Firm Valuation When AI Shapes the Business Model

Introduction

Artificial intelligence has moved from pilot budgets to strategic capital allocation. In the United States, AI-focused firms have reached private-market valuations of hundreds of billions of dollars within years of founding,(Anthropic 2026; Databricks 2025) while no European company created in the last fifty years has achieved a market capitalisation above EUR 100 billion.(Draghi 2024) The speed and magnitude of this repricing are not explained by technology alone. Value realisation remains highly heterogeneous across firms, sectors, and jurisdictions,(Singla, Alex, Sukharevsky, Alexander, Yee, Lareina, et al. 2025; Singla, Alex, Sukharevsky, Alexander, Hall, Bryce, et al. 2025; Brynjolfsson et al. 2023; Eisfeldt et al. 2023) and the pattern is consistent: most deployments remain at limited integration depth, while a smaller group restructures workflows, decision rights, and data feedback loops.(Singla, Alex, Sukharevsky, Alexander, Yee, Lareina, et al. 2025; Singla, Alex, Sukharevsky, Alexander, Hall, Bryce, et al. 2025) The difference between experimenting with AI and restructuring around it is a structural divide, and standard valuation practice has yet to absorb it. The standard discounted cash-flow framework values a firm as

The accuracy of any DCF valuation depends on the data available to the valuator at the valuation date, and that data is a function of operating history. 3 illustrates the valuator’s problem. Before founding, there is no observable data: projection uncertainty $\sigma$ is at its maximum. The act of establishing a firm marks a transition; from that point, costs, contracts, and revenue begin to accumulate, and $\sigma$ decreases as the projection gains empirical grounding. The longer the firm operates, the lower $\sigma$ becomes, but it never reaches zero because DCF remains a forecast at every point. When an operational firm introduces AI into its processes, $\sigma$ spikes: the business model is being reconfigured, and the historical patterns the valuator relied on lose predictive power. The spike is not gradual; at sufficient integration depth, it is discontinuous. As AI milestones clear, $\sigma$ decreases again, but the rate of recovery depends on the depth and scope of the integration, and $\sigma$ may not return to its pre-AI level because the business model has structurally changed.

What makes AI integration particularly resistant to ex ante valuation is the confluence of mechanisms it activates. Even within a given integration regime, deployment choices exhibit sensitive dependence on initial conditions: small differences in sequencing produce divergent firm-level trajectories because each decision reshapes the capability base available for subsequent ones.(Lorenz 1963) The resulting system exhibits non-linear feedback loops across functions, costs, and competitive position, characteristic of complexity-economic regimes where equilibrium assumptions break down and the distribution of outcomes is fat-tailed rather than normally distributed around a central projection.(Arthur 2021) Strategic interaction amplifies this further: platform economics, winner-take-most dynamics, and moat construction turn each firm’s AI decisions into moves in an evolving competitive game whose payoff structure changes as other players act.(Von Neumann and Morgenstern 1944; Rochet and Tirole 2003) For a valuator, these mechanisms operate simultaneously. The future configuration of an AI-integrated firm carries two stacked layers of uncertainty: magnitude uncertainty driven by the firm’s own decisions, and kind uncertainty driven by what competitors do.

Within DCF, three mechanisms compound the problem. First, staged optionality collapses into the long-term growth rate: a contingent, milestone-gated payoff that management may abandon or expand is priced as if it were a deterministic cash-flow trajectory. Second, risk concentrates in a single discount rate that cannot represent the structural shift from high uncertainty at early integration stages to lower uncertainty after milestone clearance. Third, management flexibility (the right to defer, expand, or abandon at each decision node) carries positive value that DCF assigns implicitly at zero.

U.S. private-market practice already reveals this tension. Venture rounds price preferred claims with contractual protections, producing headline post-money valuations that average roughly 48% above common-share fair value.(Gornall and Strebulaev 2020) Tax and disclosure workflows value common shares separately under dedicated standards such as 409A and pre-IPO scenario methods,(U.S. Treasury Department and Internal Revenue Service 2026; Maplebear Inc. 2023) so the same firm carries two structurally different valuations at the same date. The very existence of this dual-track architecture signals that practitioners already distinguish between option-loaded headline figures and fair-value estimates. A European practitioner reading a U.S. post-money figure as a common-equity equivalent makes a structural category error: the figure prices contractual optionality alongside operating forecasts.

In European and Swiss settings, the mismatch runs deeper. Dominant valuation standards such as IDW S 1 mandate discounted earnings as the primary method and require demonstrable historical patterns.(Institut der Wirtschaftsprüfer in Deutschland (IDW) 2016) EU and Swiss AI governance frameworks formalise compliance gates that affect execution cost, admissibility risk, and time-to-scale.(European Parliament and Council of the European Union 2024; Swiss Financial Market Supervisory Authority (FINMA) 2024; Swiss Federal Council 2025) These instruments were built for firms with established cash-flow histories, and for those firms they work well. The problem arises when the same instruments are applied to milestone-gated AI ventures: not as a deliberate analytical choice, but because no structured alternative exists within the European valuation ecosystem. The result is a measurement mismatch, not conservative valuation. European venture capital stands at less than one-third of U.S. levels as a share of GDP, and comparable startups are valued approximately seven times higher in the United States at the same development stage.(Arnold et al. 2024; Draghi 2024) Part of this gap reflects genuine differences in market structure and risk appetite. Part of it may reflect the absence of a valuation framework that can price what these firms actually are.

The core problem is not whether AI matters for firm value, but how to value integration pathways whose pay-offs depend on depth, sequence, and jurisdiction-conditioned feasibility. Existing literature provides the necessary components: real options for staged technology investment, multi-criteria methods for structured probability assessment, maturity models for depth classification, and market-pricing studies for empirical calibration.(Schwartz and Moon 2000; Benaroch and Kauffman 1999; Saaty 1980; Hansen et al. 2024; Eisfeldt et al. 2023) The unresolved task is to integrate these components into a single valuation logic that identifies the relevant milestones, specifies how each probability is estimated, and indicates where integration depth changes the parameter set. Within the present work, a structured, auditable protocol is developed that decomposes AI integration into option-level valuation inputs, comparable across analysts, applicable across integration depths, and consistent with existing valuation standards.

Related Work

Valuation research is still dominated by discounted cash-flow logic calibrated on relatively stable business models.(Koller et al. 2020) The young-firm literature shows why that architecture weakens under negative earnings, high failure probability, and limited historical depth: conventional DCF and relative valuation break down precisely where the firm profile is dominated by intangible investment, survival uncertainty, and non-linear growth trajectories.(Damodaran 2018, 2009) Intangible-intensive firms compound this problem because their core value drivers resist balance-sheet measurement.(Lev 2001; Haskel and Westlake 2017) The unresolved issue is not whether DCF remains useful, but where its assumptions become structurally incomplete for staged AI integration.

Market-based and early-stage methods face structurally different but equally limiting constraints. Revenue and ARR multiples, the dominant approach for private AI firm valuation in practice, compress option value into a comparable figure without identifying which milestone generates the premium or how sensitive the valuation is to individual probability assumptions; their application is also constrained by the scarcity of comparables for firms at Integration Depth Level (IDL) 2–3.(Damodaran 2009; Eisfeldt et al. 2023) Early-stage methods such as the VC method and First Chicago introduce scenario weighting but stop short of per-milestone probability estimation and depth-conditioned option architecture.(Lerner and Nanda 2020) The income approach under IDW S 1, the primary European statutory method, shares DCF’s structural limitation: it projects a single income path and discounts it, compressing milestone probabilities and continuation-option value into the same opaque aggregate parameter.(Institut der Wirtschaftsprüfer in Deutschland (IDW) 2016) The framework developed here does not replace these methods; it adds a decomposition layer in the form of option-level inputs with independent probability estimates linked to identifiable milestone assumptions.

Real-options research addresses staged commitment under uncertainty and supplies the formal language for sequential exercise, deferral, expansion, and abandonment.(Myers 1977; Dixit and Pindyck 1994; Trigeorgis 1996) Compound-option formulations extend this to multi-stage investment chains where each decision gate conditions the next.(Geske 1979; Cassimon et al. 2004) Applied work in internet and IT contexts confirms that technology investments are path-dependent and decision-contingent, with value concentrated in future optionality rather than static baseline projections.(Schwartz and Moon 2000, 2001; Benaroch and Kauffman 1999, 2000) This research programme has also extended to natural-resource investment, where empirical evidence confirms that firms hold loss-making assets open when waiting option value exceeds closure value,(Moel and Tufano 2002) and to strategic investment contexts where option value interacts with competitive pre-emption and deterrence.(Smit and Trigeorgis 2004) Across all applied domains, the consistent finding is that option value is substantial relative to static NPV and that standard discounted-earnings methods systematically underestimate flexibility value. These models treat technology generically: none addresses AI integration specifically or distinguishes integration depth as a structural variable that reshapes the option parameters themselves. Stage-level probabilities in applied implementations are typically derived from market-implied volatility or unstructured expert judgement, with no structured protocol for estimating each option’s likelihood independently or varying individual probabilities in scenario analysis.

AI adoption and maturity studies document heterogeneous operational effects and depth-dependent capability formation.(Hansen et al. 2024; Dreyling et al. 2024) Productivity evidence shows strong heterogeneity by worker skill and organisational complementarity,(Brynjolfsson et al. 2023) delayed measured gains from costly intangible co-investments,(Brynjolfsson et al. 2018) and market repricing that varies with AI exposure.(Eisfeldt et al. 2023) Industry surveys confirm near-ubiquitous experimentation but limited deep workflow redesign.(Singla, Alex, Sukharevsky, Alexander, Yee, Lareina, et al. 2025; Singla, Alex, Sukharevsky, Alexander, Hall, Bryce, et al. 2025) This stream identifies the empirical phenomenon (integration depth matters) but does not provide a valuation operator that transforms depth and milestone progression into auditable firm-value components.

Existing maturity scales classify depth within individual organisations; none distinguishes firms that embed AI into existing operations from firms whose core product is AI itself. The two cases generate structurally different risk profiles, option architectures, and competitive dynamics. No classification in this literature provides an industry-agnostic taxonomy that maps these distinctions to valuation mechanics.

Multi-criteria decision analysis provides structured tools for aggregating expert judgement under complexity. The Analytic Hierarchy Process decomposes assessment into pairwise comparisons with a built-in consistency diagnostic,(Saaty 1980) and has been combined with real options for telecom and IT infrastructure investment decisions.(Angelou and Economides 2008) Alternative methods (the Analytic Network Process, the Best-Worst Method,(Rezaei 2015) fuzzy extensions) relax specific AHP assumptions but share the core logic of structured criterion weighting. This stream has not been applied to AI-integration-depth-conditioned probability estimation: existing MCDM–real-options hybrids treat technology type generically without modelling how integration depth restructures the assessment criteria themselves.

Venture-finance and private-market evidence adds a further missing element: pricing mechanics. Staged financing is modelled as milestone-contingent commitment under uncertainty,(Lerner and Nanda 2020; Talmor and Cuny 2005) and security design drives systematic divergence between headline round valuations and common-share fair values in unicorn markets.(Gornall and Strebulaev 2020) Regulatory and tax frameworks formalise this divergence through distinct valuation tracks for preferred and common equity.(U.S. Treasury Department and Internal Revenue Service 2026; Maplebear Inc. 2023) In AI markets, winner-take-most dynamics and complementary-asset moats concentrate option value in a small number of firms,(Korinek and Vipra 2025; Azoulay et al. 2024; Rochet and Tirole 2003) compounding the disconnect between headline figures and operating-value fundamentals.

Cross-jurisdiction evidence indicates that valuation outcomes are shaped by financing ecosystems and regulatory institutions. European competitiveness diagnostics identify persistent scale-up constraints rooted in capital-market structure, exit availability, and risk-capital depth.(Draghi 2024; Arnold et al. 2024) EU and Swiss AI governance frameworks add a distinct layer: jurisdiction-specific compliance requirements that condition which milestones are feasible, at what cost, and on what timeline.(European Parliament and Council of the European Union 2024; Swiss Financial Market Supervisory Authority (FINMA) 2024; Swiss Federal Council 2025)

For the valuation framework, these requirements enter directly as $C_{\mathrm{r},s}$ in the $V_s$ decomposition and as constraints on milestone feasibility that vary by jurisdiction. A firm pursuing IDL 3 integration under EU AI Act oversight faces compliance gates that alter both the timing and the probability of milestone clearance relative to the same firm operating in a lighter regulatory environment. Regulatory jurisdiction must therefore enter the framework as a parameter that varies across cases.

No existing framework provides an industry-agnostic taxonomy that distinguishes AI integration from AI production, links both categories to milestone-gated option structures, and estimates each option’s probability independently so that scenario analysis can isolate individual sensitivities. The required synthesis spans depth-conditioned classification, real-options architecture, structured per-option probability estimation, security-design effects, and jurisdiction-conditioned feasibility. Each literature stream supplies components; none combines them into a single auditable valuation logic.

Contribution

DCF remains the necessary foundation for firm valuation; the issue is which protocol governs the assumptions embedded in it when AI integration is priced. The framework developed here decomposes the opaque growth-rate assumption into identifiable options and stated probabilities, thereby making the source of the valuator’s projection uncertainty ($\sigma$ in 3) explicit. The contribution is structural rather than predictive. The framework extends DCF with a decomposition protocol for AI-related inputs, without any claim to estimate option values precisely.

The scope is deliberately diagnostic and aims to make assumptions comparable across analysts, firms, and regulatory environments. The paper contributes:

1. An industry-agnostic taxonomy separating AI Integrators (IDL 0–3) from AI Providers (AI Wrapper and AI Native). IDL 0 and IDL 1 are valuation-neutral under this framework; IDL 2 and IDL 3 generate distinct option structures.

2. A milestone-gated real-options overlay where each milestone along the integration path gates a sequential investment decision with depth-specific and category-specific probabilities and payoffs, producing the additive structure $V_0 = V_{\mathrm{DCF}} + V_{\mathrm{AI}}$.

3. A $V_s$ decomposition splitting each milestone’s opaque value into five auditable components: baseline value, operating uplift, continuation optionality, regulatory cost, and execution cost.

4. A per-option probability architecture where each option carries an independently estimated $p_s$ derived from AHP-based intersubjective scoring (the Success Readiness Index), calibrated against depth-specific base rates and adjustable individually in scenario analysis. The SRI protocol accommodates mixed rater panels of internal and external experts; rater pool composition is an explicit, documented design choice governed by two credentialing dimensions (domain knowledge and conflict of interest), making rater governance itself auditable and contestable.

5. A scenario simulation layer producing structured valuation bands where each bound traces to identifiable option-level assumptions, enabling the valuator to locate where sensitivity concentrates.

6. Testable predictions on when markets react to AI milestones, how analyst dispersion changes with integration depth, and which milestone types move value most.

The framework applies at any lifecycle stage, from pre-investment planning to M&A due diligence on firms with fully completed AI integration, where option value has migrated into the DCF baseline.

Research

The valuation problem differs structurally depending on whether a firm uses AI to transform an existing business or sells AI as a product. A bank deploying machine-learning models for credit scoring faces integration risk, organisational adaptation cost, and regulatory compliance gates; a firm that develops and licenses the underlying model faces technology risk, platform economics, and competitive moat dynamics.

A single firm may contain elements of both, but the valuation logic for each component differs. Collapsing both into one classification obscures the mechanisms that drive option value. 4 introduces a taxonomy that separates these categories and defines the sub-classifications used throughout this paper.

AI Integrators are classified by Integration Depth Level (IDL 0–3). IDL 0 denotes no AI involvement and serves as a boundary condition. IDL 1 (Baseline) captures firms using off-the-shelf AI tools such as general-purpose chatbots or embedded copilot features in standard software; this creates no switching costs, no proprietary capability, and no measurable cash-flow restructuring. Because neither IDL 0 nor IDL 1 generates a decision node with identifiable conditional value, the option overlay is zero by construction and the framework’s analytical contribution begins at IDL 2. IDL 2 (Partial) marks the threshold where AI shapes specific business processes, producing measurable efficiency gains and the first real options. IDL 3 (Full) describes firms where AI is embedded across core operations, trained on proprietary data, and generates strategic feedback loops. At this depth, continuation-option value dominates operating gains, and the valuation problem shifts from measuring efficiency improvements to pricing future strategic optionality. The classification is cross-sectional; firms at any IDL level may remain there indefinitely.

AI Providers are subdivided into AI Wrapper and AI Native. A wrapper builds on third-party models, adding domain-specific interfaces or vertical integration without developing the underlying technology. A native develops proprietary models or infrastructure as its core product. The risk profiles diverge: wrappers face upstream dependency and compete on application-layer differentiation; natives face deep technology risk but can build platform-level moats. Progression from wrapper to native is rare but real, and the distinction matters for valuation. A permanent wrapper carries a narrower option set. A wrapper with a credible plan to develop proprietary models carries substantially more option value, because the transition unlocks moats, pricing power, and reduced dependency on upstream providers.

The unit of analysis for the valuation model is the milestone state: an observable, dateable point in the AI integration path at which the firm faces a decision node (continue, expand, pivot, or abandon). Each milestone state $s$ carries a conditional value $V_s$ and an independently estimated probability of success $p_s$; the model prices the firm’s AI component as a portfolio of such milestone-gated options. 1 summarises the valuation implications by integration depth.

The framework proposed here builds on DCF and makes the assumptions an analyst already embeds in it explicit and decomposable. In standard practice, the AI-related growth premium is a single opaque parameter. The option overlay replaces that parameter with a structured set of milestone-level inputs: which milestones are priced, at what probability, with what conditional value, and under which IDL classification. A sufficiently rich multi-scenario DCF with probability-weighted stage gates approximates this structure; at that point, the analyst is performing real-options analysis with DCF mechanics. The contribution is a transparent architecture that makes the assumptions behind the number auditable and comparable across analysts.

Illustrative application

The framework generates testable predictions: (a) milestone type and IDL depth predict the magnitude of market reactions to AI integration announcements, (b) higher-IDL milestones produce larger abnormal returns, (c) later-stage milestones within a given IDL level produce diminishing incremental effects as option value is progressively priced in, and (d) valuation dispersion increases monotonically with integration depth.

The framework is illustrated through a single evidence-anchored case reconstruction, parameterized from a real firm. The goal is to show that the model produces coherent, decomposable valuation bands under transparent, citable inputs, not to claim causal validation against market prices. All inputs are stated explicitly so readers can verify the arithmetic and substitute their own assumptions.

Case Firm E12 is a Swiss AI-native SaaS startup building an automated energy portfolio optimisation platform.¹ Its core product, E12-eAI, trades simultaneously across Day-Ahead, Intraday, Futures, power purchase agreements, flexibility markets, and CO₂ certificates, replacing the manual spreadsheets and fragmented legacy tools that most energy portfolio operators still rely on. Revenues combine platform licences, performance fees tied to realised customer margin uplift, data products, and integration services. Under the framework taxonomy, Case Firm E12 is an AI Provider (Native): E12-eAI is the product sold to customers, not a tool deployed internally; the energy portfolio operators who adopt the platform are the AI Integrators.

At the valuation date (Q2 2026), the AI core is at TRL 2–3: the optimisation algorithm has been validated on real market data, but the agent-orchestration layer is still in pre-development. The firm is targeting a CHF 15 million Seed round at a pre-money valuation of CHF 100 million. Four milestone gates over 24 months structure both the technical roadmap and investor decision rights (5): Gate 1 (month 3) confirms system stability for customer-facing pilots; Gate 2 (month 6) is the financing trigger, releasing the remaining CHF 13.5 million tranche only if two pilots have converted to paid agreements; Gate 3 (month 12) validates product-market fit through first recurring revenues (CHF 400k ARR); Gate 4 (month 24) establishes a reproducible go-to-market model with a proprietary TRL 7 model in production and 10+ active customers. Gate failure closes capital access and terminates all downstream options.

Each gate maps directly onto the option structure in Eq. 3: milestone state $s$, conditional value $V_s$, and sequential dependency such that failure at any gate terminates downstream options. The $V_s$ decomposition (Eq. 4) applies at each gate. In 6, $w_s = p_s V_s / (1+r_s)^{T_s}$ denotes the present-value-weighted contribution of each option to $V_{\mathrm{AI}}$, and $|\delta w_s|$ captures the absolute change in $w_s$ from a uniform $\pm 0.10$ shift in $p_s$, holding other milestones at base. $V_s$ combines all five components of Eq. 4; $\Delta V_{\mathrm{o},s}$ and $V_{\mathrm{c},s}$ are shown separately as the dominant drivers. All values are in CHF millions.

Table 6Milestone-gated valuation reconstruction, Case Firm E12 (Q2 2026). Source: (Case Firm E12 2026).

Milestone	$T_s$	$p_s$	$\Delta V_{\mathrm{o},s}$	$V_{\mathrm{c},s}$	$V_s$	$r_s$	$w_s$	$|\delta w_s|$
	(yr)		(CHF m)	(CHF m)	(CHF m)		(CHF m)	(CHF m)
Table  – continued
Milestone	$T_s$	$p_s$	$\Delta V_{\mathrm{o},s}$	$V_{\mathrm{c},s}$	$V_s$	$r_s$	$w_s$	$|\delta w_s|$
	(yr)		(CHF m)	(CHF m)	(CHF m)		(CHF m)	(CHF m)
M1: Shadow mode stable; 2 API pilots in onboarding (TRL 4)	0.25	0.82	2.0	5.0	13.9	15%	11.0	1.3
M2: 2 active paid pilots; CHF 13.5m tranche unlocked (TRL 5)	0.50	0.67	6.0	15.0	34.9	20%	21.4	3.2
M3: First recurring revenues (CHF 400k ARR); TRL 6+	1.00	0.52	18.0	35.0	79.3	25%	33.0	6.3
M4: Reproducible GTM; 10$+$ customers; TRL 7 model	2.00	0.38	48.0	95.0	198.5	30%	44.6	11.7
$\sum w_s$							110.0
$-I_0$ (seed capital committed)							$-$15.0
$V_{\mathrm{AI}}$							95.0
$V_{\mathrm{DCF}}$ (IP/team baseline)							5.0
$V_0 = V_{\mathrm{DCF}} + V_{\mathrm{AI}}$							100.0

Case Firm E12 has no revenue history at the valuation date. That means $V_{\mathrm{b},s}$, the baseline value given sunk costs, is carried by team credentials and IP rather than by cashflow extrapolation, which is why the $\Delta V_{\mathrm{o},s}$ values at M1 and M2 are small: no sustained customer margin uplift has been demonstrated yet. Value is almost entirely forward-looking, concentrated in $V_{\mathrm{c},s}$, the competitive-moat option that unlocks once proprietary TRL 7 models are in production and deliver continuously improving performance advantages over competitors still dependent on external-model wrappers. The option-killing gate structure has real teeth: the CHF 13.5 million tranche is contractually tied to Gate 2 pilot conversion, so two missed pilots terminate most available capital and effectively kill M3 and M4. The AHP–SRI probability priors reflect this structure directly: $p_{\mathrm{M1}} = 0.82$ captures near-complete technical control over Gate 1 conditions; $p_{\mathrm{M4}} = 0.38$ captures the compounded uncertainty of product-market fit, competitive dynamics, and 24 months of model maturation.

The base case yields CHF 100 million, matching the disclosed pre-money exactly. The inputs were derived independently from the masterplan’s financial model and gate definitions, so the convergence reflects genuine internal consistency. For context, the masterplan itself arrives at the same figure through a distinct methodology: a milestone-scoring overlay across 13 strategic components (readiness score 36.9% in Q2 2026; strategic value CHF 102.1m) is added to a negative DCF contribution (WACC 30%, pre-revenue; $-$CHF 1.3m), yielding CHF 100.8m, rounded to CHF 100m as the pre-money anchor ((Case Firm E12 2026), Ch. 8). Two independent methods converging on the same value provides stronger evidence of internal consistency than either alone. The full valuation band runs from CHF 77 million (downside: all $p_s - 0.10$) to CHF 123 million (upside: all $p_s + 0.10$), a range of CHF 46 million. As 6 shows, M4 drives more than 60% of that range ($|\delta w_s| = 11.7$): the proprietary model maturation and reproducible GTM option concentrate risk in the later stage, exactly as the structural prediction for AI Providers anticipates. Removing M4 entirely collapses $V_{\mathrm{AI}}$ from CHF 95 million to approximately CHF 50 million, meaning that roughly half of firm value rests on 24-month execution. Two analysts who disagree on the CHF 100 million pre-money can now identify whether the disagreement sits in $p_{\mathrm{M4}}$, in $V_{\mathrm{c},\mathrm{M4}}$, or in the 30% discount rate, each of which is individually challengeable against observable proxies.

Discussion

Standard valuation approaches do not ignore AI integration; they absorb it into an undifferentiated growth-rate assumption. The five-component $V_s$ decomposition (Eq. 4) separates option value at each integration stage. $\Delta V_{\mathrm{o},s}$ captures the structural efficiency gain realised at the IDL 2 transition, $V_{\mathrm{c},s}$ captures the competitive-moat value that becomes accessible at IDL 3, and $C_{\mathrm{r},s}$ and $C_{\mathrm{e},s}$ record compliance and execution costs instead of netting them silently against projected revenues. The contribution does not lie in a different total valuation, but in the decomposition itself. Disagreement over $V_{\mathrm{AI}}$ can then be located in a specific component, probability estimate, input value, or discount rate.

This decomposition points to two systematic mispricing channels under conventional approaches. IDL 1–2 firms are likely to be undervalued when option value is ignored: the operational efficiency gains captured by $\Delta V_{\mathrm{o},s}$ are real and staged, but a single-stage DCF with a generic AI premium does not separate them from baseline revenues. IDL 3 firms face the opposite risk: markets may price network effects and platform control, represented here by $V_{\mathrm{c},s}$, as already realised although they remain contingent on milestone completion. The framework does not remove that contingency, but it does isolate the probability assumptions ($p_s$) that drive the valuation instead of embedding them in a single discount rate.

These mispricing channels are compounded by financial reporting standards. Under HGB (§248 II), development costs carry an optional capitalisation right (Aktivierungswahlrecht) that conservative preparers typically do not exercise; AI development expenditure hits the income statement in full and the balance sheet carries nothing. Under IFRS (IAS 38.57), development costs must be capitalised once six criteria are met: (i) technical feasibility of completing the asset, (ii) intention to complete and use or sell it, (iii) ability to use or sell it, (iv) demonstration of probable future economic benefits, (v) availability of adequate technical and financial resources, and (vi) ability to measure reliably the attributable expenditure; research costs are always expensed. The capitalisation boundary is judgement-intensive and varies across firms and auditors. Under US GAAP (ASC 350-40 for internal-use software), capitalisation begins only at the application development stage and stops at post-implementation, with preliminary project work always expensed. The result: two firms committing CHF 10 million to economically identical AI development programmes may report dramatically different earnings, asset values, and return metrics depending solely on their reporting regime (7). A value driver that generates CHF 95 million in real-option value may simultaneously produce a multi-year operating loss under HGB and a partially capitalised intangible asset under IFRS. The economic reality is identical; the reported financial statements are not.

The framework addresses this directly. The cost term $C_{\mathrm{e},s}$ in the $V_s$ decomposition (Eq. 4) records AI development expenditure as an economic cash outflow, regardless of whether the applicable standard classifies it as an asset or an expense. Two analysts applying the framework to the same firm under different reporting regimes will produce the same option-level inputs, because the decomposition operates on the economic layer beneath the reporting layer. This is precisely why it produces comparable valuations where unadjusted multiple-based methods produce incomparable ones.

The structural shift from $\Delta V_{\mathrm{o},s}$-dominance at IDL 2 to $V_{\mathrm{c},s}$-dominance at IDL 3 generates a directional empirical prediction: analyst forecast dispersion should increase monotonically with IDL depth. Competitive-moat value contingent on full integration success, represented here by $V_{\mathrm{c},s}$, is inherently harder to estimate than $\Delta V_{\mathrm{o},s}$, which has observable operational proxies. This prediction is testable from analyst forecast data without access to the internal valuations that produced it. The exploratory AHP hypothesis runs in parallel: if pairwise-comparison consistency ratios increase with IDL depth, that constitutes independent evidence that the estimation problem is structurally harder at higher integration levels, beyond what individual rater heterogeneity alone would generate.

Single-rater probability estimates are vulnerable to anchoring and overconfidence, both well-documented patterns in expert elicitation. The AHP-based SRI protocol constrains expert judgement rather than replacing it. Pairwise comparisons expose internal inconsistency before aggregation, and the credentialed weighting scheme (Eq. 5) discounts estimates from raters with domain gaps or conflict-of-interest exposure. The resulting gain over current practice is narrow and concrete: disagreement can be assigned to a specific $p_s$ estimate for a specific option at a specific IDL stage, where a negotiated discount rate would otherwise absorb it.

$C_{\mathrm{r},s}$ enters the $V_s$ decomposition (Eq. 4) as a cost term, but regulatory compliance may also carry option value that the present framework does not model fully. Firms that complete EU AI Act compliance milestones early may acquire a market-access advantage that late compliers cannot replicate quickly; in regulated industries that advantage may persist. In the Swiss context, FINMA’s 2024 guidance on AI-driven financial services (Swiss Financial Market Supervisory Authority (FINMA) 2024) and the Federal Council’s sector-specific approach (Swiss Federal Council 2025) create distinct regulatory milestone paths whose option structure differs from the EU model. A fuller treatment of regulatory compliance as a staged option remains outside the present scope.

The lifecycle model (Eq. 6 and Table 5) and the synergy decomposition (Eq. 7) reframe M&A negotiation. Under current practice, acquirer and target disagree on a headline premium whose component drivers are opaque to both parties. The framework reduces this to a structured comparison of specific options, probabilities, and continuation values. An acquirer who assigns high probability to completing the IDL 3 transition can justify a higher premium transparently; a target that believes its $p_s$ estimates are conservative can argue that point at the option level rather than relitigating the discount rate. This does not eliminate negotiation difficulty; it replaces narrative disagreement with structured disagreement that due diligence can address.

Three additional finance-side gaps are worth making explicit before turning to the literature.

First, valuation multiples. EV/Revenue and P/E are the standard plausibility checks on any DCF, but they presuppose a homogeneous peer group. An IDL 3 firm with a platform-level data moat trades at fundamentally different multiples than an IDL 1 firm using off-the-shelf AI tools, even within the same sector and the same GICS sub-industry classification. Assigning a technology-sector median multiple to both produces a spurious comparison. The IDL taxonomy makes peer selection explicit and IDL-conditioned: the right comparable set is determined by integration depth, not sector label alone.

Second, terminal value. In a standard DCF, 60–80% of total present value typically concentrates in the terminal-value term, a single growth rate applied to a normalised cash flow in perpetuity. For a firm at an IDL transition boundary, this growth rate implicitly bets on whether IDL 3 continuation value materialises. The lifecycle model (Eq. 6) addresses this structurally: as milestones clear, option value migrates into the DCF baseline and the terminal-value calculation is applied to an already-adjusted cash flow rather than one that embeds unresolved option payoffs. The framework does not eliminate terminal-value sensitivity, but it confines it to the post-exercise DCF rather than allowing it to absorb milestone uncertainty silently.

Third, preferred equity and dilution. The pre-money headline valuation of an AI-native firm diverges from its common-equity value. Liquidation preferences, ESOP pools, and anti-dilution provisions systematically decouple the two: Gornall and Strebulaev find that US unicorn post-money valuations exceed fair common-equity value by 48% on average due to the economic rights embedded in preferred shares (Gornall and Strebulaev 2020). The Case Firm E12 illustrates this directly: the CHF 100 million pre-money is a Series seed headline. The $V_s$ decomposition operates on economic cash flows attributed to each option, which means dilution and liquidation-preference effects must be modelled separately in the capital-structure layer rather than embedded in either $p_s$ or $r_s$. Conflating them produces a valuation that is internally consistent at the option level but misleading at the equity level.

7 maps each gap to its framework response; the common pattern is that the decomposition moves the problem to a layer where economic cash flows can be compared directly, rather than leaving it embedded in a multiple, a growth rate, or a headline number.

The Case Firm E12 functions as a demonstration, not as an empirical validation. The case was chosen for disclosure quality and for a clean AI-native structure, conditions that favour the framework’s applicability and that bias case selection toward firms where the decomposition fits. One case establishes that the protocol can be applied coherently and yields a traceable band; it does not establish that the framework outperforms a conventional DCF on any observable metric, nor that the decomposition generalises to hybrid integrators, low-disclosure firms, or integration failures. A proper test requires a multi-case portfolio covering IDL levels, integration successes and failures, scored against realised post-exit valuations. That test is left to future research.

The framework connects three bodies of work. It introduces option structure into intangible-asset valuation (Lev 2001), where the disclosure problem is recognised but no mechanism is provided for staging value realisation. It introduces integration-depth architecture into real-options applications in technology valuation (Schwartz and Moon 2000), where uncertainty is modelled at the firm level without conditioning option structure on the depth of technology embedding. It also links AI adoption research (Hansen et al. 2024; Brynjolfsson et al. 2018) to valuation by connecting milestone achievement to firm value. The AHP component draws on established multi-criteria methodology (Saaty 1980) and uses it for IDL-conditioned probability estimation in staged AI integration options.

Conclusion

The central claim of this paper is modest in scope but consequential in implication: standard valuation methods are not wrong about AI-integrated firms, but they are structurally incomplete. DCF, IDW S 1’s income approach, and market multiples all remain valid instruments for established cash-flow businesses. When applied to firms whose value is dominated by staged, contingent, and depth-dependent AI integration decisions, these methods compress option value, milestone risk, and continuation payoffs into parameters that obscure the very assumptions an analyst needs to challenge. The framework proposed here does not discard that machinery. It adds a decomposition layer, the IDL 0–3 taxonomy, the five-component $V_s$ structure, and the AHP-based SRI protocol, that makes the embedded assumptions explicit and individually auditable.

The framework’s contribution lies in the informativeness of the disagreement it surfaces, beyond any improvement in point estimates relative to DCF or multiples. A conventional DCF applied to Case Firm E12 would yield a single present value whose sensitivity concentrates in the terminal growth rate and the discount rate, two parameters that absorb everything the analyst believes about AI integration without revealing what those beliefs are. The framework decomposes that belief into four milestone-level probabilities, five value components per milestone, and a gate structure whose failure conditions are contractually defined. Two analysts who disagree on the CHF 100 million pre-money can now locate their disagreement in $p_{\mathrm{M4}}$, in $V_{\mathrm{c},\mathrm{M4}}$, or in the 30% discount rate. That precision is the contribution: a valuation whose internal structure can be interrogated.

Real-options theory has made similar promises before, and the track record warrants scepticism. Schwartz and Moon’s option models for internet firms in 2000 were formally elegant but empirically uncalibrated; the biotech pipeline analogy that underpins much of staged-option pricing works well in pharmaceuticals, where phase-gate probabilities have decades of actuarial data, but lacks comparable base rates for AI integration milestones. The AHP-based probability architecture proposed here addresses part of this calibration gap. It replaces unstructured expert guesses with pairwise-comparison discipline, consistency checks, and credentialed aggregation; the resulting $p_s$ estimates still sit in the domain of expert judgement, since observed frequencies for AI integration milestones do not yet exist at scale. The framework’s honesty about this point is deliberate: it surfaces where judgement enters the valuation, exposing what a discount rate would otherwise absorb. Whether that exposure leads to better decisions is an empirical question the framework itself cannot answer.

The Case Firm E12 illustrates both the promise and the boundary. On one hand, the framework produces a coherent valuation band (CHF 77m to CHF 123m) from transparent, citable inputs, and the concentration of risk in M4 ($|\delta w_s| = 11.7$, more than 60% of total sensitivity) confirms the structural prediction that AI Provider valuations are dominated by late-stage continuation options. On the other hand, roughly half the firm’s value rests on a single 24-month execution bet, a finding that no amount of decomposition makes less uncertain. The framework clarifies what the bet is and where it sits; it does not make the bet safer.

For European valuation practice, the implications extend beyond methodology. The measurement mismatch between how AI-integrated firms create value and how statutory standards capture it is not a temporary gap that market maturation will close. IDW S 1’s income approach, built for firms with demonstrable earnings histories, systematically underweights optionality; HGB’s treatment of development costs ensures that balance sheets remain silent about the very investments that drive AI value. The framework proposed here operates on the economic layer beneath these reporting conventions: $C_{\mathrm{e},s}$ records AI development expenditure as economic cost regardless of its accounting classification, and the $V_s$ decomposition yields identical inputs under HGB, IFRS, and US GAAP. If European venture capital is to close even part of the sevenfold valuation gap with U.S. peers at comparable development stages, the valuation infrastructure will need to accommodate option-loaded, milestone-gated firm profiles. This paper proposes one architecture for doing so.

Limitations and Future Research

Milestone identification depends on disclosure quality, and firms may reveal neither their integration depth nor the timing of intermediate gates with sufficient precision for external reconstruction. Parameter estimation for $p_s$, $V_s$, and $r_s$ remains sensitive to assumptions, especially at IDL 2–3 where comparables are scarce; the framework exposes that sensitivity rather than resolving it. IDL classification also requires judgement for hybrid firms operating at multiple levels across functions. The AHP component introduces further constraints: pairwise-comparison burden grows with criterion count, the IDL-specific calibration anchors remain illustrative priors rather than empirically derived base rates, and geometric-mean aggregation assumes a level of rater independence that may weaken when internal and external experts share the same organisational context. The credentialed weighting scheme partially addresses rater-quality and independence concerns, but assigning $w_i^{\mathrm{comp}}$ and $b_i$ is itself a judgement that must be documented and remains open to challenge. Option interdependency handling through attribution by assignment likewise depends on human judgement, because the interaction matrix is qualitative and does not replace formal coalition analysis. The case design is illustrative rather than inferential: a single firm with detailed milestone disclosure can test framework coherence, but not predictive accuracy. Case selection was also purposive: the Case Firm E12 offered rich investor disclosure, an AI-native business model, and a clean milestone architecture, conditions under which the framework is expected to fit well. Cases with weaker disclosure, hybrid integration across IDL levels, or post-hoc failed transitions remain outside the present evidence base, and generalisation requires a larger and more varied sample. Regulatory conditions are also moving targets, because the EU AI Act is not yet fully in force and the Swiss sector-specific approach is still evolving.

Future research should move in five directions.

The first direction is empirical validation. Applying the framework to a multi-case portfolio across IDL levels and integration outcomes, with option-level estimates scored against realised post-exit valuations, would test whether the decomposition improves on conventional alternatives rather than merely describing them.

The second direction is longitudinal. Tracking dynamic IDL transitions would show how value migrates from the option overlay into the DCF baseline as milestones clear over time.

The third direction is empirical calibration. Panel data on AI project outcomes across industries could anchor the SRI-to-$p_s$ mapping and refine milestone payoffs by sector.

The fourth direction is methodological. Bayesian updating for $p_{s+1 \mid s\;\text{cleared}}$, Shapley-value decomposition for option interdependencies, and robustness checks with alternative MCDM methods such as fuzzy AHP, ANP, or BWM would test how sensitive the framework is to its current simplifying choices.

The fifth direction is regulatory. As EU AI Act provisions take effect and Swiss guidance evolves, cross-jurisdiction evidence could show how compliance regimes alter milestone feasibility, risk, and continuation-option value.