Choosing the five KPIs that actually run the business

The question is usually asked as 'which five KPIs should I track', which has no answer, and is better asked as 'which five slots must be filled'. The slots are demand, throughput, margin, cash and people. Demand tells you what is coming. Throughput tells you whether the business can convert it. Margin tells you whether conversion is worth doing. Cash tells you whether you survive the conversion. People tells you whether next quarter still has the capacity to do any of it.

Filling the slots is model-specific and the slots themselves are not. That is what makes this framing portable across a trades company, a distributor and a professional-services firm, and it is why a generic list of operational metrics is useless to someone who just inherited a specific business with a specific constraint.

Three artifacts get confused here and they are different objects. A dashboard is continuous and self-serve — you look at it when you want to know something. A scorecard is weekly, reviewed in a meeting, and every row has an owner. A report is periodic and narrative. The scorecard is the one that changes behaviour, because it is the only one where a human is accountable for a number in front of other humans on a fixed schedule.

At least one of the five should be a leading measure. A scorecard composed entirely of lagging outcomes tells you accurately what has already happened and gives nobody anything to do differently on Wednesday. This is the single most transferable idea in the entire cadence literature and it is the one most often dropped when a team builds its first scorecard, because lagging numbers are easier to get.

One owner per number, and the owner is a named human rather than a department. Shared ownership of a metric reliably produces no ownership of it, and it is the most-violated rule in the category. It is worth being structural about: if you cannot name the person, the row does not go on the scorecard.

Every number needs a written definition before it needs a target. Two people will otherwise report the same metric differently and the disagreement will surface in month three, in a meeting, about a number that mattered. The definition dictionary is a dull artifact that quietly determines whether the scorecard is still trusted in the second quarter. It also protects you from a specific trap: the same word meaning different quantities in different traditions.

Map each number to its source system and budget the manual entries explicitly. Some numbers come out of the ledger, some out of the field-service or inventory system, some out of the customer database, and some out of a person typing. Every manual entry is a future failure, and knowing which rows are manual tells you where the scorecard will decay first.

Adoption is a separate problem from design and it is the one nobody publishes. Every framework steward publishes the artifact; none of them publishes the failure modes. What makes a team fill a scorecard in is that the number is reviewed in a meeting the owner attends, that the result is publicly visible, and that missing it has a consequence — in that order. Cascading the scorecard down to departments and individuals is a later move, and below roughly thirty headcount it usually produces ceremony rather than signal. That threshold is an editorial judgement, not a sourced finding, and it is offered as one.

One dependency runs underneath all of this: the chart of accounts. Margin by service line cannot be produced from a chart that grew by accretion under the previous owner. If the margin slot on your scorecard is going to mean anything, the account structure has to be redesigned to feed it, and that work belongs earlier in the year than most operators put it.

A trades business fills demand with booked-call rate, throughput with completed jobs per technician per day, margin with gross margin by job type, cash with collections timing, and people with technician availability. A professional-services firm fills demand with pipeline coverage, throughput with utilisation, margin with the effective bill rate against loaded cost, cash with collection realisation, and people with bench cost. A distributor fills throughput with lines shipped per labour hour and margin with gross margin dollars against average stock held. A shop fills throughput with cycle time against takt and margin with estimated-versus-actual job cost. A small subscription business fills demand with new bookings and margin with gross revenue retention.

Three same-word traps live inside those lists and each one has cost somebody a quarter. Utilisation means billable hours over available hours in one benchmarking tradition and a dollars-based ratio in another, so two firms reporting the same utilisation figure may not be comparable at all. Overhead rate is a multiple of direct labour in architecture, engineering and professional-services benchmarking, and a percentage of revenue in the trades — same phrase, incompatible scales. And a fifty per cent markup is a thirty-three per cent margin, which is the most expensive arithmetic confusion in trades pricing and the one that quietly breaks a flat-rate pricing conversion.

This is why the definition dictionary is not administrative overhead. It is the thing that stops a number from meaning one thing in the ledger, another in the field system and a third in the weekly meeting.

The next question is always what the number should be, and here the category performs a trick. Take a metric whose value you want, follow the figure back, and the chain runs: a field-service-software blog states a number, a listicle cites the blog, a consultant cites the listicle, and the number becomes 'the industry average'. There is no study at the bottom. Publishing that chain break is more useful to you than publishing another link in it.

The distribution of citable data is also perverse. Free, rigorous, methodology-disclosed benchmark corpora do exist — an architecture and engineering study covering 896 firms for financial year 2025, and an accounting-firm survey covering 1,073 firms for financial year 2024 — and they cover categories that are acquired comparatively rarely. The two models most commonly bought by operator-acquirers, the trades and distribution, are precisely the two where no such corpus is reachable: the trade association's financial survey is 2020-vintage and paywalled, and every primary distribution source we attempted returned a 403 or a 404.

So the standing rule on these pages is simple. A metric definition and its formula are safe to publish. A metric's typical value is not, unless it traces to a government or academic source we actually retrieved. Where it does not, the page ships the formula, the interpretation, and a plain statement that no verified benchmark exists — plus, where we traced it, where the circulating number came from. That is a more useful product than a fabricated target, and it is one no vendor-funded competitor can copy, because their marketing depends on the numbers.

The practical substitute for a benchmark is your own baseline plus a survivability calculation. Take three to four quarters of your own history on a stable definition and treat the trend as the target. Then compute how much revenue decline the business survives with debt service treated as a fixed cost — break-even, re-derived for a leveraged owner. That number is specific to your business, it is checkable, and it is the one that actually governs decisions.