Instrumenting the Objective Before Building the Model

From WFM Labs
Three terms in the objective, one of them measured: the optimizer bends toward the only term it can see.

Instrumenting the objective before building the model is the design rule that a multi-objective placement or sourcing model can optimize only the objectives for which every node carries a measured attribute, and that an objective for which no node carries a measured attribute is not weighted low — whatever weight the objective statement assigns it, it discriminates between no two candidates, and its effective contribution is zero. The rule sounds obvious and is violated routinely: sourcing programs are launched with a stated ambition to balance cost, customer experience and employee experience, and deliver a cost-only answer, because cost was the only attribute measured for every node under consideration. The collapse is then read as a choice the program made when it was a property of the program's inputs. The rule is the input-side condition on the supply cards specified at Placement Engine Architecture, which requires at least one measured attribute per objective family per pool; what this page adds is the mechanism by which a missing attribute collapses the comparison, and the governance of the objective itself as a published, ordered, owned list. The three-objective governance frame — cost, customer experience, employee experience — is at Value-Based Planning Model; the sourcing decision under cost pressure at Sourcing Strategy Under Cost Pressure.

The collapse mechanism

A model that chooses among nodes — locations, delivery arrangements, human or automated handlers — compares them on attributes. If every candidate node carries a measured cost and none carries a measured customer-experience or employee-experience attribute, the comparison is on cost, whatever the objective statement says. The other objectives do not lose the argument; they never enter it. Three features make the collapse hard to see from inside the program.[1]

  • The objective is rarely written down as a sentence. Programs state ambitions — "best-cost footprint," "location-agnostic quality" — and never a sentence that begins "our objective is," so there is no text against which the model's actual behavior can be checked.
  • Scenarios are rejected against an unstated ambition. Candidate designs are evaluated and discarded, and the criterion doing the discarding is whichever attribute the spreadsheet holds for every row. The rejections look like judgment and are arithmetic.
  • The measured attribute is the one that already existed. Cost is instrumented because finance instruments it for other reasons; customer experience and employee experience by node are instrumented by nobody, because no function needed them before the placement question arose. The collapse is toward whatever the estate happened to measure, which is why it is almost always toward cost.[1]

The decision-analysis literature has long argued that the first act of any decision is to articulate the values the decision serves and to derive measurable attributes for each, before alternatives are generated or compared;[2] the reverse — generating alternatives first and reading the criteria off them — is what Keeney terms alternative-focused thinking, and a placement model built that way inherits the attributes its alternatives happened to carry.

One exception to the collapse is worth stating.An objective can be carried as an objective carried a screen rather than a scored term — a floor that disqualifies a node, as in the gate tests at Placement Engine Architecture — which keeps the term alive without a continuous attribute, at the cost of being unable to trade it off against the others.

The instrumenting rule

One measured attribute per objective, per node, before the model is built. The terms are whatever the objective names; where the frame is the cost, customer experience and employee experience triple of Value-Based Planning Model, the rule requires three measured attributes per node, and where a fourth term such as revenue or flexibility is added, four. Every candidate node, including any automated or agentic node, carries a measured value for each term before it is admitted to the comparison. The rule has four consequences.

  1. It forces the objective to be stated. An attribute cannot be instrumented for an objective that has not been named, so the rule produces the missing sentence as a side effect.
  2. It applies to the agentic node. An automated handler admitted to a placement comparison on cost alone — with no measured experience attribute and no measured completion or escalation attribute — collapses the comparison toward cost as surely as an unmeasured offshore node does. The escalation attribute is at The Escalation Tax; the completion attribute at Interior Optimum (containment rate); the full card an automated pool must carry, including drift terms, at Placement Engine Architecture.
  3. It exposes the not-assessable. Where a node cannot produce the attribute, the finding is recorded as not-assessable rather than imputed, and the instrumentation gap becomes a program prerequisite; the interpretive rule is at Interpreting WFM Maturity Assessments.
  4. It is cheaper than the collapse. One measured attribute per node per objective is a bounded instrumentation task. A program that delivers a cost-only answer under a three-objective ambition, and is then challenged on it, repeats the whole program.

The attribute need not be perfect. A coarse, comparable measure of experience by node — the same instrument, case-mix adjusted, per Comparing Delivery Arrangements — is enough to keep the term in the objective; the collapse happens at zero, not at coarse. Where a revenue term is carried it is the one most often absent entirely, and a coarse proxy — conversion or save rate on the node's own routed contacts, at equal case mix — is enough to keep it in the comparison.

The objective as a governed artifact

Instrumenting the terms is half the rule; governing their order is the other half. The governance layer can weigh only what the nodes report, and an objective with several terms has weights; in most programs the weights are set implicitly by whoever holds the model.[1] The weighting machinery itself — weighted sums, dynamic weights — is at Multi-Objective Optimization in Contact Center; what changes with governance is the source of the weights. The corrective is to publish the objective as an ordered, dated, owned list that a named authority can reorder.

The objective as a published artifact
Element What it fixes
Ordered The terms are ranked or weighted explicitly, so the model's trade-offs are the organization's rather than the modeler's
Dated The order is stamped, so a result can be read against the priorities in force when it was produced
Owned A named authority — not the model's custodian — holds the order and is the only party who may change it
Reorderable by signature A change in priority becomes a recorded reordering event rather than a modeling crisis; the same model, re-run under the new order, is the new answer

The last element is the operational point. Owner priorities change — a cost target arrives, an experience commitment is made to a client — and a model whose objective is implicit either breaks or is quietly re-tuned. A model whose objective is a signed list is re-run. It also removes the default that otherwise operates: whoever holds the model chooses the objective by omission, which is the same pattern Platform-Imposed Commercial Constraints: When the Platform Decides the Product describes for configuration.

Failure modes

  • Building the model first. Alternatives are generated and compared on the attributes they happen to carry, and the objective is inferred afterwards from what the model did.
  • Admitting a node on cost alone. A new location or an automated handler enters the comparison with one attribute, and the comparison collapses around it.
  • Imputing the missing attribute. A plausible experience figure is assumed for the unmeasured node, and the model's answer inherits the assumption as if it were a measurement.
  • Leaving the weights with the modeler. The objective's order is whatever the spreadsheet's structure implies, and a change in owner priority becomes a rebuild.
  • Stating the ambition and not the objective. "Balanced" appears in the charter and nowhere in the arithmetic.[1]

Maturity Model Position

The instrumenting rule is a Level 3 prerequisite on the WFM Labs Maturity Model™ — it is measurement discipline applied to the inputs of a decision — and the condition without which the Level 4 goal architecture of primary objectives, hard constraints and risk metrics described at Level 4: Planning in Distributions cannot be solved. A Level 4 optimizer over a Level 2 instrument set is the same failure that page names as Level 2 planning in Level 4 notation.

See Also

References

  1. ↑ 1.0 1.1 1.2 1.3 Practitioner observation from sourcing and placement programmes in multi-node service estates; a consistent pattern rather than a measured result.
  2. ↑ Keeney, R. L. (1992). Value-Focused Thinking: A Path to Creative Decisionmaking. Harvard University Press.
Retrieved from "https://wiki.wfmlabs.org/w/index.php?title=Instrumenting_the_Objective_Before_Building_the_Model&oldid=4358"