representation makes it honest.
Truncated axes exaggerate change and are sometimes appropriate for showing small movements in a large base. They should be labelled clearly, and a truncated axis presented without indication is misleading whether or not it was intended to be.
Dual axes invite spurious correlation by allowing two unrelated series to appear to move together, and they should be avoided unless the relationship is genuinely being examined.
Pie charts are poor at conveying proportions beyond a few categories and are particularly bad for the long-tail distributions this industry produces.
Line charts across periods of different length distort trend, which happens with monthly data where months differ in days and in weekend count.
Aggregation that hides the tail presents a summary of a distribution as though it were the distribution.
Colour used to indicate performance carries a judgement that may not be warranted, and red on a figure that is within normal variation prompts a response to noise.
The general guidance is to represent the distribution rather than a summary of it wherever the distribution matters, which in this industry is most of the time, and to make any choice that emphasises or de-emphasises a movement visible to the reader.
The narrative
The element that turns figures into decisions, and the one most often omitted.
A chart shows a movement. Whether it matters, what caused it, what is uncertain about it and what should be done are interpretive questions. If the analyst does not answer them the reader will, usually with less context and more confidence.
A useful narrative states what changed, in plain terms. Whether it matters, given normal variation. What appears to have caused it, with appropriate hedging. What is uncertain. What would resolve the uncertainty. And what decision it bears on.
That is a short paragraph, it takes minutes, and it is the difference between a report that informs and a report that is circulated.
The related discipline is being willing to say that a movement is not significant. Reporting that treats every fluctuation as meaningful trains readers to react to noise, and reporting that identifies which movements matter is considerably more useful, though it requires the analyst to have a view.
Communicating to different audiences
The same finding needs different presentation depending on who receives it.
Operational teams need specificity and immediacy: what is wrong, where, and what to do. Detail is welcome and context can be assumed.
Managers need performance against expectation with enough segmentation to act, and interpretation rather than raw figures.
Executives need the small number of things that bear on decisions they will take, with the analysis behind them available and not presented. The most common failure with this audience is volume, and the second most common is burying the finding in the methodology.
Boards need strategic measures, exception reporting and honest statements of what is uncertain, and the governance material in the Leadership course indicates what they should be asking for that they usually are not.
Regulators, where reporting is required, need accuracy and completeness above all, since a submission that later proves incomplete is worse than one that was unfavourable.
The general adaptation is that seniority should reduce volume and increase interpretation, not the reverse, and that the instinct to include more when presenting upwards is precisely wrong.
Removing reporting
The discipline that keeps a reporting set useful, and the one nobody performs.
Reporting accumulates because adding is easy and removing requires telling someone their report is not being read.
The practices that work: usage measurement, since most tools record who opened what and the answer is frequently nobody; periodic review with an explicit question about which decision each report supports; sunset dates on reports created for specific purposes; and an owner for each report who is accountable for whether it is still needed.
The argument for making the effort is that attention is the scarce resource. A person facing forty metrics attends properly to none, and the important ones are indistinguishable from the rest. Removing thirty of them is an improvement to the ten that remain.
What good reporting looks like
To close, the characteristics.
It is short enough to be read by the people it is for.
It is built for decisions that someone actually takes.
Its cadence matches the rhythm of those decisions.
It represents the distribution rather than summarising it away.
It carries a narrative that says what changed, whether it matters and what it bears on.
It flags exceptions rather than requiring the reader to find them.
It is honest about uncertainty, and says when a movement is not meaningful.
It is reviewed and pruned, so that what remains has earned its place.
And it changes something, which is the only test that ultimately matters and the one worth applying annually to every report an operator produces.
Self-service and its limits
A structural question most operators face, since demand for analysis exceeds the capacity of any analytics function.
Self-service tooling lets people answer their own questions, which reduces request volume and increases the speed at which questions get answered.
Its limits are specific.
Definitions drift unless enforced in a semantic layer. A person building their own query will construct their own version of active players, and the divergence described in the metrics lesson multiplies.
Causal errors proliferate. Someone comparing groups that formed themselves will draw a causal conclusion, and self-service tooling makes that easy and fast.
Skew traps catch people. A tool that returns an average will return one, and a user without the distributional habits described in this course will use it.
Context is missing. A figure retrieved without knowledge of a data quality issue, a definitional caveat or a period anomaly will be used as though it were clean.
The arrangements that work provide self-service within guardrails: metrics defined centrally and enforced in the layer, curated datasets rather than raw tables, documentation attached to the data rather than held separately, and analysts available to consult rather than only to produce.
The complementary discipline is that the analytics function should watch what people are querying, since it indicates which questions matter and where curated reporting is missing.
Common reporting failures
A catalogue, drawn from the sections above.
Built for nobody, with no defined audience or decision.
Cadence mismatched to the decision rhythm, generating reactions to noise.
Volume increasing with seniority, when it should decrease.
Averages without distributions, which for this data misrepresents almost everything.
No narrative, leaving interpretation to the reader.
Every movement treated as significant, which trains readers to react to variance.
Truncated axes and dual axes unlabelled.
Metrics without definitions attached, so readers apply their own understanding.
Nothing ever removed, so attention is diluted across an accumulating set.
No usage measurement, so nobody knows what is read.
Exception reporting absent, requiring readers to find the problems themselves.
Uncertainty omitted, so provisional figures are treated as settled.
Each is straightforward to correct and most persist because reporting is produced under time pressure and reviewed rarely.
Writing up an analysis
Distinct from routine reporting, and the format in which most consequential findings are delivered.
The structure that works.
The finding first. One or two sentences stating what was established. Readers who go no further should have the answer.
What it means for the decision. Immediately after, since that is why the analysis was commissioned.
The evidence, summarised, with the strongest element rather than all of it.
The uncertainty, stated plainly. What is not established, what assumptions were made and what would change the conclusion.
The recommendation, if one is warranted, distinguished clearly from the finding so that a reader can accept the second while disputing the first.
The detail, available and separate, for anyone who wants it.
The common failures are burying the finding beneath the methodology, presenting the analysis in the order it was conducted rather than the order the reader needs, omitting the uncertainty because it weakens the case, and merging the finding with the recommendation so that disagreeing with one appears to require disputing the other.
The related discipline is length. A finding that requires twenty pages to convey has usually not been distilled, and the distillation is part of the analytical work rather than a presentational afterthought.
When the finding is unwelcome
A situation every analyst encounters and the one that determines the function's value.
Analysis frequently contradicts what someone hoped. A campaign that did nothing. A channel delivering poor customers. A change that made things worse. A metric that has been reported wrongly.
The approaches that work.
Check it properly first. An unwelcome finding will be scrutinised harder than a welcome one, and it should be able to withstand that.
Deliver it to the person who commissioned it before it circulates, so they are not surprised in a meeting.
Separate the finding from the blame. A campaign that did nothing is information about the campaign, and presenting it as such rather than as a judgement on whoever ran it determines whether it is accepted.
Offer the next question. A finding that something did not work is more useful accompanied by what might.
Be precise about what was and was not established, since an overstated unwelcome finding is easily dismissed on the overstatement.
Do not soften it into ambiguity. A finding hedged until it says nothing has protected the analyst and failed the organisation.
The organisational conditions that determine whether this is possible are covered in the Leadership course. Where they are absent, an analyst can still deliver honestly and should understand that they are doing so against the grain rather than assuming the problem is their communication.
Reporting on protective measures
A category that belongs in an operator's reporting set and is frequently absent, drawing on the compliance material in the first batch of courses.
The measures worth reporting regularly.
Interaction volume and outcome, meaning how many customers were contacted about their play and what changed afterwards. The second half is the one regulators examine and the one operators frequently do not record.
Time from indicator to action, which measures whether detection produces timely response.
Tool adoption, meaning what proportion of customers have set limits, which indicates whether the tooling is findable.
Self-exclusion volumes and processing time.
Suppression verification, meaning tested evidence that excluded customers do not receive marketing.
Value concentration, which as noted throughout carries a protective reading alongside its commercial one.
Affordability assessment coverage, meaning what proportion of customers above defined thresholds have actually been assessed.
Alert queue status, since a queue larger than the team can work is a documented failure to act.
Reporting these alongside commercial measures does two things. It puts the information in front of people who would otherwise not see it, and it establishes a record that the operator was monitoring rather than assuming.
The presentational note is that these should not be targeted in either direction. Interaction volume driven upward produces automated contacts that achieve nothing; driven downward produces silence. They should be tracked, investigated when they move, and left untargeted.
An annual reporting review
To close, an exercise worth running once a year.
List every recurring report the analytics function produces.
For each, identify the audience and the decision it supports. Anything without both is a candidate for removal.
Check usage, since most tools record it and the answer is frequently that nobody opens it.
Ask the recipients what they actually use, which produces a different answer from what they say they need.
Identify the gaps, meaning decisions being taken without supporting reporting, which is usually a longer list than the removal candidates.
Check the definitions used in each, against the dictionary.
Check the cadence against the decision rhythm.
Remove ruthlessly, and communicate the removals so that anyone who genuinely needed something can say so.
Operators running this exercise typically remove a substantial proportion of their reporting and find that almost nobody notices, which is itself the finding. The capacity released goes to the gaps identified in the fifth step, which are the decisions currently being taken without evidence.
That reallocation, from producing unread reporting to supporting unsupported decisions, is available annually at the cost of an uncomfortable afternoon.