Picture your last quarterly review.  Fourteen AI initiatives on the slide.  All of them amber-green and none of them finished.  None of them killed.  Everyone nods and the meeting moves on.

Now reflect on the industry analysis.  The global research and advisory firm Gartner describes 2026 as the year “AI is in the Trough of Disillusionment”, noting that CIOs still struggle to prove value from AI investments and demonstrate tangible business outcomes (Gartner, 2026a; 2026b).   You may also have seen the widely quoted MIT-NANDA claim that 95% of organisations studied were getting zero return from their GenAI investments (Challapally et al., 2025).  That figure is contested by critics who have questioned the sample and the six-month observation window (Futuriom, 2025; UC Berkeley Executive Education, 2025) and the authors themselves acknowledge these limitations.  But notice what nobody disputes: spend is compounding far faster than demonstrated returns.

Why?  Because most organisations have quietly become excellent at starting AI initiatives and remain terrible at stopping them.

The zombie pilot

A pilot with no owner and no exit criteria is not an investment but hope dressed as governance.  Zombie pilots are not born; they are made.  They are funded on enthusiasm, measured on activity and never given a condition under which anyone would pull the plug.  Six months later they are still “showing promise”, still consuming budget, engineers and executive attention – and still contributing nothing to the P&L.  Multiply by fourteen and you have a portfolio whose expected return is roughly zero – not because AI does not work, but because nothing was ever run as a portfolio.

What actual investors do

Compare that with how professional investors behave.  Venture capital assumes most bets will fail.  The discipline is not in avoiding failure but in refusing to keep paying for it: place multiple small bets, kill the losers fast, and double down on the winners.  This is precisely the model that current portfolio guidance recommends for the age of AI, in place of annual planning cycles that produce strategies which are outdated before they are executed.

Let us be clear about what this does not mean.  The objection is not to the number of bets.  AI has collapsed the cost of exploration – a working prototype that once took a quarter can now be built in days – so running more cheap experiments is entirely rational.  The objection is to bets with no exit criteria for stopping them.  Twenty pilots with written ‘kill conditions’ is a portfolio. Twenty pilots without them are a subscription to hope, at scale.

The conditions for killing a pilot must be written at funding time, because that is when it is the cheapest and easiest time to do so.  Before any money flows, “we stop if adoption is below X by month six” is an easy sentence to write.  A year in, the same sentence must fight sunk costs, personal reputations, and a steering committee that approved the thing. That’s a losing proposition and a win for the zombie pilot.

Where is your kill budget?

Here is the uncomfortable question.  Every organisation has an innovation budget.  Very few can point to a kill budget – often known as ‘Horizon 0’.

Portfolio investment thinking distributes funding across horizons: exploratory options (Horizon 3), emerging value (Horizon 2), and the core business (Horizon 1) that pays for everything (Baghai, Coley and White, 1999).  Modern portfolio guidance adds a fourth, often-forgotten horizon: retiring and decommissioning products that no longer earn their keep – Horizon 0 – explicitly to free up money and people for more valuable work (Scaled Agile, 2026c).

The AI twist makes this sharper.  Because exploration is now faster and cheaper, promising initiatives arrive at the core sooner than before – which means marginal performers must leave sooner too.  The guidance is explicit: the accelerated arrival of new opportunities requires accelerating decommissioning, with AI itself reducing the cost of migration and retirement (Scaled Agile, 2026a). One could argue that the entire solution lifecycle is accelerated.  As a result, a portfolio that only ever adds is not a portfolio – it is a hoard.

You cannot kill what cannot fail

There is a deeper reason zombies survive, and it hides inside the business case.

“This initiative will deliver €2M in savings.”  No range.  No confidence level.  No date on which forecast meets actuals.  Single-point estimates like this do not convey precision; they hide uncertainty and create false confidence, and decision science has warned about exactly this failure mode for years (Savage, 2009; Hubbard, 2014).  A business case that cannot fail cannot succeed either.  It is theatre – and you cannot hold an initiative accountable to theatre.

The fix is not more paperwork. It is three things: a forecast expressed as a 90% confidence range, a named owner and a review date on which actuals are compared to that range. When the actual falls outside the range, someone decides – maintain, increase, reduce, or stop (Scaled Agile, 2026b). That comparison is the kill condition, expressed as economics rather than governance ritual.

Two further traps deserve a warning.  First, when AI makes building nearly free, ROI percentages balloon and stop discriminating between investments; the discipline has to come from modelling operating costs, and for AI that means token consumption – we pay humans in salary, but we pay agents in tokens, and that cost scales with adoption (Scaled Agile, 2026a; 2026b).  Success, in other words, is exactly when AI gets expensive.  Second, the last increment of an ambitious target is often the value-destroying one: marginal analysis can show a healthy overall return concealing a final tier that will not break even for years.  Sometimes the most profitable decision a portfolio makes is to stop before the ambition runs out.

Discipline is speed

One final reframe – most organisations still budget annually: allocate 100% of the portfolio budget at the start of the year, then live with it.  A disciplined portfolio works differently. It withholds a small percentage of the total budget as a strategic reserve and reviews investment decisions quarterly rather than annually, comparing actual outcomes against the ranges in the original business case.  At each review, that reserve flows toward winners and away from initiatives that missed forecast or triggered their kill criteria.

This is not about adding bureaucracy. Quarterly reallocation with a liquid strategic reserve is faster than the annual budget cycle, precisely because it lets you kill sooner and double down sooner. The disciplined portfolio is not the slow one. It is the one that gets to reinvest the zombies’ budget while its competitors are still nodding at amber-green slides.

So, a question to end on: could you name three AI initiatives you would kill tomorrow – and who owns that decision?

References

• Baghai, M., Coley, S. and White, D. (1999) The alchemy of growth: kickstarting and sustaining growth in your company. London: Orion Business Books.

  Challapally, A., Pease, C., Raskar, R. and Chari, P. (2025) The GenAI divide: state of AI in business 2025.  Cambridge, MA: MIT NANDA.  Available at: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf.

  Futuriom (2025) ‘Why we don’t believe MIT NANDA’s weird AI study’, Futuriom, August.  Available at: https://www.futuriom.com/articles/news/why-we-dont-believe-mit-nandas-werid-ai-study/2025/08.

  Gartner (2026a) Gartner says worldwide AI spending will total $2.5 trillion in 2026.  Press release, 15 January.  Available at: https://www.gartner.com/en/newsroom/press-releases/2026-1-15-gartner-says-worldwide-ai-spending-will-total-2-point-5-trillion-dollars-in-2026.

  Gartner (2026b) Gartner forecasts worldwide AI spending to grow 47% in 2026.  Press release, 19 May.  Available at: https://www.gartner.com/en/newsroom/press-releases/2026-05-19-gartner-forecasts-worldwide-ai-spending-to-grow-47-percent-in-2026.

  Hubbard, D.W. (2014) How to measure anything: finding the value of intangibles in business.  3rd edn.  Hoboken, NJ: Wiley.

  Savage, S.L. (2009) The flaw of averages: why we underestimate risk in the face of uncertainty.  Hoboken, NJ: Wiley.

  Scaled Agile (2026a) ‘Portfolio outcomes and investment strategy’, Scaled Agile Framework.  Available at: https://framework.scaledagile.com/ain-safe-portfolio-outcomes-and-investment-strategy.

  Scaled Agile (2026b) ‘Return on investment’, Scaled Agile Framework.  Available at: https://framework.scaledagile.com/ain-safe-return-on-investment/.

  Scaled Agile (2026c) ‘Lean Portfolio Management Discipline’, Scaled Agile Framework. Available at: https://framework.scaledagile.com/lean-budgets/.

  UC Berkeley Executive Education (2025) ‘Beyond ROI: are we using the wrong metric in measuring AI success?’, UC Berkeley Executive Education, September.  Available at: https://exec-ed.berkeley.edu/2025/09/beyond-roi-are-we-using-the-wrong-metric-in-measuring-ai-success/.