InsightsPoint of View

When Metrics Become Instructions

How the numbers a business chooses begin to govern the business it gets

A mobile-restroom rental delivery at a commercial construction site, showing service access and delivery logistics.
By
Cam Myers
Content type
Point of View
Reading time
7 min read
Status
Author’s perspective
Topics
Demand Generation

In brief

  • A measure becomes an operating instruction when it directs spending, rewards, or automated action.
  • Transaction value, retained revenue, and contribution can lead to different commercial choices.
  • Define what the score omits, when outcomes mature, and which obligations must hold regardless.

The argument

An owner approves a plan to lower the cost of acquiring customers. The marketing team improves its numbers. Sales handles more inquiries. The service team gets busier. Yet the business has little more money left at the end of the month.

Nobody necessarily made a bad decision within the assignment. The problem may be in the assignment itself.

A measure changes its role when a business begins acting to improve it. Cost per lead becomes a spending instruction. Bookings per employee becomes a staffing instruction. Inventory turnover becomes a purchasing instruction. The number is no longer simply describing performance. It is helping determine what the business does next.

That transition deserves more attention as companies put AI and automated decisions into daily operations. A system can become exceptionally effective at pursuing an objective that only partly represents what the owner wants. Better execution then makes the gap more consequential.

The decision underneath

If the system improved this measure very successfully, would it improve the business?

The number chooses the work

Consider a hypothetical regional equipment-rental company in the Midwest comparing two campaigns. Each spends $1,000 and produces $10,000 in booking revenue.

The first produces 100 leads and 20 bookings. Each booking contributes $120 after the costs directly associated with serving it, before advertising and shared overhead. The second produces 50 leads and 10 bookings, each contributing $350 on the same basis.

The first campaign wins on cost per lead and cost per booking. Both report the same revenue return on advertising spend. But the first leaves $1,400 after advertising, while the second leaves $2,500. Those figures are before shared overhead, and assume the estimated service costs and completed bookings hold.

This is not a dispute over arithmetic. It is a choice about which arithmetic should direct the next dollar.

Tell the system to find inexpensive leads and it has reason to favor the first pattern. Reward booking volume and it has another reason to do so. Give it a dependable estimate of contribution and the preferred allocation could change. Capacity constraints or customer retention could change the decision again.

The familiar concern is that people will manipulate the numbers. That happens, but it is only part of the problem. Honest employees and well-functioning software can pursue an incomplete objective exactly as instructed.

Steven Kerr documented this mismatch between desired behavior and rewarded behavior in 1975.[1] Automation adds another route through which the mismatch operates: the measure can become an input to repeated decisions without anyone consciously reconsidering the tradeoff each time.

What the value actually represents

In work with a regional auction operator in the Upper Midwest, I examined fee structures and caps alongside the economics of growth. The value sold under the hammer was not the same as the revenue earned by the auction company. Buyer premiums, seller commissions, and their terms changed how additional sales value translated into company revenue.

We modeled alternative fee arrangements. Their eventual commercial effect would also depend on bidder and seller responses. The important point for measurement was already clear: a larger transaction and a more valuable transaction for the operator were different questions.

That distinction belongs inside the information used to direct spending. Sending gross transaction value into an advertising system may be useful, but it does not make that value equivalent to retained revenue or profit.

The same problem appears in rentals. The initial invoice, the minimum committed contract, expected rental duration, and eventual collections represent different things. For a distributor, sales value does not deduct the purchase cost of stock or the costs of serving the order. Calling all these values “revenue” can hide differences large enough to reverse a decision.

Google's own documentation states that automated bidding uses the conversion values supplied to it and can optimize toward values such as revenue or profit margins.[2] The business therefore has a consequential responsibility before the bid is placed: decide what its supplied value means.

A prediction does not resolve that responsibility. A model may estimate how likely a customer is to book, or how much a product will sell for. Management still has to decide which outcomes deserve resources, subject to which limits. Only then should an execution system act on the recommendation.

Reporting, prediction, and instruction can use the same data while serving different purposes. Confusing them makes a technical setting carry a commercial decision nobody remembers approving.

What the measure leaves outside

Changing the main objective to profit does not remove the difficulty. Profit must still be defined over a period, assigned to a unit, and estimated from information that may arrive late.

A job that looks attractive before delivery may require extra travel or service. An order may be returned. A newly acquired customer may buy again. A decision that improves this month's contribution may consume the capacity needed for a better customer next month.

Some of these effects can be measured. Others require judgment. Research on incentives across multiple tasks explains why concentrating rewards on readily measured work can divert effort from valuable work that is harder to measure.[3] The practical response is to make the omitted responsibilities explicit, rather than assuming a more sophisticated score captures everything.

For example, a rental business may pursue contribution while limiting delivery distance and preserving availability for committed customers. An auction business may allocate advertising toward expected fee revenue while retaining obligations to sellers and a sufficient window for bidders to discover inventory. These limits belong in the operating design even when they are absent from the main score.

Timing creates another problem. If longer contracts report their value slowly while short jobs report immediately, the fast signal can dominate early comparisons. Waiting for every outcome to mature can make decisions too slow. Using estimates allows earlier action, but those estimates need to be checked against completed outcomes.

Attribution introduces a separate distinction. A transaction credited to an advertising campaign is not automatically a transaction caused by it. A system can get better at finding customers who would have purchased anyway. Where the investment warrants it, the business needs comparisons or experiments that help separate additional value from reassigned credit.

No single dashboard column resolves these issues. They are decisions about what the business values and what evidence it will accept.

Approve the instruction behind the metric

Before connecting a performance measure to compensation, spending, or an automated action, ask what would happen if the system improved that measure very successfully.

Which customers would it seek? Which jobs would it avoid? What work would increase elsewhere? What could deteriorate while the reported result improved?

Then examine the instruction itself. Define what counts, whose economics it represents, when it is recorded, and whether it describes a completed outcome or an estimate. State the obligations that must hold even when meeting them reduces the score. Assign someone responsibility for checking whether improving the measure still improves the business.

This need not become a long governance exercise. Start with one consequential decision and trace the number that drives it. Compare the action it currently encourages with the action the owner would choose using the fuller operating picture.

That comparison often reveals a more useful technology project than adding another dashboard or upgrading a model. The missing work may be a cost connection, a corrected contract value, a different decision rule, or a limit the system currently cannot see.

When metrics become instructions, choosing them becomes part of designing the business. The test is whether a system that follows those instructions well will create the results the owner actually intended.

Sources

  1. Steven Kerr. On the Folly of Rewarding A While Hoping for B. Academy of Management Journal 18 no 4 1975 pages 769 to 783.
  2. Google Ads Help. About conversion values. Accessed November 10 2026.
  3. Bengt Holmstrom and Paul Milgrom. Multitask Principal Agent Analyses Incentive Contracts Asset Ownership and Job Design. Journal of Law Economics and Organization 7 special issue 1991 pages 24 to 52.

For the leadership team

In practice

  • Trace one consequential spending or operating decision back to the number that directs it.
  • Define whose economics the value represents and distinguish estimates from completed outcomes.
  • Keep customer obligations, capacity limits, and credible evidence of additional value outside a single score where necessary.

Questions worth asking

  1. What could deteriorate while the reported measure improves?
  2. Would an owner with the fuller operating picture make the same allocation?

Weighing this decision now? Tell us what you are considering and we will say plainly whether we can help.

Discuss your business