The Portfolio and the Bill

Before the detail — Artifact: the portfolio sheet. Status of what follows: the public inventory binding where M-25-21 applies; reusable guidance elsewhere.

Prompt first: draft the sheet, leave every hour blank

Here is ai-inventory.yaml [paste] and the register's current fields
per system [paste, or NOT YET RECORDED per system].

Draft the PORTFOLIO SHEET, one row per system, columns:
  owner (a role with a person in it, or UNOWNED) · tier · gate path
  · lifecycle state · valid-through · expires this quarter (Y/N) ·
  open conditions · exceptions outstanding · currency per 6.1's
  three checks (CURRENT / SUSPEND-DUE / CLOCK-OVERDUE / STALE) ·
  capability class (so duplicates
  surface) · claimed benefit measure from section 1 · realized
  value (last reading, or NONE)

Then a CAPACITY block with every hour figure as a bracket I fill:
  [H_light] [H_full]  reviewer hours per decision, by gate path
  [N_light] [N_full]  decisions due this quarter, by gate path
  [R]                 reviewer hours available this quarter
  [A] [V]             eval-authoring and reviewer hours per re-run

Estimate nothing I did not give you; every missing figure is
NEEDS-OWNER.

Hour figures are where an agent does the most damage: a plausible number in a capacity table looks like planning, and a leader will fund against it. The structure and the arithmetic are the agent’s; every hour, reading, and count comes from the reviewer bench, or stays a bracket.

Records are per system; the budget is per portfolio

The record is complete for one system. Leadership’s questions run across all ten: how many are gated, what expires this quarter, which have no owner, what is open past deadline, and where two systems do the same job.

MU-AI-005 summarizes procurement documents and MU-AI-008 summarizes meetings — one capability class, two vendors, two re-evaluation bills, and a question only the sheet can raise. The example record supplies one dated row: MU-AI-004, valid through 2027-02-27; the other nine read NOT YET RECORDED until their records exist. Cloud Modernization 5.3 (a separate training in this series) puts an owner on every system — and an owner line naming an office with nobody in the seat reads UNOWNED here.

A portfolio nobody can see is the audit finding “no organization-level view of AI risk,” made once about all ten systems at the same time.

Capacity in assessor-hours

Federal Delivery 1.3 treats assessor capacity, in hours, as the binding constraint — the same arithmetic governs release decisions, with your figures:

CAPACITY - this quarter (fill every bracket; none has a default)
  demand   = [N_light] x [H_light]  +  [N_full] x [H_full]
  supply   = [R]   reviewer hours available after their other work
  result   demand <= supply  the quarter's decisions land on time
           demand  > supply  a queue: valid-through dates pass
                             unrenewed; STALE becomes the default
  levers   fewer systems (1.3's STOP exit) · a lighter path where
           the tier honestly allows it · more reviewer hours ·
           never a longer valid-through to make the sum close

H. Lindqvist can review a bounded number of full-path decisions in a quarter while running Regulatory Affairs, and the sheet is where that number meets the decisions due.

The bill a provider sends when it retires a version

Every provider-pinned system carries a recurring cost the project budget never listed. When the provider retires that version — the provider’s published retirement policy sets how often, the retirement date the record names sets the deadline — the decision suspends on the new boundary and re-runs: [A] hours to replenish the sealed set if the gate’s set must change, [V] reviewer hours, and the risk acceptor’s time to sign again.

Off the sheet, the re-run lands as unplanned reviewer hours in the provider’s chosen quarter.

Realized value, read against the claim

Section 1 of every record names the expected benefit and its measure. Data Products 6.2 measures cost in your own environment rather than from a published figure, and a benefit claim is no different — the sheet carries the claim beside the last realized reading, and a system whose realized column reads NONE across consecutive reviews is a sunset candidate. M-25-22 (the White House budget office’s April 2025 memo governing AI purchases) says that where practicable agencies should determine sunset criteria; writing them at release is this training’s practice, not the memo’s deadline. This column reads them.

Lesson 1.2 owns the build, buy, or integrate checklist; this sheet is its acceptance test — whatever you buy must produce these columns from your register without re-keying.

Seam: portfolio reporting to the public inventory

The commercial starting practice is an internal portfolio report to a risk committee. The federal delta is the annual public use-case inventory: every use case with a high-impact flag, a retired entry reported once more. The handoff artifact is this sheet, with Meridian’s federal-task-order row exportable in the inventory’s schema. Not equivalent: the federal inventory is public and annual, and its high-impact rows carry fields a commercial sheet lacks — testing, impact assessment, independent review, monitoring, training, fail-safe, appeals, feedback.

Stop and escalate when demand exceeds supply and every honest lever is spent. Which business areas keep their AI is leadership’s decision, not the reviewer’s whose hours ran out.

Practice status — among organizations running several AI systems under signed release decisions, commercial and federal

Scale: required | common baseline | strong optional | reference-shop (seen only at organizations that publish their own practice) | emerging

Key takeaway

Records are per system; the budget, the reviewer bench, and the provider’s retirement calendar are per portfolio — the sheet, with NOT YET RECORDED and UNOWNED written where true. Capacity is Federal Delivery 1.3’s arithmetic with your figures in every bracket; the honest levers are fewer systems, a lighter path the tier allows, or more hours — never a longer expiry. The re-run a retirement forces is a recurring line; the realized-value column reads section 1’s claim against M-25-22’s sunset criteria. Lesson 6.4 takes one row through the whole loop.

LEADERSHIP DECISION   fund the reviewer hours the capacity block
                      shows or run fewer systems - and read the
                      realized-value column before renewing a claim
PRACTITIONER ACTION   build the sheet from the register with every
                      hour figure supplied by the bench, surface
                      duplicates and UNOWNED rows, price each
                      provider retirement as a re-run
SUCCESS MEASURE       every quarter's decisions due matched to
                      reviewer hours before the quarter starts; zero
                      valid-through dates passed for lack of capacity