Three Bands of Work
Three bands of work is a segmentation principle for service operations and the workforce functions that plan them. It sorts every unit of work into one of three bands — discretionary, overseen and specified — according to two properties of the process itself: how completely it is documented, and whether its outcome can be measured. The band is a property of the process as written down, not of the person or system currently performing it. Work moves from discretionary toward specified only as its documentation and measurement mature, and it can move back. The principle is the sorting rule beneath the agentic handover gate and the agent overseer role. It is the third stage of the agentic journey map, and the first at which a workforce function can say which of its own processes are candidates for agents at all.
The three bands
| Band | What defines it | Who performs it | Examples in a workforce function | Examples in servicing |
|---|---|---|---|---|
| Discretionary | Judgment-bearing and consequential; the objective itself may be contested; not fully specifiable by nature | People, permanently | Long-range and strategic forecasting · ordering competing objectives · arbitrating between definitions · negotiating with suppliers · deciding a capacity trade a business leader has declined | A complaint with regulatory exposure · a claim that turns on intent · a patient or traveler in a disruption · any contact where the customer's state, not the task, sets the difficulty |
| Overseen | Specified enough to run without a person performing each step, but not yet trusted to run alone; exceptions carry stakes | An overseer with agents | Schedules generated and reviewed before publication · reallocation proposals approved · exception queues · sampled verification · agent evaluation | A claim adjusted within authority with a human sign-off · an account change within policy with sampled checks |
| Specified | Decomposed to a step table and work instructions; deterministic inputs; a measurable definition of done | Agents, then conventional software | Schedule publication · intraday reallocation within guardrails · assembling the monthly outlook · reconciling supplier invoices to forecast | Case creation and routing · documentation · address and payment changes inside rules · refunds inside rules |
The bands are deliberately few. Finer segmentations tend to reproduce the org chart; three bands force the question that matters for automation — can this process be written down and measured? — to be answered before any question about who or what should do it.
Three rules that make the bands usable
The band is set by the documentation, not by the worker
The L0–L3 documentation standard gives a process four levels: an L0 identity, an L1 one-page flow, an L2 step table naming the tools each step touches, and L3 work instructions. Documentation level is the necessary condition for each band; documentation level and a measured outcome together are the sufficient one. A process is discretionary until it has an L1 flow. It is eligible to be overseen once it has an L2 step table and a measured definition of done. It is eligible to be specified once it has L3 work instructions and an overseen phase has produced a published catch rate for its errors. The documentation standard therefore does double duty as a classification, with the catch rate as the one measurement the standard itself does not contain. The alternative — sorting work by who does it today — reproduces every inherited habit as if it were a law of the work.
The rule has a consequence that is easy to miss. Most of the work a service operation regards as "judgment" is in fact undocumented specified work. The act of decomposing it to a step table often reveals that the judgment lives in two or three steps, not the whole process, and those steps can be isolated as a discretionary sub-process while the rest moves. This is the same decomposition logic as shrinking a specialist's scope rather than relocating the specialist.
Stakes and complexity are independent axes, and the band test uses both
Stakes and complexity are independent dimensions of case mix. A process can be simple and consequential at once — a refund inside rules is specified work; a refund inside rules for a customer stranded overnight is not, because the stakes have changed the definition of done. A specified process with high stakes therefore stays overseen regardless of how deterministic its steps are. A complex process with low stakes may still move. Treating complexity as the only axis is how an operation hands high-stakes work to automation by accident.
Movement is in both directions, and it is logged
A band change is a placement decision and is recorded as one: who made it, on what date, on what evidence, and under what conditions it is re-posed. When an agentic process's catch rate slips, or the domain it was measured in shifts, the process moves back a band and the record says why. Without the log, band membership drifts toward whatever is convenient, and the operation loses the ability to say why a process is where it is.
Relation to other decompositions
The bands are analogous to, but not the same as, two decompositions the wiki already holds. The decomposition of a single contact into intake, transaction authority and asynchronous fulfillment, and of a service chain into its nodes (Service Chain Decomposition and Node Sourcing), cuts work by where in the chain it sits. Fulfillment is the specified band of servicing by construction, but intake and transaction authority do not map one-to-one onto the other two bands, and neither decomposition has a stage corresponding to overseen. The bands cut by documentation and measurement, and the overseen band exists precisely because it is not a place in a chain but a state a process passes through. Applying the bands to a workforce function's own processes, and to back office, is what is new.
The bands also reconcile two readings of agentic capacity that the wiki holds together: at decision altitude an agentic capability is a node that stands in the placement gate like any other, and at strategy altitude it is a force that takes the easiest work first and hardens the residual (The Hardening Residual). The specified band is where the node stands; the discretionary band is what hardens.
Relation to task-based models of automation
The bands are an operational form of the task-based view of technological change, in which a job is a bundle of tasks and automation substitutes for the codifiable ones while complementing the rest.[1][2] Two refinements matter for service operations. Codifiability is not fixed: it is a function of how much effort an organization has spent writing the task down, which is why the band is tied to documentation level rather than to an intrinsic property of the work. And field evidence on generative AI shows the boundary of what automation handles well is jagged rather than smooth, so it has to be measured per process rather than inferred from a difficulty ranking.[3] The overseen band exists to do that measuring. The layering of quality-system documentation from policy through procedure to work instruction is the antecedent of the L0–L3 standard the band test reads.[4]
How the bands are used
- Sorting a workforce function's own processes. The first application is to the planning function itself: which forecasting, scheduling, intraday and capacity processes are specified, which are overseen, and which are genuinely discretionary. The sort is usually a surprise; the volume of specified work is larger than the function believed, and the discretionary core is smaller and more concentrated.
- Placing servicing work. Asynchronous fulfillment is the specified band of servicing by construction, which is why it is the first candidate for consolidation and then automation.
- Sizing the overseer. The overseen band is the home of a specific worker type, sized by exceptions and stakes rather than by agent count.
- Re-sizing after each move. Every process moved to specified changes what remains in the other two bands — fewer runs, higher stakes per run, more judgment per head. The hardening residual is the reason the sort is repeated at every planning cycle rather than done once.
Failure modes
| Failure mode | What it looks like | Countermeasure |
|---|---|---|
| Sorting by worker | "Analysts do judgment work" — the band inherits the org chart | Sort by documentation level; decompose before classifying |
| Single-axis test | High-stakes simple work handed to automation because it is simple | Stakes as an independent axis; high stakes stays overseen |
| One-way movement | No process ever returns to a higher band; catch rates decay unread | Band changes logged with re-posing conditions; catch rate reviewed each cycle |
| Band inflation | Everything undocumented is declared discretionary to protect it | The band is a claim about documentation, not about value; undocumented is a state to leave, not a status |
Maturity Model Position
At Level 2 the bands are a diagnostic: sorting the function's own processes reveals how little is written down. At Level 3 the specified band is where rule-based automation already lives, and the sort names its boundary. At Level 4 the overseen band becomes a managed population with published catch rates, and at Level 5 band membership is re-evaluated continuously as part of the operating loop.
See Also
- The Agentic Handover Gate — the tests a process passes to move from overseen to specified
- The Agent Overseer — the worker type that owns the overseen band
- Process Decomposition (L0–L3) — the documentation standard the band test reads
- Stakes and Complexity: Two Axes of Case Mix — why the band test needs two axes
- Service Chain Decomposition and Node Sourcing — the same cut applied to a service chain
- Conservation of Labor — where displaced specified work goes
- The Agentic Journey Map — the sequence of stages this principle sits within
References
- ↑ Autor, D. H., Levy, F., & Murnane, R. J. (2003). "The Skill Content of Recent Technological Change: An Empirical Exploration". Quarterly Journal of Economics 118 (4), 1279–1333. doi:10.1162/003355303322552801.
- ↑ Autor, D. H. (2015). "Why Are There Still So Many Jobs? The History and Future of Workplace Automation". Journal of Economic Perspectives 29 (3), 3–30. doi:10.1257/jep.29.3.3.
- ↑ Dell'Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality". Harvard Business School Working Paper 24-013.
- ↑ International Organization for Standardization (2001). ISO/TR 10013:2001 — Guidelines for quality management system documentation. Geneva: ISO. (The 2021 revision, ISO 10013:2021, retired the explicit documentation hierarchy; the layered form cited here is the 2001 technical report's.)
