Workforce Planning as a Control System

From WFM Labs

Workforce planning as a control system is a way of reading a service organization's planning process as a set of feedback loops rather than as a sequence of documents. In the conventional reading, an annual plan is built once, decomposed downward into monthly, weekly and daily plans, and executed; variance is explained upward at the end of each period. In the control-system reading, each horizon is a loop with its own sampling rate, lever, delay and frozen zone, and the horizons exchange state in both directions. The two readings describe the same organization. They differ in what they make visible: the first makes the plan visible and the loops invisible, the second does the opposite.

The page sets out the reading, the five loops it implies, the couplings between them, the pathologies that appear when only the fastest loop is closed, and the evidence from system dynamics and planning science that the reading rests on. A companion page, Feedback Loops in Workforce Planning, catalogues every loop with its polarity and signature. A browser simulation at loops.wfmlabs.com runs the two readings side by side on synthetic demand.

The linear cascade, read as a controller

Five controller loops stacked by horizon, each with its own sampling rate, lever, delay and frozen zone, and the five couplings that carry state between them in both directions. Index of structure, not a process map.

A controller measures a quantity, compares it with a goal, and acts through a lever after a delay. Whether the loop is closed depends on one thing: does the measurement change the decision inside the same cycle, or only at the next planning event?

Read that way, the conventional cascade has one closed loop and several open ones.

Horizon Sampling rate today Lever Feedback path today What goes wrong
Annual plan, one to three years once a year headcount envelope, budget variance explanation, annual the goal is frozen and reality is explained; a step change in demand is answered by defending the number
Monthly capacity plan monthly requisitions, moves reconciliation to the annual plan the forecast is bent toward the budget; the hiring pipeline is not counted in the gap, so hiring overshoots and freezes in turn
Weekly plan weekly schedules, overtime, time off forecast-versus-actual review nervous when re-planned, rigid when frozen; no route for its findings upward
Daily and intraday every interval real-time actions closed absorbs variance that belongs to a slower loop, so the slow loop never learns

Three structural facts follow. First, only the interval loop is closed, because it is the only horizon at which a measurement changes a decision before the next planning event. Second, variance has no owner: when demand runs eight percent above plan for six weeks [estimated], the interval loop carries it through overtime and occupancy because no slower loop is armed to take it. Third, the annual plan is a target that is treated as a forecast. Because it is one number, every re-forecast reads as a challenge to the commitment, so the forecast is bent toward the budget or the decision is deferred, and the pipeline delay turns deferral into oscillation.

Alan Kay's line that "point of view is worth 80 IQ points" describes the move.[1] A plan is not a number. It is a set of stocks and the policies that move them, and the annual plan and the intraday plan are the same system observed at different sampling rates. Once that is the point of view, the question "how do we reconcile to the annual plan?" is replaced by "which loop owns this variance, and is it closed?"

The five controller loops

Each horizon becomes a balancing loop. The table gives the stock each loop controls, its lever, the delay that sets its natural period, and the frozen zone inside which it does not reverse itself.

Loop Controls Lever Dominant delay Samples Frozen zone
B1 Interval correction the queue or occupancy gap now real-time actions, skill flex, voluntary overtime and time off offers minutes every 15 to 30 minutes none; this is the committed step
B2 Short-term reforecast the hours gap in the next one to two weeks the hours band, overtime and time off, schedule trades, cross-skill routing days daily the published week
B3 Pipeline-aware capacity the rolling thirteen-week surplus or deficit by pool, net of the supply line moves, requisitions, the partner's band, class starts eight to sixteen weeks, the hiring pipeline weekly roll, monthly decision the plan of record
B4 Structural elasticity the cost of variance; the share of the swing carried by the cheap levers contract bands, cross-trained groups, working-time averaging, location and automation mix quarters quarterly contract notice
B5 Finance envelope spend against a rolling envelope the envelope itself, authority bands a month monthly the fiscal quarter

Three design rules come with the table: one from control engineering, one from the research on planning nervousness, one from system dynamics. The first is from model predictive control: at each cycle, optimize over the horizon, commit only the first step, measure, and re-solve.[2] The second is from the research on planning nervousness: re-planning on a rolling horizon has a cost of its own, and the dominant remedies are a frozen zone and a change-cost rule.[3] Continuous planning does not mean updating everything all the time. It means every loop commits only its first step, keeps a frozen zone shorter than the delay it protects (this wiki's design rule; the production-scheduling studies found freezing up to half the horizon cost little[4]), and pays a stated cost to break the freeze.

The third rule is from the beer game. Loop B3 must count the supply line, the requisitions open, offers out and classes in training, in the gap it is closing, or it hires into a surplus that lands twelve weeks later.[5]

The couplings

The loops become one system through five couplings. Two carry state upward (L1, L2), one across (L3), two downward (L4, L5).

  • L1 Vintage learning (up). Forecast error, measured by the horizon at which the forecast was made (see Forecast Vintages), retunes the model or assumption that owns that horizon. Daily bias tunes the short-term model; thirteen-week bias tunes the long-term drivers. Reconciliation across temporal hierarchies keeps the levels coherent while each moves.[6]
  • L2 Assumption ledger (up). The annual plan is decomposed into named assumptions, volume, mix, handle time, shrinkage, attrition, ramp and productivity, each held as a stock with a tolerance band. Actuals update the ledger continuously, and the plan is recomputed from it, so "annual" becomes a horizon rather than an event. Breaching a band is what opens a decision, not the calendar.
  • L3 Variance routing (across). Each week's gap is decomposed into level, mix, handle time, shrinkage and supply, and each component is assigned to the slowest loop whose delay matches the component's persistence. Transient variance stays with B1 and B2; variance that has persisted for the persistence window (four weekly cycles in the simulation) goes to B3; variance that has persisted longer than one pipeline goes to B4. Nothing persistent is allowed to live in B1. The method is on Variance Routing by Horizon.
  • L4 Envelope feed-down. The finance loop publishes a rolling envelope and authority bands inside which B3 may act without escalation. This replaces "reconcile to the annual plan" with "stay inside the band or trigger the decision".
  • L5 Anticipation (down). Each slow loop carries a model of how the fast loop will respond: B4 sizes bands knowing how much B2 can absorb, and B3 sizes hires knowing B4's bands. Hierarchical production planning requires the upper level's plan to be feasible once disaggregated at the lower level, and without a model of the lower level's response it often is not.[7]

Three ledgers, never one number

The conventional annual plan carries three things in one number: what the organization expects (a forecast), what it wants (a target), and what it has bought (a commitment), and tying them together rewards bending the forecast.[8] The control-system reading keeps them as three ledgers, with the annual plan becoming the commitment ledger, an envelope with authority bands maintained by B5; the argument is made in full at The Annual Plan as an Open-Loop Controller.

What happens when only the fastest loop is closed

Four pathologies the linear cascade exhibits, drawn as a causal-loop diagram: hiring oscillation from the ignored supply line, the overtime–attrition spiral, quality erosion when the aspiration follows performance, and the capability trap.

Four pathologies appear, each with a published pedigree: one delayed balancing loop that oscillates, and three reinforcing loops. The catalogue numbers them R6, R2, R3 and R4.

  • Hiring oscillation (R6). When the supply line is not counted in the gap, the organization keeps hiring while earlier hires are still in the pipeline, overshoots, freezes, loses people to attrition during the freeze, and is short again. In the beer-game experiments the overwhelming majority of subjects produced costly oscillations of this kind, and the estimated decision rules placed little weight on the supply line.[5] The natural period is a multiple of the pipeline delay; in the companion simulation it runs at about three times the pipeline for the default parameters [estimated].
  • Quality erosion (R3). A temporary gap between capacity and demand, combined with decision rules for effort, overtime and the quality aspiration, produces permanent erosion of service standards: the goal drifts down to meet performance. The result was calibrated on a back-office service centre.[9] A budget that does not move is exactly the temporary gap made permanent.
  • The overtime–attrition spiral (R2). A gap raises overtime and occupancy, which raise attrition, which widens the gap. The field's own signature loop; see Systems Thinking and Recruiting Pipeline and Capacity Planning.
  • The capability trap (R4). Under pressure, time spent improving the planning system is cut to fight fires, capability erodes, and the fires multiply. The authors name the short-term relief a "Shortcuts" loop and the long-term loss a "Reinvestment" loop.[10] Intraday firefighting that carries persistent variance is the Shortcuts loop; fixing the slow loops is Reinvestment.

Service supply chains amplify variance even though they hold no inventory: with nothing to buffer, backlog can only be managed by changing capacity, and capacity-adjustment delays amplify it upstream. Uniform targets across stages make the amplification worse, and the backlog is best held nearest the customer.[11] A planning cascade with one target cascaded through every tier is a uniform-target chain.

The reading in one application

A step in demand in a flat world: agents on the floor, the supply line and overtime under the linear cascade and the closed loop. Synthetic demand, index only [estimated].

The companion simulation runs one weekly stock-and-flow model under two policies on identical synthetic demand. Policy A is the linear cascade: a forecast refreshed monthly and blended toward the annual plan, the supply line ignored, the whole gap hired monthly inside a fixed annual envelope, and a quality aspiration that follows performance. Policy B closes the five loops: a weekly forecast on actuals, the supply line counted, only persisted variance released to hiring after a frozen zone, a rolling envelope, the aspiration held. Both face the same pipeline, the same operational absorbers and the same cost function, and both pay a charge for every change to the plan of record.

Four observations held across the parameters tried and are stated as indices rather than quantities [estimated]. In a flat world with one step change, the two policies cost about the same, because overtime bridges a shortfall cheaply for a long time; the cascade then rings for the rest of the run at a period near three pipelines while the closed loop settles. Seasonality, a slow trend and disruption spikes open a cost gap of about three percent per contact in favour of the closed loop, with service a statistical tie and about a quarter less hiring churn; a slow decline on its own widens the gap to about nine percent. With the pipeline delay removed, the two policies converge to within about two percent, the residual being hiring-churn smoothing; roughly half of the seasonal gap is attributable to delay, and the rest to forecast anchoring and churn smoothing. An ablation ladder that turns the cascade into the closed loop one switch at a time shows that no single switch carries the result: counting the supply line reduces the ringing modestly; forecasting weekly on actuals is often the largest single gain; slowing the adjustment and routing only persisted variance to hiring is what removes the ringing and bounds the churn, while the routing and frozen-zone switches lose under fast upward shocks and long pipelines. Some switches hurt until the ones after them are thrown.

Prerequisites

The reading asks for four things that are not planning decisions. One definition of a productive hour and one headcount, read from the same data by every loop (see Workforce Data Definitions Framework). A version on every forecast, without which L1 cannot run. A decomposition of variance by cause, without which L3 has nothing to route. And a commitment ledger separate from the forecast, without which the three-ledger split is a slogan.

Limitations

The reading is a model, and the five loops are a design rather than a description of any organization. The couplings are the least developed part: hierarchical planning has long described feedback between levels and seldom specified it, and the same is true here. The quantitative results come from one simulation with round, illustrative parameters and should be read as a demonstration of structure, not as a forecast of savings. No peer-reviewed system-dynamics model calibrated on contact-centre staffing was found by this review; the nearest is the service-centre calibration above. And nothing in the reading removes the information asymmetry between the planning function and the operation it plans for; it moves the argument from the number to the definition.

Maturity Model Position

At lower levels of the WFM Labs Maturity Model™ the cascade is the operating model and the interval loop is the only one closed, usually by a real-time automation platform. Level 4: Planning in Distributions closes B2 and B3 by planning on ranges and re-forecasting on a rolling cadence. Level 5: Adaptive Orchestration describes the operation "run as a loop" at the interaction level; this page describes the same four moves, sense, decide, act, learn, at the planning level, and the two meet when L3 routes variance between them.

See Also

References

  1. ↑ Quote Investigator, "Point of View Is Worth 80 IQ Points", 29 May 2018, tracing the line to Alan Kay's closing talk at the Creative Think seminar, 20 July 1982. https://quoteinvestigator.com/2018/05/29/pov/
  2. ↑ Rawlings, J. B., Mayne, D. Q., Diehl, M., Model Predictive Control: Theory, Computation, and Design, 2nd ed., Nob Hill Publishing, 2017.
  3. ↑ Blackburn, J. D., Kropp, D. H., Millen, R. A., "A comparison of strategies to dampen nervousness in MRP systems", Management Science 32(4), 1986, pp. 413–429.
  4. ↑ Sridharan, V., Berry, W. L., Udayabhanu, V., "Freezing the master production schedule under rolling planning horizons", Management Science 33(9), 1987, pp. 1137–1149.
  5. ↑ 5.0 5.1 Sterman, J. D., "Modeling managerial behavior: misperceptions of feedback in a dynamic decision making experiment", Management Science 35(3), 1989, pp. 321–339.
  6. ↑ Athanasopoulos, G., Hyndman, R. J., Kourentzes, N., Petropoulos, F., "Forecasting with temporal hierarchies", European Journal of Operational Research 262(1), 2017, pp. 60–74.
  7. ↑ Hax, A. C., Meal, H. C., "Hierarchical integration of production planning and scheduling", in Geisler, M. A. (ed.), Studies in the Management Sciences, vol. 1: Logistics, North-Holland, 1975.
  8. ↑ Jensen, M. C., "Corporate budgeting is broken—let's fix it", Harvard Business Review 79(10), 2001, pp. 94–101.
  9. ↑ Oliva, R., Sterman, J. D., "Cutting corners and working overtime: quality erosion in the service industry", Management Science 47(7), 2001, pp. 894–914.
  10. ↑ Repenning, N. P., Sterman, J. D., "Nobody ever gets credit for fixing problems that never happened: creating and sustaining process improvement", California Management Review 43(4), 2001, pp. 64–88.
  11. ↑ Anderson, E. G., Morrice, D. J., Lundeen, G., "The 'physics' of capacity and backlog management in service and custom manufacturing supply chains", System Dynamics Review 21(3), 2005, pp. 217–247.
Retrieved from "https://wiki.wfmlabs.org/w/index.php?title=Workforce_Planning_as_a_Control_System&oldid=5674"