PRIISM — The Financial Authority Layer

Executive Summary

PRIISM is the financial authority layer of the AIPM Toolkit: it tests whether an AI initiative is financially defensible at every stage, from selection through retirement. At every gate it asks one question: is this still financially defensible, right now, at this stage? The default answer is no. An initiative does not continue because it was approved once; it earns yes at every stage, on the evidence, or it stops.

Five elements are judged at every gate: a Probabilistic Gate of five viability questions, one named owner with the authority to stop (Responsibility), an Investment test anchored on the downside, the full lifecycle cost (Sustainability), and a funded plan against Model decay. The second I in the name is the Interlock, the mechanism that runs the gate the same way regardless of who operates it. The rule sets the verdict; one named person owns and acts on it.

The discipline that is rare rather than borrowed comes from how disciplined traders manage risk: the exit is agreed before the position is taken. PRIISM applies it through pre-committed stop conditions, signed before any funding is released, so no initiative survives on sunk cost alone. A one-page visual companion, the PRIISM Master Infographic, presents this document’s content in a single view.

1. The One Question and the Default

At every stage of an AI initiative, PRIISM asks one question: is this still financially defensible, right now, at this stage? The default answer is NO. Past approval is not authorization; the default resets to no at every gate. PRIISM maintains a continuous financial review of an AI initiative across its full lifecycle, anchored on the downside, with the stop lines agreed before funds are committed.

2. The Six Letters: Five Elements and the Interlock

2.1 P · Probabilistic Gate

The five viability questions, each answered yes on evidence: a quantified problem, AI justified over a simpler option, the pessimistic case clears the hurdle, one owner named, and the data exists.

2.2 R · Responsibility

One named person is accountable for the investment, holds written authority to stop the initiative, and a named adoption owner is in place.

2.3 I · Investment

The probability-weighted expected value must clear the hurdle rate, and the worst case must stay above the loss floor. The pessimistic number is always shown on its own, so the downside is never hidden inside the average. The hurdle and floor are illustrative defaults; every adopting organization sets its own.

2.4 I · Interlock

Not a risk it checks: the second I is the mechanism that runs the gate. It is described in Section 3.

2.5 S · Sustainability

Can the organization afford to run and maintain the system, not just build it? The full lifecycle cost is present, with nothing hidden or deferred.

2.6 M · Model decay

AI models can lose accuracy over time, and it cannot be known in advance whether a given model will. Retraining is funded from day one as insurance, and any decline is kept inside an agreed limit.

No averaging. Each element passes or fails on its own; a strong result on one element never compensates for a failure on another. A single failure holds the gate at no.

3. The Interlock: Four Locks

The Interlock forces the four contributing functions’ evidence to combine into one verdict and runs the gate the same way regardless of who operates it. The five elements judge the initiative; the Interlock runs the machine. It operates through four named locks. The Evidence lock: inputs are confirmed in advance, or the gate does not convene. The Format lock: handoffs arrive usable, or they are flagged. The Conflict lock: when the contributing functions disagree, the conservative value governs. The Currency lock: the alignment record is kept current, and a stale record is itself a flag.

4. The Gates and the Five Verdicts

The one question is asked at every gate across the lifecycle: Gate 1, the business case; Gate 2, data and build; Gate 3, pre-deployment; Gate 4, go-live; and the live-operations reviews at 30, 60 and 90 days, six months, and the annual re-buy.

Every gate produces one of five verdicts. APPROVED: all five elements pass; the next gate still opens at no. CONTINUE WITH CONTROLS: caution signals; named mitigations, owners and dates; also the highest verdict available when Sustainability, Model decay or adoption owner check fails. REWORK OR DEFER: a fixable failure; the initiative pauses and a return condition is named. STOP: before or during delivery, an element fails and cannot be resolved; work ceases and the remaining budget is released. RETIRE: after go-live only; a live system no longer clears the hurdle and is decommissioned on a plan, not halted abruptly. A quick-reference table is given in Appendix B.

5. What Is Borrowed, and What Is Rare

PRIISM does not claim to invent. Stage gates, a single accountable owner, three financial scenarios, and model-drift monitoring are proven disciplines, drawn from project governance, capital budgeting, and machine-learning operations. The rare move is borrowed from how disciplined traders manage risk: a trader sets the exit before entering the trade, while judgement is clear, so that emotion and capital already committed cannot rescue a losing position later. PRIISM brings that same discipline into AI through the pre-committed stop conditions.

6. The Pre-Committed Stop Conditions

Before any funding is released, the named owner signs the exact conditions that will later stop or re-open the initiative: the specific lines, agreed in writing while judgement is clear. Hard lines are absolute; cross one and the initiative stops. Trend lines are slopes, not walls; breached across consecutive reviews they return the initiative to the funding gate, so a single poor month never ends a healthy initiative. Because the owner agreed to both before any funds were committed, there is no argument when one is crossed. This removes the strongest pressure to continue a failing initiative: the sunk cost.

7. Why It Is Not One Person’s Opinion

Two things are true at once, and they are not in tension: the rule sets the verdict, so it is not one person’s opinion, and one named person is accountable for that verdict and holds the authority to act on it, so it is never a faceless committee. The first protects against bias; the second against no one being answerable. The criteria are explicit, so the same evidence produces the same verdict no matter who applies it. The verdict is fixed by the evidence; a named owner may override the resulting action, but only in the open, with the element, the reason, a named risk owner and a mitigation on the record. At Gate 1, before funding, there is no action to override; the only cure for a failed viability check is the missing evidence. An undocumented override fails Responsibility and blocks the next gate from opening. No individual is the framework.

8. Gate 1 Stop Conditions: the Instrument

The instrument below is signed while judgement is clear, before any funds are committed. Every line is sourced from the signed business case, the cost model, the benefits baseline, or the monitoring KPIs, or it carries a labelled illustrative default. An unsourced line is invalid; blanks are set by the named owner, never assumed.

8.1 The invalidation condition

The single condition under which this investment would no longer be defensible is: _________________________________________________

If no such condition can be named, the initiative does not proceed: an investment case that cannot state what would invalidate it has not been tested.

8.2 Absolute lines

One breach acts immediately. Used only where failure is binary.

8.3 Trend lines

Strike one is a logged flag with a named fix; strike two the next period is a formal warning; strike three returns the initiative to the funding gate. Every strike carries an owner and a mitigation.

8.4 The confidence rule

Every number carries a tag set by the function that supplied it. High is hard evidence. Medium is a sound estimate. Low is a working assumption. A pass built on a Low number is a pass on probation, and it automatically tightens the review cadence until the evidence firms up.

8.5 Signature

9. The Instrument Applied: the LexAI Case

The stop conditions were applied to a real business case, LexAI, a legal due-diligence assistant: net benefit 432,788 against a lifecycle cost of 19.0M, an 88 percent recall requirement, a base of 100 deals and 35 active users, tested against the illustrative 8 percent hurdle and the −2M loss floor. The finding was not in the filled cells but in the blanks: where the case could not state a line, the blank itself exposed what the investment case had not yet tested. The full worked example is carried in the Gate 1 Stop Conditions document.

Appendix A. Glossary

Appendix B. Gate and Verdict Quick Reference

Appendix C. The Four Locks Quick Reference

Visual Companion — PRIISM Master Infographics