Why the forward deployed engineer lowers the cost of integration and can raise the cost of imputation
Third piece in the Frontier Company series, this article picks deployment up where the two earlier ones left it and asks the question Microsoft, Amazon and OpenAI leave unspoken: once integration is solved, what is the binding factor, and can the instrument chosen to solve it, the forward deployed engineer, reach that far? The answer is compact enough to serve as the thesis. Under ex ante allocation of responsibility, the binding factor stops being technical and becomes institutional. The family of FDE strategies, superb at building, is neither the authority that institutes nor the party that answers. The European regime (Regulation EU 2024/1689) is not the only place the mechanism operates, it is the place it is written down. Validity domain: three conditions must hold together, responsibility allocated before deployment, juridical roles named in advance, a system whose decisions carry high stakes. Remove any one and the forward deployed engineer may well be enough.
The deployment turn rests on a correct observation and a hasty inference: the model is no longer the problem, therefore deployment is integration. It is not. In a regulated regime, to deploy is three distinct jobs. To build is to produce the working system: integrations, harness, evaluations, observability. To institute is to establish the rule system under which it may operate: the conformity file, the risk management system, human oversight, the allocation of who is accountable for what. To answer is to carry the responsibility for the decisions the system makes. The market prices only the first and treats the other two as a downstream chore.
The third job splits, and this split carries the whole argument. The line that matters is not moral but economic: transferability. Attesting produces a transferable object, the declaration, the marking, the technical file, the formal risk acceptance. It is a deliverable, delegable, insurable, partly codable as policy as code and evidence as code. Because it makes a market, it carries rent. Answering is the other half, to be read in its technical sense and not as a moral flourish: remaining the duty holder exposed, in civil and criminal law, once attestation, insurance and documentation have done all they can. It is not insured, not delegated, not coded. It makes no market, so it carries no rent, only exposure.
It would be weak to claim the forward deployed engineer does not institute: he can document an institution, accelerate its artefacts, reveal its gaps. What he cannot do is constitute the authority from which the institution draws its force. Worse, the depth that makes his value is also his trigger. By tuning the model’s local weights to the customer’s specifics, the embedded team does exactly what it was paid to do, and in doing so moves the customer toward the substantial modification that Article 25 treats as turning a deployer into a provider. The scene is unremarkable: the connector ships to production, the load tests pass, then the chief compliance officer refuses to sign the fundamental rights impact assessment (Article 27). Integration is complete, deployment is paralysed. Requalification is not mechanical, it is a matter of law and fact; which is precisely why no supplier clause waives it. An unpriced risk manufactured by the go-to-market motion itself.
The predictable objection: conformity will be automated. In part, yes. Continuous compliance, policy as code, automated conformity assessment: the market is already industrialising the evidentiary, documentary and procedural layers. What automation reaches is attesting, the transferable instrument. What it does not reach is answering. Bainbridge’s ironies of automation, written for control rooms in 1983, transpose exactly, provided one aims at the right residue: the more of building, integrating and even attesting is automated, the rarer, the more critical and the more concentrated the remaining act becomes. And that act is not a signature on a form, which delegates and insures. One codes the proof, one transfers the attestation, one does not automate the act of answering.
The regulatory calendar moved while the series was being written, and the move strengthens the thesis rather than contradicting it. Under the Digital Omnibus on AI (Council-Parliament political agreement of 7 May 2026, awaiting publication in the Official Journal at the time of writing), the high-risk obligations for standalone Annex III systems (recruitment, credit scoring, education, border control) are deferred from 2 August 2026 to 2 December 2027, and those for Annex I systems embedded in regulated products to 2 August 2028. The transparency obligations of Article 50 hold at 2 August 2026. The deferral was granted because the machinery that makes the obligations operable (harmonised CEN-CENELEC standards, notified bodies, competent authorities) was not ready. The binding factor for deployment at scale is therefore neither the model, which shipped, nor the integration, which the engineers ship: it is the infrastructure of institution, and the regulator concedes its absence by conceding the calendar. The AI Act does not create the mechanism, it reveals it; it is a microscope, not the phenomenon. Guardrail: a deferred calendar is not a soft constraint. The Omnibus moves the date of attesting, it does not move answering. Sectoral law bites in 2026, and personal exposure does not defer with the annex.
The consequence needs one layer more than the earlier notes gave it. Rent does not migrate toward attestation, it migrates toward power, and the relevant power is the authority to authorise operation. That authority is the scarce asset; attestation is only its marketable face, the instrument through which the authority is exercised and evidenced. This is old ground in the economics of organisation: institutions, in North’s sense, are the rules of the game, and the rules here confer a right to authorise; value accrues to the holder of the conferred right, not to the technician who prepares the file. The boundary the vendor cannot cross is Williamson’s: attestation is contractible, therefore a market; answering is the residual no contract completes, therefore it stays inside the party who holds it. Vendors will reach for the transferable part through separate assurance vehicles that monetise attestation without costing them their neutrality. What no vehicle captures is answering. The CTO frames the integration, the CFO budgets the operation, a third party is paid to attest, but the Decision Owner alone answers, and the vendor sent no one to hold that pen.
The transatlantic asymmetry is the engine that explains the thesis. The forward deployed model was born in the United States, where responsibility is produced after the fact and from many sources (jurisprudence, class actions, insurers, contracts), arbitrated at the level of use rather than of ex ante attestation. There the embedded engineer is closer to a complete solution, because the institution forms later, around the use. But the asymmetry is one of mechanism and degree, not of presence: the fear of a class action already stops many an American committee. Europe is not the only place the mechanism operates, it is the place it is legible. The thesis is falsifiable twice: in principle, the day answering becomes transferable (personal immunity in exchange for a licensed third party’s attestation); in fact, if firms with embedded teams deploy regulated, high-impact AI systematically faster, with no increase in institutional cost and no reallocation of responsibility. The second test is the one to watch. What holds for a credit model holds for a medical device, a grid controller, an aircraft system: the technical layer industrialises, fast and with great capital, and then the work stops where authority begins. Software builds. Rules institute. A person answers.
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.