Portfolio Investment Models for AI: When the ROI Curve Is Exponential
An investment that looks like the worst performer in the portfolio at month nine can be the best performer in the portfolio by month eighteen, and the stage gate in between cannot tell the difference.
Executive Summary
Most portfolio investment models were built for a world of linear value delivery. A capital project, a systems upgrade, a process improvement initiative — each of these tends to return value in rough proportion to effort expended. Spend the first quarter of the budget, get roughly a quarter of the benefit. This assumption is baked so deeply into portfolio governance that few practitioners ever question it. Stage gates, benefit-realisation curves, traffic-light reporting, the entire discipline of portfolio prioritisation — all of it quietly assumes that value accrues at a pace that tracks spend.
Artificial intelligence investments break this assumption, and they break it in a specific and predictable way. AI initiatives — particularly those involving large language models, agentic workflows, and the data and integration work that surrounds them — follow an exponential value curve. Early stages consume disproportionate effort for very little visible return: data has to be cleaned, models have to be evaluated, workflows have to be redesigned, and people have to be retrained. Then, often abruptly, capability compounds. The same model that answered questions passably in month three is orchestrating multi-step workflows by month ten, and the return accelerates far faster than the underlying spend.
The consequence is a governance failure mode that is now recurring across the industry: AI investments look like failures at every early stage gate, get treated as underperforming, and are starved of funding or cancelled outright at precisely the point they are about to become the best-performing line in the portfolio. This paper sets out why the exponential curve occurs, why conventional stage-gating actively selects against it, and what a portfolio investment model needs to look like to fund AI properly. It closes with a specific set of recommendations for portfolio directors and investment committees: separate evaluation criteria for exponential-curve investments, milestone design that measures compounding rather than output, staged funding structures that tolerate a longer proving period, and a portfolio-level allocation that treats a small number of exponential bets as a distinct asset class rather than folding them into the same gates as everything else.
None of this argues for suspending discipline. It argues for the opposite: a more rigorous discipline, tailored to the shape of the curve actually in front of us.
The Shape of the Problem
Portfolio management as a discipline emerged to solve a resource allocation problem: too many candidate investments, too little capital, and a need for a defensible, comparable way to choose between them. The tools built for this — net present value comparisons, stage-gate reviews, benefit-realisation tracking, RAG status reporting — all share an implicit model of how value arrives. They assume that if an investment is going to deliver, it will show visible signs of delivering early, and that the rate of visible progress is a reasonable proxy for the rate of underlying value creation.
That proxy holds reasonably well for most traditional investment types. It does not hold for AI.
Why AI Investments Are Different
There are structural reasons the AI value curve looks different from almost anything portfolio offices have funded before.
- The groundwork is invisible and unglamorous. Before any AI system produces a usable output, an organisation typically has to address data quality, access, and structure; this work produces no visible output and is easy to mistake for stalling.
- Data cleansing, entitlement mapping, and pipeline construction do not show up as demonstrable capability, so status reports on these investments read as “amber” for long stretches even when the work is on track.
- Capability improves in steps, not increments. Model and workflow performance does not creep upward smoothly; it tends to sit flat while underlying components (retrieval quality, prompt structure, orchestration logic, guardrails) are tuned, then jump when several of those components cross a usability threshold together.
- Organisational learning is a hidden multiplier. The people using the system are themselves on a learning curve — discovering what to ask, what to trust, and what to hand off — and this human-side adoption curve compounds with the technical curve rather than running in parallel to it.
- Early outputs understate later capability. An AI agent that can answer a narrow question today may, with the same underlying model and three more months of workflow integration, be handling an entire multi-step process — but nothing in a month-three demo signals that trajectory.
“An investment that looks like the worst performer in the portfolio at month nine can be the best performer in the portfolio by month eighteen, and the stage gate in between cannot tell the difference.”
The Governance Mismatch
Stage gates were designed to kill bad investments early and cheaply — a genuinely sound principle for the majority of a portfolio. The difficulty is that the criteria used to identify a bad investment (slow visible progress, benefits below plan, repeated slippage against milestones) are exactly the symptoms an exponential-curve investment will display in its early phase, indistinguishable from genuine underperformance. A portfolio office applying uniform stage-gate criteria across a mixed portfolio will therefore systematically defund its highest-potential AI investments while continuing to fund linear investments that are merely mediocre but predictable.
This is not a hypothetical risk. It is the single most common pattern observed across organisations attempting to govern AI investment through unmodified traditional portfolio processes: promising initiatives cancelled at the exact trough of the curve, while safer, lower-ceiling projects sail through gate after gate on the strength of steady, forecastable — but modest — returns.
Why Conventional Models Fail Specifically
It is worth being precise about the mechanisms of failure, because the fix depends on identifying which assumption breaks.
- NPV and payback-period models assume a roughly consistent discount relationship between time and cash flow. An exponential curve inverts the normal risk-adjustment logic: the further out the return, the larger it becomes relative to near-term numbers, which standard discounting mathematics is not built to reward.
- RAG status reporting treats time-in-amber as a leading indicator of failure. For a linear investment this is often true. For an exponential investment, a long amber phase followed by a rapid green is the expected pattern, not a warning sign.
- Benefit-realisation tracking assumes benefits arrive in the phase in which they are planned. AI benefits frequently arrive later than planned but larger than planned, and a tracking model built around phase-by-phase realisation records this as delay and underperformance rather than as a different — and ultimately superior — profile.
- Portfolio prioritisation frameworks that rank by near-term ROI will structurally deprioritise every exponential-curve investment relative to every linear one, because near-term ROI is, by definition, the weakest part of the exponential curve.
- This is compounded when prioritisation is done annually: a single budget cycle rarely spans enough of the curve to capture the inflection point, so the investment can be cancelled in year one on data that would have reversed itself in year two.
| Dimension | Linear Investment Model | Exponential (AI) Investment Model |
|---|---|---|
| Early-stage signal | Predictive of eventual return | Weakly predictive; often misleading |
| “Stalled” period | Reliable warning sign | Expected phase (groundwork/threshold-building) |
| Benefit timing | Matches phase plan | Later than plan, often larger than plan |
| Risk of early cancellation | Low cost if wrong | High cost if wrong — kills investments pre-inflection |
| Appropriate gate cadence | Standard quarterly/stage gates | Milestone-based, compounding-indicator gates |
What a Fit-for-Purpose Portfolio Model Looks Like
The purpose of this section is not to argue that AI investments should be exempt from scrutiny. It is to argue that the scrutiny needs to test for the right thing — compounding potential rather than linear output — and that this requires deliberate changes to portfolio design, not just a softer attitude toward underperformance.
Recommendation 1: Classify Investments by Curve Shape, Not Just by Category
Rather than grouping investments purely by business unit or technology type, portfolio offices should classify each candidate investment by its expected value curve: linear, stepped, or exponential. This classification should be made explicit at business-case stage, with the investment sponsor required to justify which curve applies and why.
- Investments classified as exponential should be routed to a different governance track from the outset, not reclassified retroactively once they appear to be underperforming.
- Curve classification should be revisited at each gate, but the burden of proof for reclassifying an exponential investment as a failure should be higher than for a linear one, precisely because the early symptoms are similar.
Recommendation 2: Replace Output Milestones with Compounding Indicators
Standard milestones ask “what did we deliver.” Exponential-curve investments need milestones that ask “what is now possible that was not possible before, even if it has not yet been exploited.”
- Track leading indicators of compounding: data pipeline completeness, the number of workflow steps a system can now handle unattended, the rate of reduction in human review required, and adoption depth among the people who will ultimately extract the value.
- Treat a rising trend in these leading indicators as equivalent to financial benefit realisation for gating purposes, even where the pound-and-pence benefit has not yet materialised.
- Where possible, instrument the system itself to produce these indicators automatically, rather than relying on subjective sponsor reporting, which is where optimism bias and gate-survival incentives distort the picture most.
Recommendation 3: Fund in Stages That Match the Curve, Not the Calendar
Annual budget cycles are a poor match for a curve that may take five, six, or more quarters to inflect. Staged funding structures — commit a modest tranche to prove the groundwork, a larger tranche to prove the threshold crossing, and the largest tranche only once compounding is visible — allow the organisation to limit downside exposure without forcing a premature go/no-go decision at the worst possible point on the curve.
The single highest-value governance change available to most portfolio offices right now is simple to state and hard to implement: stop asking exponential-curve investments to justify themselves against linear-curve milestones. The gate criteria, not the investment, is usually what is broken.
Recommendation 4: Create a Protected Allocation for Exponential Bets
A portfolio that funds AI investments purely by competing them against every other line item, using identical criteria, will rationally starve them, because the near-term numbers will not compete. A small, explicitly ring-fenced allocation — a defined percentage of the portfolio reserved for investments with exponential curve classification — protects a handful of genuinely high-potential initiatives from being outcompeted on the wrong metric.
- This allocation should be sized deliberately small (in the range most organisations use for genuine innovation or venture-style funding), so that it does not become a loophole for undisciplined spending.
- Investments within this allocation should still face rigorous review, but against the compounding indicators described above, not against near-term financial return.
Recommendation 5: Build Explicit Off-Ramps, Not Just On-Ramps
Protecting exponential investments from premature cancellation is not the same as protecting them from ever being cancelled. Genuine failures do occur, and a portfolio model that cannot kill a bad exponential-curve bet is as dangerous as one that kills every good one too early.
- Define, at the outset, what a genuine failure signature looks like for an exponential investment — for example, leading indicators that remain flat well past the point where comparable investments have historically inflected, or adoption that continues to decline despite groundwork being complete.
- Distinguish this explicitly from a merely slow inflection, which is the expected pattern and not, on its own, a failure signal.
Objections Worth Taking Seriously
A reasonable portfolio director will raise two objections to this argument, and both deserve a direct answer rather than dismissal.
The first objection is that classifying an investment as “exponential” is exactly the kind of soft, self-serving claim that every underperforming sponsor will make to avoid scrutiny. This is a fair concern, and the answer is not to take sponsor assertions at face value but to require the compounding indicators described above as the evidence base for the claim, assessed independently of the sponsor’s own reporting. A genuine exponential curve leaves measurable traces in leading indicators; a sponsor covering for a failing project usually cannot produce them.
The second objection is that ring-fencing capital for exponential bets simply moves the problem — it protects some investments from competition while starving others that might also have merit. This is true, and it is precisely why the recommendation is to size the ring-fenced allocation deliberately small and to apply real scrutiny within it, rather than to exempt AI investment from portfolio discipline altogether. The goal is not less governance. It is governance calibrated to the actual shape of the return.
Conclusion
The organisations that will extract disproportionate value from AI over the next several years are unlikely to be the ones with the most enthusiasm for the technology. They are more likely to be the ones whose portfolio governance can tell the difference between an investment that is failing and one that is simply early on a curve that has not yet inflected. That distinction is not visible from the outside using the tools most portfolio offices currently apply. It becomes visible only when the governance model is redesigned around the shape of the curve itself: separate classification, compounding-based milestones, staged funding aligned to proof points rather than the calendar, a protected but disciplined allocation, and an honest, pre-defined test for genuine failure.
None of this is exotic. Venture investors have operated portfolios shaped this way for decades. What is new is the need to bring that discipline inside mainstream corporate portfolio management, where the muscle memory of linear stage-gating is deeply entrenched. The organisations that make this adjustment early will be the ones still funding their best AI investment in month fifteen, at the exact moment their peers have already cancelled the equivalent line in their own portfolio.