A Portfolio Office With No Office: Why Stopping Is Harder Than Ranking

The problem

At the end of March 2026, the UK's Government Major Projects Portfolio held 189 projects with a combined whole-life cost of £924.2bn, run by 20 departments and their arm's-length bodies¹ (official data, March 2026). Delivery confidence, the portfolio's own measure of how likely a project is to deliver its objectives to time and cost, was rated Green for 29 projects (15%), Amber for 109 (58%) and Red for 34 (18%), with 17 (9%) exempt¹ (official data, March 2026; most ratings are self-reported each quarter by the project's Senior Responsible Owner, and replaced by an independent rating where the central authority has reviewed the project in the previous six months).

Of the 42 projects that left the portfolio in the year to March 2026, seven were brought to early closure¹. Twenty-six left having reported delivery against their objectives, eight no longer met the portfolio's criteria and one was replaced by another project¹ (official data, 2025–26). The report stresses that a Red rating does not mean failure and often reflects the scale, complexity or early stage of a programme¹. That caution is fair, and it points to the gap this paper explores. The portfolio's main signal is a confidence rating, not a recommendation about what to keep funding.

Private companies show a similar pattern in a different currency. A McKinsey analysis of 1,616 US-listed multi-business companies between 1990 and 2005 found that, for one-third of business units, the capital received in a year was almost exactly the capital received the year before, with a mean correlation of 0.99³ (consultancy analysis of company filings, published March 2012; measured without a control group; more than three years old). Companies in the top third for reallocation earned on average 30% higher annual total returns to shareholders than those in the bottom third³ (same basis). In plain terms: most budgets are last year's budgets with small adjustments, and the companies that moved money more did better, although the analysis cannot show that moving money caused the better returns.

Why does stopping happen so rarely? McKinsey points to two forces: anchoring on last year's allocation, and internal politics³. In an executive game it ran, participants given the previous year's allocation aligned their choices far more closely to it than participants who were not³ (consultancy-reported; no figures published). Management research has a name for the tendency to keep backing a failing course of action: escalation of commitment⁶ (search extract).

The evidence gathered here comes from the United Kingdom and the United States only. The public-sector evidence is recent; the private-sector evidence is more than a decade old.

What already works

Forcing a decision when the signal turns bad

Executive reviews with the power to stop. From January 2010, the US Office of Management and Budget (OMB) ran "TechStat" sessions: face-to-face meetings with agency leadership intended to turn around or terminate IT investments that were failing or not producing results². By April 2013, OMB reported 79 TechStats covering 55 investments at 23 agencies² (regulator-reported, cited in a GAO audit of June 2013; more than three years old). In June 2015, following a GAO recommendation, OMB began requiring agencies to hold a TechStat on any investment rated high-risk (red) for three consecutive months² (official audit record). This is the nearest working precedent for the possibility: a rule that turns a bad signal into a mandatory decision meeting.

Rules that shift the burden of proof. McKinsey reports that when Lee Raymond was chief executive of Exxon Mobil, the planning team had to identify 3 to 5% of the company's assets for potential disposal every year, and divisions could keep them only by showing a credible turnaround plan³ (consultancy-reported, published 2012; period of the practice not stated).

Neither of these mechanisms uses AI. Both work by changing the default: an initiative must earn its place, rather than a challenger having to prove it should go.

Forecasting which initiatives are drifting

An operating early-warning tool. In 2025–26, the UK's National Infrastructure and Service Transformation Authority (NISTA) launched what it describes as a digital AI tool, the Early Warning System, which uses existing portfolio data to flag projects at risk of moving to a Red rating¹. NISTA says the tool is being embedded into review processes and has already proved helpful in forecasting project health, particularly for projects it is not actively monitoring¹ (operator-reported; no method or accuracy figures published).

Research forecasts of overruns. Machine learning, meaning statistical models that learn patterns from past cases, has been tested on project data. A 2024 Air Force Institute of Technology study of US Air Force military construction projects found that its best model predicted schedule overrun with 82% accuracy, 20% higher than the no-information rate (the accuracy of always guessing the most common outcome), and cost overrun with 70% accuracy, 11% higher than that rate⁴ (abstract; measured on historical project data, not in a live trial; master's thesis, March 2024). The author concluded that the models were not consistently accurate enough to recommend as an aid for decisions⁴.

In plain terms: the models did better than guessing, but not well enough for anyone to act on them alone.

What does not exist

The research found no system, in either sector, that continuously re-ranks a whole portfolio on value and evidence and issues recommendations to stop. NISTA's tool forecasts a rating; the report does not describe it ranking projects against each other or recommending closure¹. Joining forecasting, ranking and a stop recommendation into one standing process would have to be invented.

The possibility

Picture a portfolio office with no office. Instead of a team that assembles a quarterly pack and runs an annual prioritisation round, an agent, meaning an AI system that can carry out a sequence of tasks on its own such as reading, calculating and drafting, would watch the portfolio continuously. Each time new evidence arrived, such as a quarterly return, a revised cost estimate, a missed milestone, a benefits report or an assurance review, it would re-score every initiative against criteria the leadership had agreed in advance, re-rank the whole portfolio, and flag any initiative that had fallen below an agreed line.

For each initiative below the line, the agent would draft a stop recommendation: what the evidence shows, what continuing would cost, what stopping would release, and what would have to change for the initiative to keep its place. The memo would go to a named decision forum, which would have to accept, reject or defer it and record why. Over time, the record of overruled recommendations would itself become evidence of whether the people or the model had judged better.

Where AI would do the work. Three parts would depend on AI. Forecasting: machine learning models estimating each initiative's chance of overrunning or slipping to Red, as NISTA's tool already attempts¹. Reading: language models extracting evidence from unstructured material such as assurance reports, business cases and progress narratives, where many early warnings sit. Orchestration and drafting: an agent running the cycle and writing the first draft of each recommendation.

Where it would not. Much of the possibility is ordinary. The data plumbing that brings returns in on time and in a common format is conventional software. The ranking criteria and their weights are a policy choice made by leaders. The money, people and contracts released by a stop decision are handled by finance, workforce and commercial teams. The decision stays with accountable people; in UK central government, departments remain accountable for delivering their projects¹. On an honest reading, AI would be a significant but minority share of the system; the hardest parts are agreement and follow-through.

What would have to be true

1. The evidence would have to be timely and honest. A model that ranks on self-reported data inherits its biases. In the UK portfolio, most delivery confidence ratings come each quarter from the project's own Senior Responsible Owner¹. If a low score could trigger a stop memo, the incentive to report optimistically would rise.
2. Leaders would have to agree the ranking rule before seeing the ranking. Value is hard to score. Of 38 UK military capability projects with a combined whole-life cost of £305.1bn, only one reported benefits in 2025–26, which the report attributes to the difficulty of monetising national security benefits¹ (official data). Without an agreed way to compare unlike things, the ranking would be contested at the first unwelcome result.
3. Forecasts would have to beat current judgement, measurably. The research example above was better than guessing but not good enough to act on⁴. A recommendation resting on a forecast would need published accuracy, tested on the organisation's own past initiatives.
4. Every recommendation would have to reach a forum obliged to answer it. OMB's three-months-red rule shows what such a trigger looks like². Without an obligation to decide, a recommendation is one more report.
5. Stopping would have to be safe for the people who stop. McKinsey describes internal politics in which any leader who wins fewer resources than the year before is seen as weak³. If closing an initiative damaged careers, recommendations would be ignored however good they were.
6. The system would have to be open to scrutiny. The UK's Algorithmic Transparency Recording Standard was mandated across government departments on 6 February 2024⁷ (official policy). It applies to algorithmic tools that significantly influence a decision with public effect or that interact directly with the public, while analytical models supporting broad policy-making typically fall outside it⁷. Whether a portfolio-ranking agent would be in scope is not settled by the text; even if it were not, auditors, legislators and staff would expect to see why a programme was recommended for closure.

What could stop it

What the evidence shows

• Savings claimed from stop reviews were hard to verify. OMB reported almost $4bn in life-cycle savings from TechStats; GAO could not validate the figure and found supporting documentation for about $22.2m at the four agencies it examined² (GAO audit, June 2013).
• Reviews reached only part of the problem. GAO found that the number of TechStats held on at-risk investments was small compared with the number of medium- and high-risk investments².
• People turn away from algorithms after seeing them err. In five experiments, participants who saw an algorithm make forecasts were less confident in it and less likely to choose it over a human forecaster, even when they had seen it outperform the human⁵ (abstract; controlled experiments, published 2015).
• Oversight can backfire. In commercial lending, closer monitoring and changing decision makers reduced escalation of commitment, but some decision makers resisted downgrading borrowers in the face of the intervention and escalated their commitment further⁶ (search extract; field study, published 2002).
• Reallocation pays slowly. Over periods shorter than three years, companies that reallocated more delivered lower shareholder returns than more stable peers³ (consultancy analysis, 2012).
• A rating is not a verdict. NISTA reports that 18 projects moved from Amber to Green and one from Red to Green since its previous annual report¹ (official data, 2025–26).

Reasoning

The challenge question is whether such an agent would make good stopping decisions more likely, or simply produce more documents about them. Several things point to the second.

First, the bottleneck may not be information. The TechStat record suggests the hard part was holding enough reviews and showing that they changed outcomes, not spotting troubled investments². If an organisation already knows which initiatives are drifting and still funds them, a sharper ranking would not change the politics. It would make the gap between evidence and decision more visible, which some leaders might resist.

Second, the inputs would be gamed. Once a score can trigger closure, the people reporting into it have a reason to shape what they report. An always-on system would face that pressure continuously rather than once a year.

“On an honest reading, AI would be a significant but minority share of the system; the hardest parts are agreement and follow-through.”

Third, the risk would fall unevenly. A wrong stop that is acted on harms a team, a supplier and, in the public sector, the people a service was meant to reach. A wrong decision to continue costs money quietly. A system that is right on average could still be wrong in ways that land on those least able to argue back.

Fourth, simpler options exist. A mandatory review after three consecutive red ratings², or a rule that a fixed share of the portfolio must justify its place each year³, needs no AI at all. If an organisation would not adopt those rules, it is unlikely to act on an agent's advice. The honest test is whether the AI parts, mainly better forecasting and faster reading of evidence, would add enough to such rules to justify their cost and risk.

One design choice might change the balance. If the agent's main output were not "stop this" but "this is what continuing assumes", leaders would be asked to check assumptions rather than accept a verdict. That is a hypothesis, not a finding.

Signs to watch

• NISTA's next Major Projects Annual Report (the 2025–26 edition was published on 13 July 2026): whether it reports how the Early Warning System performed, and how many projects were brought to early closure.
• An Algorithmic Transparency Recording Standard record for the Early Warning System, or for any portfolio-ranking tool, in the GOV.UK repository.
• Completion of the reforms to the control systems governing major projects, which NISTA lists among its priorities for the year ahead¹.
• A published evaluation, with methods and accuracy figures, of any AI tool that forecasts delivery risk on a live portfolio.
• Any organisation, public or private, publishing a rule that links a forecast or score to a mandatory continue-or-stop decision.