Practical guide

AI capacity plan verification checklist

A capacity plan is useful only when another person can trace its demand, supply, calculations, constraints, and decision. This checklist is for reviewing a completed AI-assisted capacity plan before leaders use it to approve work, move people, or change service commitments. It complements the planning method: the plan proposes an allocation, while this review tests whether that allocation is.

By Two Prune

A capacity plan is useful only when another person can trace its demand, supply, calculations, constraints, and decision. This checklist is for reviewing a completed AI-assisted capacity plan before leaders use it to approve work, move people, or change service commitments. It complements the planning method: the plan proposes an allocation, while this review tests whether that allocation is inspectable and decision-ready. The review is educational. It does not certify staffing adequacy, programme performance, or compliance. Apply local finance, workforce, security, and employment requirements before acting.

Assemble the review packet

Do not review the presentation alone. Ask the plan owner for a small packet containing:

OPM describes workforce planning as a systematic process for identifying and addressing gaps between current and future workforce needs. The UK Government Service Manual separately distinguishes expected support demand from the people and time available to answer it. Together, these sources support keeping demand and supply explicit rather than hiding both inside one forecast number.

  • the dated demand register, including unit definitions and the forecast horizon;
  • the available-capacity register by skill or work type, with leave and non-project time;
  • the calculation sheet or reproducible model;
  • source links, extraction dates, named owners, and known gaps;
  • the proposed allocation, rejected alternatives, and approval record;
  • a log of AI use that identifies which steps produced suggestions or transformations; and
  • the monitoring trigger that will cause the plan to be revisited.

Run the seven-part checklist

Mark each item pass, repair, unresolved, or not applicable. A pass requires retained evidence, not confidence in the author.

### 1. Scope and decision

### 2. Source traceability

### 3. Calculations

### 4. Claims and uncertainty

### 5. Contradictions and constraints

### 6. Completeness and alternatives

### 7. Approval and monitoring

  • Is the exact decision named, along with its owner and decision date?
  • Are included services, teams, work types, locations, and time periods defined?
  • Are units consistent, such as hours, cases, or full-time-equivalent weeks?
  • Does the plan distinguish a planning assumption from an approved commitment?
  • Can every material demand and capacity input be traced to an owner-maintained record?
  • Does each source have an extraction date, version, and coverage note?
  • Are copied figures reconciled with the source total?
  • Are missing, stale, or manually adjusted values visible?
  • Can a reviewer reproduce the totals from the retained inputs?
  • Are gross hours reduced for leave, operational duties, meetings, and other unavailable time?
  • Are percentages applied to the correct denominator?
  • Are rounding, conversions, and duplicated rows tested?
  • Are AI-produced formulas inspected and recalculated independently?
  • Does the wording say “estimate,” “scenario,” or “observed” where appropriate?
  • Are ranges or sensitivities shown for the assumptions that materially change the decision?
  • Does the plan avoid presenting self-reported confidence as demonstrated capability?
  • Are unsupported causal claims removed or narrowed?
  • Do source records disagree about volumes, dates, ownership, or availability?
  • Are specialist bottlenecks visible even when total hours appear sufficient?
  • Are sequencing, access, review, and dependency constraints represented?
  • Does the plan state which conflict was resolved, by whom, and on what evidence?
  • Does every priority item have an owner, required skill, effort range, and dependency?
  • Are at least two feasible allocations compared using the same assumptions?
  • Are deferred items and service consequences recorded?
  • Is residual risk assigned rather than described without an owner?
  • Did the accountable owner approve the chosen scenario and its exceptions?
  • Is there a review date and a measurable trigger, such as demand variance or missed service level?
  • Can the next review compare forecast demand, available capacity, and actual throughput?
  • Is the final artifact accessible to the people responsible for delivery and review?

Worked artifact: test a support capacity plan

Consider a synthetic plan for a four-week customer-support pilot. The plan states that 1,200 cases require 600 hours and that six specialists provide 720 hours, leaving 120 spare hours.

The reviewer reconstructs the calculation. Six specialists at 30 hours per week for four weeks do equal 720 gross hours. The availability register, however, shows 72 hours of leave, 60 hours of mandatory operational duty, and 48 hours of quality review. Net capacity is therefore 540 hours, not 720. The demand model also treats every case as 30 minutes, while a current sample separates 900 standard cases at 24 minutes from 300 complex cases at 48 minutes:

| Test | Evidence | Result | | --- | --- | --- | | Demand | 900 × 0.4 + 300 × 0.8 | 600 hours | | Net capacity | 720 - 72 - 60 - 48 | 540 hours | | Gap | 540 - 600 | 60-hour shortfall | | Constraint | Only two specialists can handle complex cases | Repair required |

The plan fails the calculation, contradiction, and constraint checks. It is not repaired by changing the headline to “near capacity.” The owner compares three actions: narrow the pilot scope, shift trained capacity from another queue, or extend the response target for complex cases. The approved artifact records the selected option, the 60-hour gap, the two-person bottleneck, and a weekly trigger if incoming case mix differs by more than 10 percentage points from the sample.

Keep a review record

Store a one-page review record with the plan:

| Field | Entry | | --- | --- | | Plan and version | Name, owner, date, horizon | | Decision | What will be approved or changed | | Reviewer | Named person independent of the main calculation | | Dispositions | Pass, repair, unresolved, not applicable by section | | Repairs | Changed cell, claim, source, or assumption | | Residual uncertainty | Remaining gap and likely decision effect | | Approval | Owner, date, exceptions | | Next review | Date and trigger |

NIST's AI Risk Management Framework says responsibilities, documentation, monitoring, and review processes should be explicit across the AI lifecycle. For an AI-assisted capacity plan, that means retaining the human decision and evidence trail rather than treating generated prose or formulas as approval evidence. Two Prune can assess how a participant frames the planning task, navigates sources, checks calculations, uses AI, and communicates the recommendation. It does not establish whether a real organisation has enough staff or whether an allocation should be implemented.

Sources

US Office of Personnel Management, workforce planning reference materials

  • UK Government Service Manual, set up and manage user support
  • NIST AI Risk Management Framework Core

Sources: [1] [2] [3]

Sources

  1. 1.US Office of Personnel Management, workforce planning reference materials · U.S. Office of Personnel Management
  2. 2.UK Government Service Manual, set up and manage user support · UK Government
  3. 3.NIST AI Risk Management Framework Core · National Institute of Standards and Technology