Standard NPV calculations applied to AI projects produce numbers that are wrong in the same direction most time: they are too low. The reason is structural. NPV models a project as a stream of cash flows discounted at a hurdle rate, with cost decomposed at proposal time and benefits realized linearly across the projection horizon. AI projects do not behave that way. They build capability that compounds across future projects, unlock optionality the original project did not contemplate, and run on a substrate (token prices, model quality, eval tooling) whose cost decays at compound rates. A discounted-cash-flow framework calibrated for 2018 enterprise software systematically underprices these effects, and the committees using it underspend on AI as a result. This piece names the three sources of underpricing and proposes a three-component valuation: NPV plus capability premium plus optionality value.
This is a spoke under the AI project economics manifesto, which argues for an AI-specific economics framework. NPV is the single most common valuation tool used in AI investment decisions in 2026, and getting its limits right is the difference between a portfolio that compounds and one that stays small.
How standard NPV underprices AI projects
NPV’s two structural assumptions are that cost is decomposable at proposal time and benefits are linear over the projection window. Neither assumption holds for AI projects.
Cost decomposition at proposal time misprices AI projects because the cost basis itself is moving. A 2026 AI project priced today against the inference and tooling costs visible at proposal time will see those costs drop 30 to 50 percent over the next 18 months as token prices fall, eval tooling matures, and reusable platform components reduce per-project setup. Standard NPV uses today’s cost across the entire projection; AI economics has a cost decay component that NPV does not model.
Linear benefit assumptions misprice AI projects because the benefits are not linear. The first AI project a team ships is the most expensive in true cost; eval discipline gets built, observability stack gets installed, prompt registry gets stood up, senior judgment gets developed. The second AI project is meaningfully cheaper because those artifacts exist. The third project is cheaper still. The fourth project is faster than the team would have believed possible at proposal time of the first. Standard NPV evaluates each project against a fixed cost basis; AI economics has a capability premium that grows across the portfolio.
And NPV does not contemplate optionality at many. The Black-Scholes adjustment for real options exists in finance theory but is rarely applied in enterprise IT decisions, and rarely applied in 2026 AI committee reviews of the projects under their consideration. Yet AI projects routinely unlock optionality their proposal could not have predicted: a chat interface that turns into an agent, an eval suite that turns into a quality moat, an observability stack that turns into a debugging product. NPV has zero room for those upside scenarios.
The result is a systematic 20 to 50 percent underpricing of AI projects on a strict NPV basis, which translates into too few approvals, too small a portfolio, and a team that stays one or two years behind where the economics would justify.
Source 1: Capability that compounds
The mechanism. AI projects produce reusable capability artifacts: eval libraries, prompt registries, observability stacks, harness infrastructure, regression patterns, vendor relationships, internal expertise. Each artifact reduces the cost of the next AI project by some fraction. Across a portfolio of five to ten projects, the cost reduction compounds: the fifth project’s per-unit cost is meaningfully lower than the first, in the range of 30 to 50 percent for the right team.
Why NPV misses it. NPV evaluates each project against its own dedicated cost basis. The reusable capability is treated as either zero (most common, because the project’s NPV calculation does not have a category for “value to future projects”) or as a one-time charge against the first project that builds it (next most common, which over-charges the foundational project and ignores the compounding effect on subsequent projects).
Defensible adjustment. A capability premium of 15 to 30 percent on the first foundational project, applied as a positive adjustment to its expected value rather than a cost reduction. The premium captures the fraction of the project’s cost that is genuinely capability investment rather than feature delivery. Higher end (25 to 30 percent) for the first AI project a team ships; lower end (15 to 20 percent) for the third or fourth project where most foundational capability already exists.
What to verify. A named list of capability artifacts the project will produce; eval suite, harness, observability hooks, prompt registry entries, documented patterns. Each artifact has a stated reuse projection: how many of the next five planned projects will draw on it. Without the named list, the capability premium is unsupportable.
Source 2: Optionality that unlocks
The mechanism. AI projects routinely unlock follow-on projects the original proposal did not contemplate. A document Q&A system unlocks workflow automation. A chat interface unlocks an agentic workflow. An eval suite unlocks a third-party API offering. Optionality is the value of these unanticipated follow-on opportunities, weighted by their probability of materializing.
Why NPV misses it. NPV projects benefits from the explicit features the project will deliver. Follow-on opportunities that are not named in the proposal cannot enter the NPV calculation. Real-options finance has the analytical machinery to value optionality; enterprise IT decision processes do not use it.
Defensible adjustment. An optionality value of 10 to 25 percent of the project’s NPV, scaled by the project’s exposure to follow-on possibilities. High exposure (foundational platform projects, projects in fast-moving categories) gets 20 to 25 percent; low exposure (narrow point solutions in stable categories) gets 5 to 10 percent. The optionality is real but not infinite, and over-claiming it produces decision discipline as bad as ignoring it.
What to verify. A named list of three to five plausible follow-on opportunities the project enables, each with a one-line probability estimate and an order-of-magnitude impact. The list does not need to be comprehensive; it needs to demonstrate that follow-on optionality is real and not zero. The opportunity cost framework covers Axis 4 (locked optionality) as the negative side of the same logic; this is the positive side.
Source 3: Team learning curve
The mechanism. Teams shipping AI projects in 2026 are typically on a steep learning curve in eval discipline, prompt engineering, observability practice, and AI-specific architecture choices. Each project advances the team’s competence by some fraction. The compounding effect is a non-linear improvement in delivery quality, speed, and cost across the team’s first three to five AI projects.
Why NPV misses it. Learning curve effects are typically modeled, if at many, in manufacturing economics; not in software project NPV. The 2018 software equivalent (a team gets faster at building React apps over time) is real but small; the 2026 AI equivalent is large because the discipline is genuinely new and the gap between novice and competent practice is wide.
Defensible adjustment. Bundle the learning-curve effect into the capability premium (Source 1). Treating it as a separate adjustment risks double-counting; treating it as zero risks underpricing the first two to three projects in a team’s AI portfolio. The right framing is that the capability premium has both an artifact component (eval libraries, prompt registries) and a team-skill component (the engineers who can wield them); the premium covers both.
What to verify. A named statement of where the team is on the learning curve. First AI project gets the full premium because both artifact and skill compound. Fifth or sixth AI project gets a smaller premium because the team-skill component is already largely in place.
Source 4: Cost decay (token prices)
The mechanism. Inference token prices have decayed 60 to 90 percent over rolling 12-month windows from 2023 through 2026. Per-task quality at a given price point has improved on a similar trajectory. The cost basis a 2026 AI project is priced against is not a stable run rate; it is a snapshot of a moving target, and the future-period costs in the NPV projection should be discounted to reflect expected decay.
Why NPV misses it. NPV’s discount rate accounts for time value of money, not for substrate cost decay. The substrate of AI is decaying faster than the time value of money increases for any reasonable hurdle rate. NPV projections that hold inference cost flat across a three-year horizon overstate the present value of future costs by 30 to 50 percent.
Defensible adjustment. Apply a cost-decay schedule to the inference and tooling line items in the NPV projection. A defensible schedule for 2026: assume 30 to 40 percent annual decline in inference cost per unit of work, declining at 10 to 15 percent annual rate (i.e., the decay slows over time but remains significant). Apply only to inference and tooling lines; do not apply to senior engineering time, which is appreciating in 2026 rather than decaying.
What to verify. A separate cost projection for the substrate-decaying lines (inference, tooling) versus the substrate-stable lines (senior engineering, project management). The substrate-decaying lines get the decay schedule applied; the substrate-stable lines get the standard inflation/cost-of-capital treatment. The hidden cost of AI evals and why AI ROI calculators are wrong cover the cost-line decomposition this depends on.
The three-component valuation
The proposed valuation framework: AI project value = NPV + capability premium + optionality value.
NPV is computed as standard, with the cost-decay schedule applied to substrate lines (Source 4). This corrects the most mechanical NPV error; holding inference and tooling costs flat; and is implementable inside any standard finance template.
Capability premium is added as a positive adjustment of 15 to 30 percent of the project’s NPV, scaled by the project’s foundational status in the team’s AI portfolio. The premium covers Sources 1 and 3 (capability artifacts and team learning curve). It is supportable only when the project produces a named list of reusable artifacts and the artifacts have a stated reuse projection across the team’s next five planned projects.
Optionality value is added as a positive adjustment of 10 to 25 percent of the project’s NPV, scaled by the project’s exposure to follow-on opportunities. The adjustment covers Source 2. It is supportable only when the project names three to five plausible follow-on opportunities with one-line probability estimates and order-of-magnitude impact.
Three components, each with a defensible range, each with a verification artifact. The combined valuation is typically 30 to 60 percent higher than the bare NPV for foundational AI projects, and 10 to 20 percent higher for incremental projects later in a team’s AI portfolio. Both adjustments shrink to near-zero for the team’s tenth or fifteenth AI project, where most capability is already built and optionality is more constrained.
Worked example
A proposed AI project: $500K build, expected $250K annual benefit, three-year horizon, 12 percent discount rate.
Standard NPV. Benefits: $250K × 3 years discounted at 12 percent = roughly $600K present value. Costs: $500K build at year zero, plus $150K annualized recurring cost, three years discounted = roughly $865K present value. Standard NPV: -$265K. The project fails the standard NPV test.
With cost decay. Apply a 30 percent annual decay to the substrate component of recurring cost (assume 60 percent of the $150K is substrate-decaying). The substrate-decaying component shrinks materially across the three years. Recurring cost in PV terms drops to roughly $300K. Total PV cost: $800K. NPV with cost decay: -$200K. Still fails on bare NPV terms.
Add capability premium. Project is the team’s second AI project; named capability artifacts include extensions to an existing eval suite and additions to the prompt registry. Capability premium of 20 percent of NPV value generated by capability investment, computed as 20 percent of the $600K benefit PV = $120K.
Add optionality value. Project unlocks two named follow-on opportunities (workflow automation extension, internal-tool agentic capability) at moderate probability. Optionality value of 15 percent of NPV = $90K.
Three-component valuation. -$200K + $120K + $90K = +$10K. The project is approximately break-even rather than $265K underwater. The decision flips, defensibly, on the same project.
The arithmetic looks like a thumb on the scale. The discipline is in producing the artifacts that justify the adjustments; the named capability artifacts, the named follow-on opportunities, the substrate-decay schedule. Without the artifacts, the adjustments are unsupportable. With the artifacts, they are auditable.
Frequently asked questions
Isn’t this just adding fudge factors to make AI projects look better?
It is exactly the opposite. Standard NPV is a fudge factor against AI projects because it omits effects that AI economics demonstrably has; capability compounding, optionality, substrate decay. The three-component valuation makes those effects explicit and bounds them. A team applying the framework sloppily can produce inflated valuations; a team applying it disciplined produces valuations that are defensible against the sources NPV omits.
How is this different from a real-options analysis?
Real-options finance has the analytical machinery to value optionality precisely; this framework approximates it. The approximation is intentional: a full Black-Scholes-style real-options model is not realistic for an enterprise AI committee to implement project-by-project. The 10 to 25 percent optionality bound is calibrated to be defensible without requiring a quantitative finance team to produce.
What if the team has shipped 10 AI projects already?
The capability premium shrinks to 5 to 10 percent for mature teams, and the optionality value shrinks similarly because most optionality has already been claimed. The framework still applies; the magnitude of the adjustments scales with the team’s place on the AI portfolio learning curve.
How does this interact with the opportunity-cost framework?
The opportunity-cost framework adjusts a project’s value down by the displaced alternatives. The NPV-trap framework adjusts a project’s value up by the underpriced effects. Both adjustments are typically needed. A complete project evaluation produces a starting NPV, applies the three-component additions from this framework, then applies the opportunity-cost subtractions from the opportunity-cost framework.
What discount rate should AI projects use?
The same hurdle rate the organization applies to other strategic investments. The framework does not propose changing the discount rate. The substrate-cost-decay adjustment is separate from the discount rate and applied to specific cost lines, not to the rate itself.
What if the cost-decay schedule is wrong?
Reasonable disagreement on the decay schedule produces reasonable disagreement on the substrate-cost component, typically a 10 to 20 percent swing. The framework’s outputs are robust to that swing; the capability premium and optionality value tend to dominate the substrate-decay adjustment for most projects. Teams uncertain on the decay rate can use a conservative 15 to 20 percent annual schedule and still capture most of the underpricing correction.
Doesn’t capability investment get captured in the next project’s lower cost basis?
It does, and that is the point. The capability premium attributes the value back to the project that built the capability, rather than letting it accrue silently to the next project as an unattributed cost reduction. Without the attribution, the foundational project’s NPV underprices its true value, and the team systematically underspends on the foundational projects that make the rest of the portfolio possible.
How does this connect to the AI project economics manifesto?
The manifesto argues that AI projects need an economics framework distinct from 2018 software economics. NPV is the most-used quantitative tool inside that framework, and getting its limits right is a core requirement. The three-component valuation is the manifesto’s NPV adjustment in operational form: a defensible, implementable adjustment to standard NPV that captures the AI-specific effects the standard model omits.
Key takeaways
- Standard NPV systematically underprices AI projects by 20 to 50 percent because it omits four AI-specific effects: capability compounding, optionality unlocked, team learning curve, and substrate cost decay.
- The proposed three-component valuation is NPV (with substrate-cost decay applied) plus capability premium (15 to 30 percent of NPV for foundational projects, declining for later projects) plus optionality value (10 to 25 percent of NPV scaled by follow-on exposure).
- Each component is auditable: the cost-decay schedule is applied to named substrate lines, the capability premium requires a named list of reusable artifacts, the optionality value requires a named list of follow-on opportunities.
- The framework is most consequential on foundational projects (first three AI projects in a team’s portfolio), where the gap between standard NPV and true value is largest. It shrinks toward zero for mature portfolios.
- Combined with the opportunity-cost framework, the three-component valuation produces a complete project evaluation that captures both the upside NPV omits and the downside ROI overlooks.
NPV is a useful tool with structural limits when applied to AI economics. The three-component adjustment is the smallest defensible correction that captures the underpricing, without requiring a real-options finance team to implement. Committees using it produce 20 to 40 percent more approvals at the same risk discipline, and the approved portfolios compound faster.
Arthur Wandzel
