Sourcing Strategy Under Imperfect Data

From WFM Labs

Sourcing strategy under imperfect data is the problem every multi-estate service operation eventually faces in its hardest form: near-term placement decisions — financial targets, client renewals, migrations — must be made now, while the measurement estate cannot yet support fair comparison between the pools the decision chooses among. Waiting for clean data forfeits the near term; deciding as if the data were clean manufactures confident wrong answers. This page describes a two-track approach: decide now, deterministically and with stated assumptions, while building the data foundation and placement engine in parallel — fed by the instrumented decisions themselves. It is the strategy companion to Placement Engine Architecture, which specifies the machinery, and to Placement Rules and the Tenure Contract, which supplies the decision instrument.

The position in five statements

  1. Start with the work — and expect it to move. The breakdown of the work, not the organisation chart, is the unit of the strategy; automation absorbs the routine first, so the work's shape shifts continuously.
  2. No single dimension decides. Revenue and cost govern in the end — beneath them sit what the service costs, how well it serves, what it earns, and how it flexes. Many measures, meaningful only when modeled against the work, together. Experience, capacity, and tenure are not outcomes but the supply levers beneath all four.
  3. Three different goods are bought — depth (retained expertise), scale (owned centres), and speed (contracted supply) — and each is the right answer to a different question.
  4. Near-term decisions cannot wait for perfect data — so they carry stated assumptions and readable results.
  5. Placement becomes a produced answer — recomputed as the work, the supply, and the objectives move.

Why two tracks

The two tracks and the data feed between them: the instrumented near-term moves generate what the engine needs.

Placement calls in most estates are educated and deterministic — made as snapshots. The people deciding weigh constraints they know deeply: ramp speed, what was sold, language floors, systems experience, client commitments. The limits are structural, not human: each call is decided once and never re-run; the context (tenure, cost basis, work type) lives in different systems, unlayered; and the constraints consulted — some measured, some not — feed no live model that tunes placement.

The two-track answer refuses the false choice between waiting and pretending:

  • Track one — decide now. Near-term moves proceed on today's data, selected on what is solid under imperfect measurement — cost arithmetic, contract facts, capacity facts, and structure, never quality instruments that cannot yet discriminate. Every assumption is stated. Every move is instrumented: one scorecard at origin and destination, receiving-team tenure recorded, ramp tracked separately from steady state.
  • Track two — build the foundation. Shared definitions, the constraint register, supply cards, the governed data layer, and the engine — with planning as the first consumer.

The anchor: the tracks are not sequential. The near-term decisions cannot wait for the model, and the model cannot be built without standardized, normalized, context-rich data — which the instrumented moves are precisely what generate. A move that banks its benefit but leaves the data as broken as it found it was half-executed.

Track one disciplines

Select on structure, not on scores. Where quality instruments cannot discriminate (see the comparability study below), near-term selection rests on structural risk — concentration, integration, chainability — and on arithmetic: coverage floors, pooling penalties, contract terms as written. This also prevents the worst near-term failure: a relocation that misses, measured on an instrument that cannot separate a new-team effect from a placement effect, manufactures a confident false lesson that outlives the data that created it.

The hand-run check. Every placement proposal is tested, by hand, against one written constraint list before signature — the gate of Placement Rules and the Tenure Contract, with its priced and dated outcomes. Constraint knowledge already exists in every estate; the check just makes one testable list of it.

State the evidence tier. Two tiers, never blended: hard data the operation can and should produce directly (counts, contracts, costs on a stated base, tenure records — where missing, the answer is get it); and calibrated estimates for the genuinely unmeasured, clearly marked, prioritised by value of information. An estimate printed as history ends arguments in the wrong direction.

Track two: the four-level data doctrine

The foundation is not "better reporting." It is four levels, each depending on the one before:

  1. Standards — one agreed meaning per measure, decided once.
  2. Hygiene — forward and backward. The under-appreciated half: a standard that coexists with its legacy predecessors is just one more definition. Every adoption names what it retires — the definitions sunset, the reports that stop. Old habits of reading past measures are hard to break; retiring the measures is what breaks them.
  3. Context — or the number lies politely. Two contexts are non-negotiable before comparison: tenure (a higher score from a five-year team against a lower score from a one-year team may describe the same capability at two costs — the experience-versus-expense trade-off is the honest decision object) and complexity (identical instruments on different work compare the difficulty of the work, not the performance of the people). Cost itself requires normalization, not judgment: contracted supply priced per productive hour and owned centres carrying loaded cost are two informative views that must be put on one basis.
  4. Data models drive strategies. The destination: placement decision trees today; value-based routing and dynamic supply-demand models tomorrow. A pool the models cannot see — telemetry-dark, definition-ambiguous — cannot participate in data-based distribution at all.

The comparability study — a reusable instrument

Before trusting any comparison, grade it. The method: inventory how quality (or any contested measure) is actually produced — every instrument, sampling rule, pass bar, and definition, cell by cell across the estate — then grade every realistic comparison path: A — same instrument, comparable coverage, case-mix adjusted, independent provenance: compare freely · B — differs on one dimension with a known adjustment: compare with the adjustment quoted · C — differs on instrument or provenance, unadjusted: directional only, never a stated delta · D — self-selected, self-scored, or definitions unestablished: report separately.

In practice, mature multi-estate operations grade few or no paths at A — multiple instruments with different pass bars wearing one metric name, sampling designs that shape the score, and statistical floors (small-team gaps within noise) are accumulation, not misconduct. The graded map does double duty: it is the evidence that standards and hygiene are prerequisites, and it becomes the trimming inventory — each standard adopted names the instruments and definitions it retires.

The work–experience fit

Experience matters most relative to the work in the seat. Judgment work compounds with years; well-specified work does not. The fit forms a quadrant with two visible failure modes and one invisible one: failure purchased (shallow experience on steep-slope work — loud, gets a meeting, and often misattributed to the receiving location) and premium wasted (deep, expensive seats quietly holding flat-slope work — appearing in no report, and at scale plausibly the larger loss). The fit also moves: automation steepens the residual while codification flattens slopes — one more reason placement is a standing answer, not a one-time sort.

Choosing from an illuminated map

The illuminated map: each cell holds the best-known distribution for that mix of cost and capability; greyed cells are constraint-infeasible; the gold cell is the chosen operating point.

A single trade-off frontier keeps only today's winners and discards everything else — so every change of objective restarts the analysis. The stronger construction borrows from quality-diversity optimisation (the MAP-Elites family): maintain the best-known solution for every combination of characteristics, not just the frontier. The planning surface becomes a grid — cost of the distribution against delivered capability — where each cell holds the strongest known way to distribute the work for that mix.

Three layers operate on the map, top down: plan — leaders choose the cell that matches the business objective · constrain — the constraint engine darkens what law, contract, and coverage rule out, and prices every cell that remains · orchestrate — supply-side routing executes the chosen cell, re-optimising as conditions move. The bright edge of the map is the classic Pareto frontier — but the whole map stays lit, so when objectives shift, the next cell is already known: the decision moves, and nothing is re-derived.

Volatility structure as a placement dimension

Demand volatility is largely imported — disruption events, client wins and losses, seasonality — and supply structures choose whether it is absorbed or amplified. Obligations that demand the most exactly when conditions are worst (minimum commitments billing in the trough, dedicated walls blocking reallocation in the surge, flexibility bounded by notice periods) amplify; genuinely variable cost shares, callable surge routes, and laddered steady intake absorb. Two disciplines follow: audit each pool's truly variable cost share at 30, 60, and 90 days, and validate flexibility against the last real disruption rather than contract language — including the correlated-surge test, since a sector-wide event reaches every client of a shared supplier at once. Hiring in waves is the internal version of the same defect: a wave ramps and departs together, a maturity cliff; steady intake is laddering.

The build: two prongs, one join

Prong one proves the logic by hand — the register seeded, the check run first retrospectively on a decision already made and going well, then on one live decision; a working mock built against acceptance cases, one of which is designed to fail (a proposal violating a hard constraint must be caught). Prong two builds the system of record above all — settled definitions, one semantic layer, metrics as governed facts: the long pole, on which everything else waits.

A commercial intelligence platform may credibly supply the governed layer, and a partner-built arrangement — the operation's logic and design assistance in exchange for the build — can be the right economics. Four conditions protect the operation regardless of vendor: the platform ingests nothing whose definition is not settled (forward-only correction makes early garbage permanent) · peer comparisons span the firm boundary · tenure is an always-on dimension, never dropped for small groups · hard constraints are never encoded as scores, and the constraint list remains the operation's own. The walk-away stays free at every gate: if the partnership disappoints, the layer gets built anyway.

Failure modes

Treating current practice as failure — placement calls today are educated and deterministic; the model formalizes existing judgment rather than replacing absent judgment, and a strategy framed as accusation loses the room that must adopt it · running the tracks sequentially in either direction · justifying near-term moves on quality instruments that cannot discriminate · blending evidence tiers · standardizing without retiring (hygiene half-done) · planning on a single frontier and discarding the rest of the map · adopting a platform before definitions settle · comparing scores without tenure and complexity context.

Maturity Model considerations

Levels 1–2: placement by episodic argument; data quality invoked to defer decisions, or ignored to force them. Level 3: the two tracks named; near-term moves instrumented; the comparability study run; definitions converging with a sunset discipline. Level 4: the hand-run check routine on live decisions; supply cards current; the graded comparability map maintained; evidence tiers enforced. Level 5: the governed layer live; the illuminated map maintained by the engine; planning cycles produced from the map; objective shifts absorbed by moving cells rather than restarting analyses.

See also

Retrieved from "https://wiki.wfmlabs.org/w/index.php?title=Sourcing_Strategy_Under_Imperfect_Data&oldid=3910"