> ## Content Index > Fetch the complete content index at: https://www.edgewisely.com/llms.txt > Use this file to discover other available public pages before exploring further. # The Model Is Free. The Installation Isn't. - URL: https://www.edgewisely.com/forward-deployed-engineer-enterprise-ai-services-margin/ - Published: 2026-09-11T05:18:10.000Z - Updated: 2026-09-11T05:19:19.000Z - Description: How the commoditisation of models and orchestration moved the only durable moat in enterprise AI into the least software-like place imaginable: people, embedded, on site. - Author: John Karpentar - Tags: Opinion, Enterprise **Our view: the forward-deployed engineer is not a support function. It is becoming the product, and the software industry has not priced that in.** Something quietly inverted in enterprise AI this year. The capability layer got cheap and abundant. The installation layer got expensive and scarce. And the companies that noticed first started hiring humans by the thousand. In July, Microsoft [committed $2.5 billion and 6,000 employees](https://www.cnbc.com/2026/07/02/microsoft-commits-2point5-billion-6000-employees-ai-implementation-unit.html?ref=edgewisely.com) to a new unit — Frontier Company — whose entire purpose is to embed engineers inside customers' operations. Judson Althoff, who runs Microsoft's commercial business, described the aim as co-designing and continuously improving AI systems against measurable business outcomes, with [Rodrigo Kede Lima leading the unit](https://the-decoder.com/microsoft-launches-2-5-billion-frontier-company-to-embed-6000-ai-engineers-inside-enterprise-clients/?ref=edgewisely.com) and an explicit ambition to go beyond the standard forward-deployed engineering model. Amazon, Anthropic and OpenAI have all stood up comparable deployment groups. Read that again as a software company would have read it in 2019: the world's largest software vendor is spending billions to put people in buildings. That is not a software gross margin. That is Accenture's business. Our argument is that this is not a temporary bridge until the models get good enough to install themselves. It is the durable shape of the market, and it means the most valuable position in enterprise AI is the one software investors have spent twenty years learning to avoid. ## The gap that created the job Start with the number everyone in enterprise AI has now internalised. MIT's *GenAI Divide* research, published in 2025, found that [roughly 95% of enterprise generative AI pilots produced no measurable P&L impact](https://www.forbes.com/sites/jaimecatmull/2025/08/22/mit-says-95-of-enterprise-ai-failsheres-what-the-5-are-doing-right/?ref=edgewisely.com), against tens of billions in spend. The study has been criticised for its sample size and methodology, and the critics have a point — 52 interviews and 153 surveys is thin, and "no measurable P&L impact" is a definitional choice that does a lot of work. But the finding survived scrutiny because it matched what everyone was seeing. And the reported causes are the interesting part: brittle workflows, no contextual learning, misalignment with how work actually gets done. Not model quality. Not accuracy benchmarks. Integration. That diagnosis has a specific implication most vendors resisted for two years. If pilots fail on workflow fit rather than capability, then no model improvement fixes them. A smarter model dropped into a badly-understood process produces a better-articulated failure. The missing input is knowledge of the customer's actual operations — which claims get escalated and why, which fields in the CRM are lies, which approval exists because of a 2017 audit finding. That knowledge is not in any training corpus. It is in people's heads, and extracting it requires someone sitting there. Hence the forward-deployed engineer: not a consultant who writes a recommendation, but an engineer who builds inside the customer's environment and stays. ## Why the capability layer stopped being the moat The second half of the argument is that the alternative moats are collapsing, and this week made that unusually visible. OpenAI [released the Agents API in public beta](https://www.edgewisely.com/openai-agents-api-codex-harness-public-beta/) with no platform fee — giving away the orchestration harness that every serious agent team had been building by hand. Orchestration was, for eighteen months, a genuine differentiator. Now it is a maintained service you get for free by buying tokens. That is the pattern, repeating. Model quality converged across the frontier labs to the point where most enterprise workloads cannot tell the tiers apart. Inference costs fell hard. [Open-weight models captured enormous token volume without capturing the revenue](https://www.edgewisely.com/open-weights-won-the-tokens-not-the-money/). Each layer that looked like a moat became a commodity within roughly a year of being identified as a moat. Deployment knowledge has not commoditised, and the reason is structural rather than temporary. It does not compress into a product because it is not general. What Microsoft's engineers learn inside one insurer's claims operation transfers only partially to the next insurer and barely at all to a logistics firm. It is expensive, non-scaling, human work — which is exactly why it resists being competed away. This is the uncomfortable part for anyone holding software multiples. The most defensible asset in enterprise AI has the economics of a services business. ## What this means if you are buying Three consequences follow, and we think all three are underweighted. **Judge vendors on deployment capacity, not benchmarks.** If integration is where pilots die, then the relevant question in a vendor evaluation is how many engineers they will put on your problem and for how long. Model performance is now close to table stakes across serious providers. Willingness to own your workflow is not. A vendor that leads with benchmark scores and offers a solutions architect for onboarding is selling you the 95% outcome. **Stop treating "build" and "buy" as the decision.** The [build-versus-buy framing has been eroding all year](https://www.edgewisely.com/why-companies-are-building-not-buying/), and the Agents API pushed it further by absorbing another layer of what "build" used to mean. The real choice is who holds the deployment knowledge: your platform team, a vendor's embedded engineers, or a systems integrator. That is a question about where institutional memory lives, and it has a much longer half-life than a procurement decision. **Budget for the installation, not the licence.** Enterprises consistently underfund the integration line and then report that AI did not work. If the empirical failure rate concentrates in workflow fit, then a programme that spends 80% on tooling and 20% on the people who embed it has its ratio inverted. The [productivity gains not reaching the P&L](https://www.edgewisely.com/ais-productivity-gains-arent-reaching-the-p-l/) are largely gains that were never installed. ## What this means if you are selling The harder implication is for vendors, and it is a strategy problem rather than an operational one. A software company that adds a large deployment organisation takes on lower gross margins, linear headcount scaling, utilisation management and a much harder hiring problem. Every instinct in SaaS says do not do this. Microsoft is doing it anyway, at $2.5 billion, because the alternative is watching customers fail to realise value and churn — and because [the gap between AI capex and AI revenue](https://www.edgewisely.com/ai-capex-revenue-gap-2026-hyperscalers/) closes through adoption or it does not close. The strategic bet underneath Frontier Company is that deployment scale is defensible in a way product features are not. Six thousand engineers with operational knowledge inside thousands of enterprises is an asset a competitor cannot replicate by shipping faster. It can only be replicated by hiring six thousand engineers and spending three years learning the same things. Where we think this ends up: the market bifurcates. A small number of players own the capability layer and give away increasing amounts of it to drive consumption. A larger set competes on deployment depth in specific industries, with services-like economics and unusually high retention. The middle — pure-play software with a thin services wrapper and a benchmark-led pitch — gets squeezed from both directions. That middle is where most AI application startups are currently standing. ## The case against us The strongest counterargument is that this is a transitional artifact, and it deserves a fair hearing. On that view, models capable enough to interrogate a customer's systems, infer the real workflow and configure themselves are a matter of a few years — and the moment they arrive, six thousand embedded engineers become a stranded cost rather than a moat. Agentic systems that can read a codebase, watch a process and propose an integration are already partway there. Microsoft would then have bought a large, low-margin organisation right before the thing it does gets automated. We think that underrates how much of the required knowledge is political rather than technical. The reason a process is broken is frequently that two executives disagree and the workflow encodes the truce. No model reads that off a database. Someone has to sit in the room. But we hold the view with appropriate humility, because "this human bottleneck is temporary" has been the correct call more often than not in this cycle. *The layer that gets commoditised is always the one everyone agrees is the moat. Watch where the humans go.* ## Frequently Asked Questions ### What is a forward-deployed engineer? A forward-deployed engineer is a technical staff member embedded inside a customer's organisation to build and operate software within that customer's actual environment, rather than shipping a product for the customer to implement. The model, popularised by Palantir, is now being adopted at scale by major AI vendors to close the gap between AI capability and realised business value. ### Why do most enterprise AI pilots fail? MIT's 2025 GenAI Divide research found roughly 95% of enterprise generative AI pilots produced no measurable profit-and-loss impact, attributing failures to brittle workflows, absent contextual learning and misalignment with daily operations rather than to model quality. The study's sample size has been criticised, but its diagnosis aligns with widely reported enterprise experience. ### What is Microsoft's Frontier Company? Frontier Company is a Microsoft unit announced in July 2026 with $2.5 billion in committed investment and 6,000 employees, led by Rodrigo Kede Lima. Its engineers embed directly with enterprise customers to co-design and continuously improve AI systems measured against business outcomes, rather than delivering software for customers to deploy themselves. ### Should companies build or buy enterprise AI systems? The framing is increasingly unhelpful, because commodity layers now cover model access and agent orchestration. The more useful question is where deployment knowledge will live — with an internal platform team, a vendor's embedded engineers, or an integrator — since that determines whether workflow expertise accumulates inside your organisation or outside it. --- *Editor's note — this is analysis and opinion; the argument and market forecast are Edgewisely's own. Reported facts are sourced from:* [*CNBC*](https://www.cnbc.com/2026/07/02/microsoft-commits-2point5-billion-6000-employees-ai-implementation-unit.html?ref=edgewisely.com) *and* [*The Decoder*](https://the-decoder.com/microsoft-launches-2-5-billion-frontier-company-to-embed-6000-ai-engineers-inside-enterprise-clients/?ref=edgewisely.com) *on Microsoft Frontier Company;* [*Forbes*](https://www.forbes.com/sites/jaimecatmull/2025/08/22/mit-says-95-of-enterprise-ai-failsheres-what-the-5-are-doing-right/?ref=edgewisely.com) *and* [*AI Magazine*](https://aimagazine.com/news/mit-why-95-of-enterprise-ai-investments-fail-to-deliver?ref=edgewisely.com) *on MIT's GenAI Divide findings;* [*OpenAI*](https://openai.com/index/introducing-the-agents-api/?ref=edgewisely.com) *on the Agents API. The 95% figure is contested on methodological grounds and is presented here with that caveat.*