From decision to KPI: derive your metrics, don't collect them

The executive team meets on Tuesday. The monthly management report runs to 34 pages and 120 metrics; three of them get discussed. For the item that actually matters – a price increase in January, yes or no – the one number that would settle it is missing: contribution margin by customer group after this year's rise in material costs. The controller knows it could be built from the ERP. He never built it, because nobody asked. The report does, however, contain twelve metrics nobody has looked at in years.

Why collecting fails

Collecting KPIs is the natural approach, and it leads reliably to the same dead end. Three forces drive it. The ERP and the BI tool deliver what is technically easy. Managers ask for what they know from earlier reports. And once a metric is in the report it stays there, because removing it might annoy someone, whereas keeping it annoys no one.

Ossola-Haring, Schlageter and Schöning describe the result with three circles: supply (technically available), demand (asked for by management) and need (required for steering). Only the intersection of all three is the target; the other areas are zombie reports, wish lists and gaps (Ossola-Haring et al. 2019, pp. 94 f.). Gladen calls the same condition «scarcity amid abundance»: data supply, data demand and information need do not coincide – and the objective need cannot be gathered by asking recipients, only derived from the problems to be solved (Gladen 2014, p. 7).

The cost comes on three levels. First, attention: whoever shows 120 numbers has not prioritised, so the executive team does it themselves – by gut feel, in the meeting. Posselt puts it bluntly: the more detailed the metrics, the less the management can say (Posselt 2018, p. 29). Second, effort: every metric has to be defined, connected, reconciled at each close and adjusted whenever the chart of accounts changes. Allow half a day per metric per year and 120 metrics cost 60 working days – for numbers of which three are discussed. Third, trust: once two reports show the same figure with different values, the meeting negotiates the numbers instead of the decision.

The most expensive effect, though, is the gap. Collecting starts from supply, so the number for the decision at hand is rarely there. «Whatever reporting is possible gets produced» (Ossola-Haring et al. 2019, p. 96, our translation) – and what is possible is set by the system, not by the question.

The reversal is simple to state: start with the decision and derive the metric from it. This is not new. The Goal Question Metric method from software engineering has done exactly that for thirty years – goal, then question, then measure (Basili/Caldiera/Rombach 1994). De Langhe and Puntoni call it «decision-driven analytics», as opposed to «data-driven»: the data does not go looking for a purpose; the decision goes looking for its data (de Langhe/Puntoni 2020; 2024). Our framework «From decision to KPI» turns this into seven steps.

The framework in seven steps

Decision → Question → Metric → Data → Report → Effect → Review. Each step has a test question that closes it and an artefact it produces. If you cannot fill in the artefact, the previous step was not done properly – that is the built-in quality control. The rule that carries everything: a metric with no decision attached is not a steering instrument. It is decoration.

Step 1: The decision

Formally, a decision is a choice between alternatives with uncertain outcomes and a value measure that makes the combinations comparable (Albright/Winston 2015, p. 224). In an executive team's practice there are two kinds: recurring decisions – prices, capacity, stock levels, credit limits, headcount – and one-off decisions such as investments or site questions. Both belong on a list.

De Langhe and Puntoni recommend starting every analytics initiative by identifying the decision and enumerating its alternatives; only then does it become clear which data could influence the choice (de Langhe/Puntoni 2020). Gartner calls the discipline «decision intelligence»: decisions modelled as objects in their own right, with context, alternatives and feedback. We keep it simpler and call the artefact the decision inventory. For each decision it records the subject, the alternatives, the decision-maker, the rhythm, the deadline and the stakes – in CHF or in exposure. An SME with a five-person executive team typically arrives at ten to twenty decisions. The ICV report on data-driven steering describes the starting point well: an «unease» felt by the manager (ICV/Horváth 2016, p. 16). The unease is the raw material; the decision is its sharpened form.

Test question: Which alternatives are on the table, who chooses, and by when? · Artefact: decision inventory · Sources: de Langhe/Puntoni 2020; Gartner; Albright/Winston 2015.

Step 2: The question

Between decision and metric lies the question, and this is the step most often skipped. Goal Question Metric is the clearest template: from a goal, derive the questions whose answers make the goal verifiable, and only from the questions derive the measures (Basili/Caldiera/Rombach 1994). Streifer observes that the questions are rarely the problem – leaders ask good questions, but they do not «deconstruct their questions into doable analyses» (Streifer 2004, p. 2). And not every good question is answerable; the check for answerability and effort belongs before the metric, not after it (Streifer 2004, p. 35).

This is also where the information requirements analysis turns subjective or objective. The demand-oriented approach assumes the user knows their own needs best (Linden 2015, p. 37); the objective need follows from the problem, not from the wish (Gladen 2014, p. 7). We combine both in one interview per executive, following the sequence goal → problem → early warning → analysis → report → effect, in the tradition of the Kimball and Ross business requirements interviews: What are you trying to achieve? What stops you? How do you know when things are going wrong? What analysis would you run then? Which report do you use today? And what do you do with the information once you have it? The early-warning question is the most productive; it leads to the few metrics somebody actually reacts to.

Test question: Can the question be answered with existing data at reasonable effort – and does the answer change the decision? · Artefact: question catalogue per decision · Sources: Basili et al. 1994; Streifer 2004; Gladen 2014; Linden 2015.

Step 3: The metric

Only now do we measure. Parmenter distinguishes result indicators, which show what came out, from performance indicators, which show what to do – and reserves the term KPI for the few measures that are tracked frequently, owned by someone and open to influence (Parmenter 2020). Kaplan and Norton supply the linkage: the strategy map arranges metrics along assumed cause-and-effect relationships, so that every leading indicator points at an outcome (Kaplan/Norton 2000). Grove adds a rule missing from almost every KPI set: every indicator needs a counterpart, or it gets optimised until the opposite happens – stock against availability, revenue against margin, lead time against quality (Grove 1983).

Two tests decide admission. First, controllability: a figure that is measurable but that nobody can influence belongs in the appendix, not in steering. Weber and colleagues demand that drivers not linked into a hierarchy be questioned and otherwise eliminated from reporting (Weber et al. 2017, p. 92). Second, depth of deployment: the decomposition ends where a figure gives immediate guidance for action in daily business (Weber et al. 2017, p. 94) – «quality» does not, «scrap rate on line 2» does. And every metric gets a cadence: how often do you look at it? A metric without a rhythm is a number without an appointment.

The artefact is the KPI profile with nine fields: decision, question, metric, definition, source, cadence, owner, threshold, response. The first three come from steps 1 to 3, the next three from step 4, the last three from step 6.

Test question: Who can move this number, and how often does that person look at it? · Artefact: KPI profile · Sources: Parmenter 2020; Kaplan/Norton 2000; Grove 1983; Weber et al. 2017.

Step 4: The data

Data reality catches up with every metric, in five questions. What is the definition, with formula and boundaries? Which source – system, table, cut-off? At which granularity – by which attributes can the figure be broken down, and which value fields does it need? How often is it refreshed? And who is the owner, accountable for it being right?

Attributes and value fields are the two halves of one large table: customer, customer group, product, legal entity, period on one side; revenue, material, freight, contribution margin on the other. Whether a metric «by customer group» is possible is decided not in the report but in the customer master – is the attribute maintained, and by whom? The ICV report places that responsibility with the controller: systematic data quality management, above all for master data (ICV/Horváth 2016, p. 18). Ries demands that a metric be «auditable», traceable to the raw data (Ries 2011, ch. 7). We phrase it as reconcilability before beauty: the contribution margin in the report must reconcile to the income statement, or the meeting discusses the difference. And because the report hangs on the close, «from which working day after month-end is the figure valid?» belongs in the measurement plan.

Two artefacts result: a glossary with one binding definition per measure and per attribute, and a measurement plan recording source, granularity, frequency and owner per metric. The glossary later becomes the template for the data model.

Test question: From which table, with which formula, at which level, how often – and who is accountable? · Artefact: glossary and measurement plan · Sources: ICV/Horváth 2016; Ries 2011.

Step 5: The report

The report is a carrier for decisions, not a showroom for data. Three rules. First, message first: the title of every page says what the number means – «CM II wholesale 3 percentage points below budget, material costs are the driver» – not «CM II by customer group». Second, variance instead of value: the target is the baseline, the actual matters only as a deviation from it, and a threshold – say ±5 % – separates attention from acknowledgement. The comparison logic is always the same: prior year, actual, budget, time running left to right. Third, notation adapted from IBCS®, pragmatically: actual filled, budget outlined, prior year grey; identical scales on one page; no pie charts. The details are in the article What is IBCS®?.

Streifer describes which data actually moves people to act: the big picture, plus the details about what would need to change (Streifer 2004, p. 7). Hence one page per decision: the variance at the top, below it the breakdown by the attribute that structures the decision. We distinguish three types of report – glasses, page, hammer: the dashboard for overview, the report for analysis, the working report as an operational tool. A board-level decision needs the page.

And: prototype before build. The first report is made in Excel or on paper, with real figures from the last close, and tested in a meeting. Only once the executive team has prepared a decision with it is the BI build worth the effort.

Test question: Is the message in the title, and can you see the variance that matters within five seconds? · Artefact: report prototype and notation concept · Sources: IBCS® Standards 1.2; Streifer 2004.

Step 6: The effect

A report does not steer. Ossola-Haring and colleagues say it plainly: an automatically generated report or a glossy cockpit steers no company (Ossola-Haring et al. 2019, p. 96). Steering only emerges when a response path stands behind every metric: threshold → owner → action. The threshold defines when a variance deserves attention. The owner is the person who then acts – not the department. The action is named in advance, at least by type: prepare a price round, shift capacity, review a credit limit. The same authors formalise this as a traffic light with an action list: green is passed over, amber discussed, red triggers an action (Ossola-Haring et al. 2019, pp. 96–99).

Cadence is part of the effect: a monthly metric belongs in the monthly meeting, a weekly metric in the weekly review, and the steering calendar fixes which metric is on the table when. The ICV report calls application to concrete decisions the «acid test» (ICV/Horváth 2016, p. 18). And Streifer reminds us that a conscious decision to do nothing can be reasonable (Streifer 2004, p. 7) – provided it is minuted.

Test question: What happens when the number breaches the threshold – who does what, by when? · Artefact: response path per metric, steering calendar · Sources: Ossola-Haring et al. 2019; ICV/Horváth 2016; Streifer 2004.

Step 7: The review

A KPI set ages. Cause-and-effect hypotheses are, as Weber and colleagues write, «prone to error» and need constant checking (Weber et al. 2017, p. 91). Decisions get taken and drop off the inventory; new ones arrive. Hence a fixed date – twice a year is enough – at which every metric must answer one question: which decision did it change in the last twelve months? Albright and Winston provide the criterion from decision analysis: «Information that has no effect on the ultimate decision is worthless» (Albright/Winston 2015, p. 262). A metric that has changed no decision is removed or moved to the appendix; that is not a loss but attention regained.

Ries describes the mechanism as «pivot or persevere»: a regular, explicit decision on whether to hold a hypothesis or switch it – without a pre-defined expectation for the metric, that decision gets postponed indefinitely (Ries 2011, ch. 7). Applied to reporting: the profile states which decision the metric is meant to carry; at the review we check whether it did. The result is an updated inventory and a removal list – the artefact by which you recognise mature steering.

Test question: Which decision did this metric change in the last twelve months? · Artefact: removal list and updated inventory · Sources: Albright/Winston 2015; Ries 2011; Weber et al. 2017.

A worked example: price increase in January

A fictitious industrial group, three legal entities, component manufacturing for four customer groups: OEM, wholesale, project business, service and spare parts. Material costs have risen six per cent this year. The decision due in October: price increase from January – yes or no, and for which customer groups? Alternatives: no increase, a flat three per cent, differentiated by customer group. Decision-maker: the executive team, on a proposal from sales. Stakes: at CHF 80 million revenue, every percentage point of price is worth about CHF 800'000 in contribution – and every lost key account roughly the same the other way.

The questions from the interviews with the head of sales and the CFO: In which customer groups has contribution margin II fallen against budget and prior year, and how much of that is the material effect? Where did the last price round stick, and where not? Which customer group withdrew volume after the last increase? The first two can be answered from ERP data, the third only partly – attrition can be measured, its cause cannot. That is noted in the question catalogue.

From this, three metrics: contribution margin II by customer group (monthly, year to date, against budget and prior year), price realisation as realised over announced increase (per price round), and volume loss among the twenty largest customers per group after the last price round (quarterly). «Revenue by customer group» is dropped: it is in the report, but it does not change this decision.

The data: contribution margin II comes from the ERP invoice lines (revenue, material at standard cost, freight) plus the monthly material price variance from the close. The attribute «customer group» is in the customer master – in two of the three entities. In the third it has to be maintained; that is the first task in the measurement plan, owner sales back office, due end of September. The figure is valid from the fifth working day after month-end.

KPI profile: contribution margin II by customer group

Field	Content
Decision	Price increase from January – yes/no, which customer groups
Question	In which customer groups has CM II fallen against budget and PY, and how much is the material effect?
Metric	CM II by customer group, year to date, TCHF and % of revenue
Definition	Net revenue − material at standard − material price variance − freight; per invoice line, aggregated by customer group
Source	ERP invoice lines, customer master (attribute customer group), closing entry material price variance
Cadence	monthly, from 5th working day; decision meeting October
Owner	Head of sales; data owner controlling
Threshold	CM II margin more than 2 percentage points below budget in two consecutive months
Response	Sales submits price-round proposal per customer group to the executive team, incl. attrition risk top-20 customers

The report: one page, titled «CM II wholesale and project business below budget, material costs are the driver; OEM and service on plan». At the top, contribution margin II by customer group, prior year – actual – budget, with variance; below, the split into price, volume and material effects. The prototype was built in Excel and tested in the September meeting.

The response path: the wholesale margin has been more than two percentage points below budget for three months – threshold breached. The owner, the head of sales, prepares the proposal for the October meeting: four per cent wholesale, three per cent project business, OEM and service unchanged, with the attrition risk of the twenty largest wholesale customers as the counter-metric in Grove's sense. The executive team decides with one page instead of 34. The spring review checks whether the metric carries the next pricing decision – or has done its job.

A thought from the library that underpins the approach: Gladen writes that in information supply, IT specialists mainly determine the limits of what is technically possible – and that an information system must first be fed with economically meaningful data before it can deliver usable management information (Gladen 2014, p. 7). That is why steps 1 to 4 come before the tool. Build the data model before the questions are settled and you build the limits of the technically possible – not the need. The purpose question we ask in every report review is the one that opens the framework: what are these numbers supposed to say, and what are they for?

What you can do tomorrow – and what the scoping sprint delivers

Start with the decision inventory. Take the agendas of the last six executive meetings and write out every decision taken there, or that should have been – with alternatives, decision-maker and deadline. Put the current monthly report next to it and assign every metric to a decision. Anything that cannot be assigned is your first candidate for removal.

If you want to run the framework with us, the scoping sprint is the format: two to three weeks, fixed price, a decision-ready target picture. It delivers the artefacts of all seven steps:

• the decision inventory with alternatives, decision-makers and rhythm,
• the question catalogue per decision, checked for answerability and effort,
• the KPI profiles with all nine fields, including threshold and response,
• the measurement plan with glossary – definition, source, granularity, frequency, owner – as the template for the data model,
• the notation concept with page template, comparison logic and thresholds, adapted from IBCS®, plus a tested prototype.

If a reporting project is already under way and you want to know whether it carries the decisions it is built for, the project review is the better fit: a few days, fixed price, a written second opinion.

Sources

• Albright, S. C.; Winston, W. L. (2015): Business Analytics: Data Analysis and Decision Making, 5th ed., Cengage Learning.
• Basili, V. R.; Caldiera, G.; Rombach, H. D. (1994): The Goal Question Metric Approach. In: Encyclopedia of Software Engineering, Wiley.
• de Langhe, B.; Puntoni, S. (2020): Leading With Decision-Driven Data Analytics. MIT Sloan Management Review, sloanreview.mit.edu.
• de Langhe, B.; Puntoni, S. (2024): Decision-Driven Analytics: Leveraging Human Intelligence to Unlock the Power of Data. Wharton School Press.
• Gartner: Decision Intelligence, Gartner Glossary, gartner.com.
• Gladen, W. (2014): Performance Measurement – Controlling mit Kennzahlen, 6th ed., Springer Gabler (in German).
• Grove, A. S. (1983): High Output Management, Random House.
• IBCS® Standards 1.2, © IBCS Association, www.ibcs.com, CC BY-SA 4.0.
• ICV/Horváth (2016): Business Analytics – Der Weg zur datengetriebenen Unternehmenssteuerung. Dream Car report of the Ideenwerkstatt, Internationaler Controller Verein (in German).
• Kaplan, R. S.; Norton, D. P. (2000): Having Trouble with Your Strategy? Then Map It. Harvard Business Review, September–October 2000.
• Kempter, P. (2023): Interview guide Data-Driven Management Accounting, master's thesis (unpublished); in the tradition of Kimball, R.; Ross, M. et al.: The Data Warehouse Lifecycle Toolkit, Wiley.
• Linden, M. (2015): Geschäftsmodellbasierte Unternehmenssteuerung mit Business-Intelligence-Technologien, Springer Gabler (in German).
• Ossola-Haring, C.; Schlageter, A.; Schöning, S. (2019): 11 Irrtümer über Kennzahlen, 2nd ed., Springer Gabler (in German).
• Parmenter, D. (2020): Key Performance Indicators: Developing, Implementing, and Using Winning KPIs, 4th ed., Wiley.
• Posselt, G. (2018): Renditeorientierte Führungsstrategien – Kennzahlen motivierend einsetzen, 2nd ed., Springer Gabler (in German).
• Ries, E. (2011): The Lean Startup, Crown Business.
• Streifer, P. A. (2004): Tools and Techniques for Effective Data-Driven Decision Making, ScarecrowEducation.
• Weber, J.; Bramsemann, U.; Heineke, C.; Hirsch, B. (2017): Wertorientierte Unternehmenssteuerung, 2nd ed., Springer Gabler (in German).