Module 6: Governing the portfolio

Watch: the module in 1:08

Captions are on by default and can be turned off in the player. The full transcript is below.

Read the transcript

Good governance does not only approve investment. It moves money away from weak investments.

Every material workload should leave each review with one of four decisions: scale, hold, redesign or stop. Two tests decide it: the return on the scarce capacity it uses, and how strongly its value is known.

Match the proof to the money. Small, reversible bets can move on light evidence. Large, hard-to-reverse ones need a holdout or better.

In a portfolio review, realised value set beside claimed value is often the most important number on the page.

Write the stop rule before the money is spent, and name the owner who can invoke it. Value is owned by the business owner, not by IT or FinOps.

In the exercise, the cheapest model per attempt turns out to be the most expensive per outcome. Then it is on to the capstone, where you bring all six modules together.

By the end of this module you can

Apply scale, hold, redesign or stop using return on scarce capacity and evidence level.
Match the evidence required to the size and reversibility of a commitment.
Write exit criteria and tell rational persistence from escalation of commitment.
Assign accountability for AI economics across finance, business owners, platform, FinOps and line managers.

Technology leadership and advisor lenses: the RACI and the review agenda are the operating model you will be asked to set up.

The argument in brief

Good governance does not only authorise investment. It moves resources away from weak investments. Most AI portfolios accumulate rather than curate: initiatives are stopped by attrition when a budget runs out or a sponsor moves on, rarely against exit criteria agreed before they started.

This module brings the course together. Each material workload ends each review in one of four decisions. Two tests decide it: return on the scarce capacity it uses, and how strongly its value is known. The portfolio review puts cost, claimed value, realised value and evidence side by side. The operating model says who holds each part of the chain, so that no single function becomes the bottleneck.

Key ideas

Scale: high value, low load, evidence at the level the commitment needs. Hold: promising, evidence too weak, named evidence and a date. Redesign: high value, high load. Stop: low value, high load, or evidence that does not hold up.
Proposed default: E0 or E1 for small reversible work with a cap, E2 for material recoverable commitments, E3 for large or hard-to-reverse ones.
Write the stop rule before the money is spent, and name the owner who can invoke it.
Realised value beside claimed value is often the most important number in a portfolio review.
Value is accountable to the business owner, not to IT or FinOps. Stop decisions are proposed by the owner and accountable to leadership.

Terms to know

Return on scarce capacity: Expected net AI value divided by the load a project places on the binding constraint, such as review hours, change slots, risk capacity, compute budget or executive attention.
Exit Criteria: Pre-agreed conditions under which an AI investment will be stopped, paused, or redesigned rather than continued.
Stage Gate (AI Investment): A defined decision point in an AI investment lifecycle at which continuation of the investment, and access to the next tranche of resources, requires meeting a specific evidence threshold.
Governance Theater: Governance processes that generate reports, reviews, and documentation without materially changing investment decisions or operating behaviour.
Value Realisation: The process of converting expected benefits from an investment into measurable business outcomes.

Read

In this order. Each opens on the main site.

Worked example: a five-initiative portfolio review

A hypothetical portfolio, money in £ thousands a year.

Steps of the worked example
Coding assistant	Cost 98, claimed 108, realised 26, E2. Hold: released time has not been shown to convert.
Workforce assistant	Cost 420, claimed 300, realised 0, E1. Hold: fund the evidence before the expansion.
Contact-centre summarisation	Cost 180, claimed 240, realised 60, E2. Redesign: review effort cancels much of the saving.
Contract review assistant	Cost 75, claimed 160, realised 120, E3. Scale and keep measuring.
Customer-service agent pilot	Cost 60, claimed 90, realised 0, E0, no named owner. Stop and record why.

The result

£833k of annual cost against £898k claimed, of which only £508k sits at E2 or above and may be monetised. £206k is realised, a net position of −£627k. 42% of spend is at E2 or above.

Realised value set beside claimed value is often the most important number on the page.

From The AI Portfolio Review Pack, where the figures are explained in full.

Try it: redesign before you scale: cost per successful outcome

A support team drafts 50,000 replies a month. A reply that fails the quality floor costs £3.00 of agent time to fix, and every reply gets a £0.10 automated check. Four setups have been piloted on the same sample.

Figures for the exercise
Frontier model	£0.40 an attempt, 92% clear the floor
Small model	£0.08 an attempt, 70% clear
Mid-tier model	£0.15 an attempt, 88% clear
Small model routed to mid-tier when unsure	£0.12 an attempt, 90% clear

Enter the setups, or open the calculator with them loaded.
Rank the four setups by cost per successful outcome.
Explain why the cheapest model per attempt is the most expensive per outcome.
The team currently runs the frontier model and has E2 evidence for its value. Which decision applies, and what would you ask for before the next review?

Open the cost per outcome calculator with these figures(opens in a new tab)

Show the worked answer

Routed £0.52, mid-tier £0.61, frontier £0.74, small model £1.08 per successful reply.

The small model fails 30% of the time, and each failure costs £3.00 to fix, which outweighs its lower price per attempt.

Redesign: the value is there, but the workflow route is costing more than it needs to. Move to the routed setup, confirm it clears the quality floor on live traffic for a set period, then decide again. The owner should name the evidence and the date.

Redesign is a decision, not a delay. Name the constraint, change it and decide again.

Check your understanding

Five practice questions. The final exam uses different questions in the same style.

Progress is saved in this browser only.