A structural asymmetry of FDE models: why supplier and client never observe their return in the same window
In the summer of 2026, infrastructure and model providers reinforced arrangements of engineers deployed at the client’s point of contact, with commitments running from hundreds of millions to a few billion dollars and headcounts in the thousands. The forward deployed engineer, which a data-analytics vendor turned twenty years ago into a distinctive model for public agencies, has become one of the most visible industrial forms of enterprise AI deployment. The ambient discourse treats it as a delivery model, a more intimate way of putting a system into production. That is a surface reading. This article proposes a different one, and names the framework it rests on: the Dual Value Horizon Model (DVHM).
The recovery mechanism sets the window within which the supplier can recognise its return. The client’s net return belongs to another window, independent and most often later, because it can be established only after observing the cost of permanence and after causal attribution of the benefits to the deployment. Supplier and client do not evaluate the same economic object, and they do not evaluate it over the same horizon. The validity domain is explicit: return here designates the net financial return after full cost, and the scope is the platform FDE, a technical and product team operating for a supplier whose return depends on adoption of its product or ecosystem. Independent consulting with no product interest is out of scope.
A point of vocabulary decides the rest. A supplier first recognises revenue, a margin, a learning or a strategic option. ROI appears only once that value is set against the full cost of the intervention. The confusion between return and ROI is not neutral: it lets what is still only a receipt be presented as a yield. The same discipline governs the distinction between the cost of the token and the price of the decision, and it is the reason adoption at go-live is not deployability.
Four mechanisms coexist and a single mission may combine them. The FDE financed as a product cost, close to quasi-R&D, where the learnings flow back into the platform. The FDE financed as an accelerator of consumption or licensing, where a contracted commitment decrements on usage. The FDE financed through a dedicated investment vehicle, the most financialised case, which forces one not to confuse the investor’s yield, the vehicle’s return, the supplier’s benefit and the value captured by the client. The FDE financed as a billed services activity, which is not consulting rebranded as long as the supplier’s return remains attached to product adoption. These families do not classify suppliers; they decompose mechanisms. Read through six variables (immediate payer, economic bearer, recovery mechanism, captured asset, return horizon, exposure to failure), they read as a single grid. A fifth actor, the integrator, traverses them all and displaces the bearer of the cost of permanence, either carrying the capacity durably or withdrawing it with its mandate.
The supplier’s captured value is written as a function that aggregates the forms of value it recognises over time: recognised revenue, the economic value of reusable learnings, and the expected value of induced future usage. This function grows sharply around the deployment, then its slope collapses. The reason is economic, not mathematical: once the system is in production, the marginal contribution attributable to the deployed engineer’s presence declines, even as ecosystem value keeps running. There is therefore a horizon, generally close to production go-live, beyond which the value directly imputable to the FDE becomes small. It is not the value that ceases; it is the share of value attributable to the FDE that saturates.
The client’s trajectory is different. Net value begins negative, since the client immediately bears the project cost and the cost of constituting capacity, then evolves under the effect of genuinely attributable benefits minus the cost of permanence: recurring operation, governance and supervision, adaptation and reversibility. Early on, attributable benefit is small and the cost of permanence is high, so cumulative net value stays below zero. Later, if benefits materialise and recurring costs stabilise, the integrand turns positive and the trajectory crosses the axis. That crossing is the client break-even, and it is late by construction.
The central property of the model is not an ROI. It is an inequality between two characteristic times: the client break-even is much greater than the supplier’s saturation horizon. The supplier reaches its recognition horizon while the client has not yet observed the full cost of permanence. High adoption at production go-live principally validates the supplier’s capture function; it does not yet demonstrate a durable net return for the client. This is the economic core of the thesis, and it no longer depends on the presence of deployed engineers.
Between the two horizons sits a mediating variable: the capacity effectively transferred and institutionalised at the client. Its effect is twofold and of opposite sign. The more capacity is transferred, the more attributable benefits rise, because the client knows how to operate, correct and extend the system, and the more governance costs fall, because supervision ceases to depend on an external operator. Transferred capacity is therefore not a variable of comfort: it is the channel through which a deployment with a short supplier horizon can, or cannot, produce a client return within a finite horizon. This is where the DVHM meets the attribution of performance across a socio-technical pipeline: without a counterfactual there is no attribution, and a before-and-after business case is a narrative, not a measurement.
The cost of permanence is not only technical. After the team withdraws, four functions must remain assigned: to decide, to operate, to finance, to answer. The theory of incomplete contracts, from Grossman and Hart to Hart and Moore, supplies the mechanism: when future states cannot all be specified, residual control rights determine who decides in unforeseen situations, and it is because those rights stay with the client that its return reveals itself only in use. The stake is not to designate a single owner, which would make one actor bear a risk without sufficient control, but to keep the allocation of responsibilities reconstructible, in the same sense that an agent that decides must remain traceable to a signatory, and that the deployed asset is no longer the model but the operating capability. Competence is not capacity, capacity is not institutionalised capacity, and a documented competence is not yet an institutional one.
The model is bilateral, and that is also its blind spot: it sees neither the regulatory, ecological or systemic value captured by no actor, nor the value of permanence symmetric to its cost. The mediating variable remains underspecified, posited through its two partial derivatives rather than a measurement protocol. And a limit of demonstration must be named without evasion: the text establishes that the recognition and demonstration windows are distinct and governed by different mechanisms, not yet that this difference causes the observed gap in return. A recognised link is not an attributed link, which is precisely the distinction the model imposes on its own objects. The testable hypothesis follows: on a population of some fifty missions, high-capacity-transfer deployments should reach break-even significantly earlier at comparable recovery mechanism. The real counter-models, a managed operator paid on persistent performance or a genuinely institutionalised transfer, do not refute the thesis; they circumscribe its validity domain. A correctly analysed deployment can show a positive client return after full cost.
The gesture is the same one at work elsewhere in this corpus: name the distinction that separates two regimes of operation, one viable under a regulated frame, the other not. Here the distinction is temporal. The FDE does not only modify the cost of deployment; it shifts when each party can claim to have created value, in the same movement that turns the death of the platform ROI into a question of who bears the cost of permanence. The business model organises the window of the supplier’s return; the cost of permanence reveals the client’s. As long as these two windows remain conflated, adoption dashboards will keep taking early revenue for demonstrated ROI. FDE models do not merely shift the distribution of value. They shift its temporal distribution.
Full text freely available in the PDF below (10 pages, with the complete formalisation and references).
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.