Practical guide

How to Create a Capacity Plan with AI

A capacity plan turns expected work into a time-bounded decision about people, specialist review, systems, and contingency. AI can help structure assumptions, compare scenarios, and find inconsistencies, but the accountable owner must approve the demand model, productive-capacity rules, constraints, and actions. The result should be a versioned planning artifact, not a confident-looking forecast.

By Two Prune

A capacity plan turns expected work into a time-bounded decision about people, specialist review, systems, and contingency. AI can help structure assumptions, compare scenarios, and find inconsistencies, but the accountable owner must approve the demand model, productive-capacity rules, constraints, and actions. The result should be a versioned planning artifact, not a confident-looking forecast with hidden inputs.

Define the planning decision and unit

Start with the decision: what period, service, work type, and resource action must this plan support? Name the owner, planning horizon, refresh trigger, and the date after which the inputs are stale. Choose a unit that can be counted consistently, such as reviewed proposals, support cases by complexity, assessment reports, or analyst hours.

Do not mix unlike work into one volume. Segment demand where handling time, skills, evidence requirements, or review burden differ materially. A “case” requiring legal review is not equivalent to a routine case merely because both close in the same system.

The U.S. Office of Personnel Management describes workforce planning as a systematic, continuous process for identifying the size and composition of a workforce needed to meet goals, assessing the current workforce, and calculating gaps. Use that logic for demand and supply alignment without treating a staffing guide as proof of a particular forecast. OPM Workforce Planning reference materials

Sources: [1]

Build a traceable demand model

Create a demand table with source, period, volume, work class, expected handling time, confidence, and owner. Use historical records where they are relevant, but record known changes such as launches, policy deadlines, seasonality, channel shifts, or planned adoption. If history is absent, use a comparable service or a bounded range and label the assumption.

GOV.UK's Service Manual recommends estimating demand by enquiry type and channel, using historical or comparable-service data, and considering handling time and staffing. That is a practical model for separating workload rather than applying one average to everything. GOV.UK: set up and manage user support

Model at least three scenarios:

For each scenario, state which input changes and why. Avoid applying a blanket percentage uplift when the real driver affects only one work class.

  • Reference: the most supportable current estimate.
  • Pressure: a plausible higher-volume or higher-complexity case.
  • Disruption: a material constraint such as reviewer absence, provider outage, new control, or rework spike.

Sources: [2]

Calculate usable capacity, not roster hours

For each resource pool, record available people or system units, working time, planned absence, non-delivery work, skill restrictions, review ratios, and a prudent operating buffer. Calculate:

usable capacity = available time − planned absence − fixed duties − expected interruption − buffer

Then convert usable capacity into work units using observed or tested handling time for the relevant class. Keep throughput from AI-assisted drafting separate from human verification and approval. Faster drafting can move the bottleneck to source checking, specialist review, or final authorization.

List every constraint explicitly: only two reviewers may approve regulated claims; one data source refreshes weekly; the model provider has a rate limit; work containing restricted information cannot use the standard assistant; or a new team needs supervised practice. A constraint register is more useful than an unexplained utilization target.

Use AI without surrendering authorship

AI can help normalize input tables, detect missing periods, produce alternative scenarios, identify contradictory assumptions, draft narrative explanations, and test arithmetic. Give it structured, approved inputs and require it to cite the row or assumption behind every output. Do not ask it to invent missing volumes, skill availability, absence, or productivity.

Retain the prompt or transformation specification, source snapshot, calculation workbook, generated comparison, reviewer corrections, and approved version. Recalculate important totals independently. When AI proposes a relationship—such as a cause for rising handling time—label it as a hypothesis until the operational owner verifies it.

NIST's AI RMF Core calls for documented roles, monitoring, mapped limitations, and risk management throughout the AI lifecycle. Apply those principles to the planning workflow: name who owns the model, record limitations, and review the plan when systems or conditions change. NIST AI RMF Core

Sources: [3]

Choose actions with thresholds

For every projected gap, specify an action, lead time, owner, trigger, and reversal condition. Options include reprioritizing work, changing service levels, cross-training, scheduling specialist review, simplifying an approved process, adding temporary capacity, deferring scope, or increasing system capacity. Do not treat hiring as the automatic answer to every gap.

Record thresholds before the period begins. For example: add a second review window when the seven-day queue exceeds 36 complex cases; pause low-priority intake if specialist-review demand exceeds 85% of usable capacity for two weeks; revisit the handling-time assumption when the median changes by more than 15% in a representative sample.

Worked plan: proposal review desk

A proposal desk expects 48 submissions next month: 30 standard, 12 complex, and six requiring specialist claim review. Historical medians are 2.0, 4.5, and 7.0 hours respectively. The initial total is 156 hours. The planner then adds 18 hours for scheduled calibration and 12 hours for predictable rework, producing 186 hours of demand.

Four reviewers provide 512 roster hours. After planned leave, meetings, administration, coaching, and a 12% interruption buffer, usable review capacity is 318 hours. The aggregate number looks comfortable, but the specialist register shows only one qualified reviewer with 52 usable hours; the six specialist cases require 42 hours before any escalation or rework.

The reference scenario is feasible with a reserved specialist block. The pressure scenario adds three complex submissions and two specialist cases, raising specialist demand to 56 hours and breaching the pool. The plan therefore assigns cross-review of two cases to an approved second specialist, schedules an early evidence check, and sets a trigger to defer non-critical advisory work if specialist demand exceeds 48 hours. AI helps compare scenario tables and flags a mismatch between the narrative and workbook; the owner corrects the narrative before approval.

Apply an acceptance checklist

Approve the plan only when the planning decision, scope, units, input sources, scenario assumptions, calculations, resource restrictions, bottlenecks, actions, triggers, owners, and refresh conditions are present and reproducible. A second person should reproduce one demand total and one capacity conversion from the retained inputs.

Two Prune can provide a controlled work simulation showing how a participant frames a capacity problem, uses evidence, works with AI, checks calculations, and communicates a plan. It does not validate the organization's forecast, supply staffing data, prove productivity gains, certify compliance, or make employment decisions.

Sources

  1. 1.OPM Workforce Planning reference materials · U.S. Office of Personnel Management
  2. 2.GOV.UK: set up and manage user support · UK Government
  3. 3.NIST AI RMF Core · National Institute of Standards and Technology

Explore this topic