Workforce Decisioning

From WFM Labs


Workforce decisioning is the practice of turning the measurements and forecasts a workforce function already produces into typed, bounded and logged decisions, each carrying a stated confidence, that software can act on inside a defined authority and that people review at named gates. It sits between two better-known layers. Workforce analytics describes the state of the operation: the forecast, the gap, the index, the adherence figure. Workforce automation executes an action on a platform: a schedule change, a routing threshold, an overtime offer. Between a number and an action there is always a decision, usually made by a person reading a dashboard or by a fixed threshold written years earlier, and decisioning treats that step as a designed component with its own inputs, outputs, failure modes and measurements.

The page sets out the division of labour the layer rests on, the three-tier pattern that implements it, and the series that develops each part, drawing on decision analysis, the human-factors literature on automation, and a class of software model released in 2026 for answering bounded questions with probabilities rather than with text.

Three layers, not two

Most descriptions of workforce technology recognise two layers. The first measures and predicts: forecasting, capacity modelling, threshold alerting, the Operational Volatility Index. The second acts: the scheduling platform, the routing engine, the intraday lever cascade. The decision that connects them is usually invisible because it is made by a person, and a person's decision does not need an interface.

Three things change when the decision is made visible.

  • The decision acquires a type. A real-time analyst looking at a service-level dashboard is answering a question: does this interval need an action, and which one? Written down, the question has a bounded answer: a proposition that is true or false, one option from a list, or a position on an ordered scale. Decision Primitives for Workforce Operations sets out these three shapes and maps the common workforce decisions onto them.
  • The decision acquires a confidence. A threshold alert is certain by construction: the number crossed the line or it did not. A decision is not. The same reading of the same dashboard produces an action from one analyst and a wait from another, and the operation has no record of how sure either was. A decision component states its confidence, and the operation can measure whether that confidence was warranted, by a reliability diagram for calibration and by the Brier score for overall probabilistic accuracy.[1]
  • The decision acquires a record. Logged with its inputs, its confidence and its outcome, a decision becomes data. Measuring noise, the unwanted variability in judgement that Kahneman, Sibony and Sunstein propose auditing directly, depends on such records; without them an operation cannot tell noise from signal in its own decisions.[2]

A decision layer is therefore not the same as more automation. An operation can run the layer with every decision still taken by a person; what changes is that the question, the confidence and the outcome are recorded. Automation enters later, class by class, when the record shows a class of decision is reliable enough to release, which is the rule Human Gates and Number Grades already applies to its action gate.

The division of labour

The layer rests on a division of labour between three kinds of component, and the most common design failure is to give one of them the work of another.

Kind of work Examples in a workforce operation Who does it Why
Calculation Erlang staffing, the hours gap, shrinkage arithmetic, date and interval comparisons, cost of a lever Deterministic code The answer is exact and auditable; a model that estimates arithmetic is slower and less reliable than code that performs it
Judgement Does this news item affect this queue? Which of these five levers fits this gap? How severe is this exception on a stated scale? A rule where one can be written; otherwise a fitted classifier or a typed decision model The question is bounded but requires reading state that a rule cannot enumerate
Decision with authority Commit overtime, publish a schedule, invoke overflow, declare an incident A person at a gate, or released automation for a class with a published catch rate Consequence and reversibility set who may sign; the calibrated probability of the answer decides only whether a released class may act without a signature

A third-party reference for the exemplar model summarises the split as: code calculates, the model judges, and people reason and decide.[3] The workforce version adds a fourth clause: the gate decides who may act, and the gate is placed by consequence. Parasuraman, Sheridan and Wickens treat the level of automation as a separate choice at each of four stages, from information acquisition through analysis and decision selection to action implementation, and recommend lower automation of decision selection and action where outcomes are costly and reliability is unproven.[4]

The three-tier pattern

Implemented, the layer has three tiers, each of which can be a separate service with a defined interface.

Tier What it publishes Workforce examples Property it must have
Signal Typed state about the world or the operation, with a grade on each number An index reading with its band and contributing events; a forecast-versus-actual decomposition; a capability record for a worker; an interval's occupancy and queue Versioned, graded, timestamped; a decision made on stale or ungraded state cannot be audited
Decision An answer to a typed question about the state, with its probability and confidence, the state version and the rule or model version it used, and a reason code assigned by the tier's own logic (which question fired, under which band), never a model-generated explanation "Does this event affect queue Q?" → yes, 0.91. "Which lever absorbs this gap?" → intraday reallocation, 0.58 Idempotent; no side effects; every output reproducible from the logged inputs and versions
Action An executed change on a platform, from a signed proposal A break resequenced, an overtime offer issued, a routing threshold moved Executes only from a proposal row that carries a signature, human or released automation, and writes the outcome back

The separation matters because each tier fails differently. A signal tier fails by feed loss and staleness. A decision tier fails by miscalibration and by answering a question it was not asked. An action tier fails by executing without authority. Merging the tiers hides which one failed. From Signal to Decision: The OVIX Case walks the pattern through one signal end to end.

What a typed decision model adds

Rules and fitted statistical classifiers have long occupied the judgement column of the division-of-labour table, joined more recently by general-purpose language models constrained to a fixed output set. Rules are auditable but cannot read unstructured state. Classifiers can, given training data, but each is built for one question. Constrained language models read anything but generate token by token and state their confidence poorly. In 2026 vendors began releasing models built only for the judgement task: state and a typed question in, a bounded answer with a probability out, no text generated. TypeSafe AI's Jev, in early access from September 2026, is the example this series uses; its vendor describes it as trained for calibration rather than fluency and as returning one of three primitives, a proposition's probability, a choice or a score, in tens to hundreds of milliseconds.[5][6]

This series uses that model as the exemplar of the class, with two cautions. The product is closed and waitlisted, and the published performance figures come from the vendor's own evaluation suite; as of November 2026 no independent benchmarks existed.[6] Nothing in the series depends on that product. The primitives, the calibration arithmetic and the gate design apply equally to a hand-written rule or to a logistic regression with a calibration layer, and a workforce function should expect to run all three side by side.

What the class changes, if its reported figures hold, is the cost of asking. When a bounded question costs little and returns in well under a second (the vendor reports tens to hundreds of milliseconds), an operation can ask many questions of every state change rather than a few questions of a daily report.[5] The design pattern that follows, fanning out every potentially useful question and letting code discard the answers that do not apply, is described on the primitives page.

What decisioning is not

  • It is not contact routing. Skill-Based Routing, the Value Routing Model and Next-Best-Action decide where a contact goes or what a customer is offered. Workforce decisioning decides what happens to supply: which lever, which pool, which horizon. The two consult the same capability record and a merged pool changes both, which is why Supply-Side Routing is treated as the first application.
  • It is not forecasting. A forecast is a signal. The decision is what to do about the gap between the forecast and the plan, and Variance Routing by Horizon is the first such decision in the control-system series.
  • It is not a replacement for the gates. A decision with high confidence is still a proposal until the class it belongs to has been released by evidence. The number grades of the agentic protocol continue to apply to every input: a decision is never better grounded than its weakest input.
  • It is not strategy. The typed-decision models read questions literally, cannot plan, cannot reason across steps and return no explanation.[3] A capacity strategy, an operating-model choice or a partner negotiation is not a bounded question and should not be forced into one.

The series

Page Question it answers Status
Decision Primitives for Workforce Operations What are the three shapes a workforce decision can take, what do confidence and calibration mean for each, and which common decisions map to which shape? Published
From Signal to Decision: The OVIX Case How does an exogenous index reading become a pre-positioning action through the three tiers, and what has to be true of each tier? Published
Supply-Side Routing Given a gap, which source of capacity absorbs it, at which horizon, under which eligibility and cost constraints? Published
Confidence-Gated Automation How does a decision's confidence map onto the act, escalate or ask-for-evidence ladder, and how is a class of decision released? In preparation
The Decision Log and Calibration Loop What is recorded for every decision, how is calibration measured on the operation's own data, and how does the result feed back? In preparation
Decision Latency Economics Where does a faster decision change the outcome and where does it not, and what does a decision cost at each horizon? In preparation

Prerequisites

The layer cannot be built on an operation that does not already have four things.

  1. Typed state. The signal tier publishes fields with definitions, not spreadsheets with headings. The Workforce Data Definitions Framework is the usual starting point.
  2. Graded numbers. Every input carries one of the four number grades, so that a decision's grounding can be stated.[7]
  3. A decision log. Decision, probability and confidence, inputs with grades, state version, rule or model version, signer, outcome. Without it there is no calibration and no release.
  4. A gate design. Which classes of action are released, at what confidence, under what catch rate, reviewed how often.

Limitations

  • Calibration is local. A model calibrated on a vendor's benchmark is not calibrated on an operation's queues. Every class of decision needs its own reliability measurement before its confidence means anything.[8]
  • A bounded answer can be confidently wrong. Constraining the output to allowed values removes malformed answers, not wrong ones.
  • The layer adds a component that can fail silently. A decision tier that drifts returns well-formed proposals of declining quality. The acceptance rate and the calibration curve are the instruments that detect it; without them the drift is invisible until an outcome is bad enough to notice.
  • Evidence for the model class is early. As of November 2026 the exemplar had been available for two months. Claims about its accuracy, latency and calibration should be read as vendor-reported until independent results exist.[6]

Maturity Model Position

  • Level 3 — Progressive: decisions are written down as typed questions and answered by rules or by people, with the answer, the confidence and the outcome logged. No automation is released; the gain is the record.
  • Level 4 — Advanced (The Ecosystem Emerges): a decision tier answers bounded questions on typed state with a stated confidence; classes of cheap, reversible action are released on evidence; the signal, decision and action tiers are separate services.
  • Level 5 — Pioneering: the calibration loop re-tunes thresholds and release decisions from outcomes; the decision tier is shared across planning, real-time and placement; agents in the workforce are themselves subjects of supply-side decisions.

See Also

References

  1. ↑ Brier, G. W. (1950). "Verification of forecasts expressed in terms of probability". Monthly Weather Review 78 (1), 1–3. doi:10.1175/1520-0493(1950)078<0001:VOFEIT>2.0.CO;2.
  2. ↑ Kahneman, D., Sibony, O., & Sunstein, C. R. (2021). Noise: A Flaw in Human Judgment. Little, Brown Spark. ISBN 978-0-316-45140-6.
  3. ↑ 3.0 3.1 Burnhill, P. (2026). "Comprehensive project reference for TypeSafe Jev: concepts, architecture, primitives, strengths, limitations, use cases, patterns, and practical guidance". GitHub Gist, reviewed 9 October 2026 against jev-1.13.0. Third-party reference, not vendor documentation. https://gist.github.com/pjburnhill/adf8d28efcad9df037bfdece178ef965
  4. ↑ Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). "A model for types and levels of human interaction with automation". IEEE Transactions on Systems, Man, and Cybernetics – Part A 30 (3), 286–297. doi:10.1109/3468.844354.
  5. ↑ 5.0 5.1 MarkTechPost (19 September 2026). "TypeSafe AI Releases Jev: A System One Model That Returns Typed, Calibrated Decisions Instead of Text". Press coverage of the vendor's launch announcement. https://www.marktechpost.com/2026/09/19/typesafe-ai-releases-jev/
  6. ↑ 6.0 6.1 6.2 TrueFoundry (2026). "TypeSafe AI's Jev: What 'System One Models' Actually Are". Third-party analysis; notes that benchmarks, latency and calibration claims are vendor-reported and unverified independently. https://www.truefoundry.com/blog/typesafe-ai-jev
  7. ↑ Human Gates and Number Grades, this wiki: the four number grades and the five gates of a planning agent team.
  8. ↑ Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). "On calibration of modern neural networks". Proceedings of the 34th International Conference on Machine Learning, PMLR 70, 1321–1330. https://proceedings.mlr.press/v70/guo17a.html